Conference Agenda

The intricate interplay between soil moisture, water retention, and carbon sequestration represents a significant challenge in comprehending and effectively managing terrestrial ecosystems. These interdependent variables play pivotal roles in shaping agricultural productivity, informing water resource management strategies, and contributing to climate change mitigation efforts. Therefore, it is crucial to accurately model and predict their dynamics and interdependencies.

This study outlines a comprehensive methodology for creating a dataset designed to explore the correlations between soil moisture, water retention, and carbon sequestration, employing satellite imagery. The primary objective is to enable the application of machine learning techniques for estimating future values based on historical trends.

The dataset construction process involves preprocessing and integration of diverse data sources, such as Earth observation data and ground-based measurements. Emphasis is placed on incorporating essential temporal and spatial features to ensure a robust analysis of soil conditions. The resulting dataset is anticipated to be instrumental to uncover patterns and correlations between soil moisture, water retention, and carbon sequestration.

This paper underscores the significance of integrating advanced technologies and multi-dimensional datasets, highlighting their potential to furnish users with a powerful tool for predicting future environmental variables. The overarching goal is to drive informed decision-making processes in the realm of sustainable land and water management. By combining cutting-edge technology with rich datasets, this approach aims to empower stakeholders in making smart and strategic decisions to address the challenges posed by changing environmental conditions.

The ongoing war in Ukraine has drawn international attention to the impact of conflict on global agricultural production and food security. On June 6 2023, the Kakhovka Dam in Southeastern Ukraine collapsed under attack, draining the 2000 square km reservoir which served as a source of water and power for 20,000 people [1]. The loss of this critical agricultural and energy infrastructure has left farmers on the left bank of the Dnieper River without irrigation in the historically arid Kherson and Zaporizhzhia oblasts [2]. In this study, we use Earth Observation to measure the change in irrigation coverage in the occupied region resulting from the loss of the dam. Given the current conflict and the lack of ground-truth data from the occupied regions, we use human-labeled validation sets derived from the greenness, thermal, and wetness characteristics of irrigated fields [3], using pre-collapse (2020-2022) imagery from PlanetScope, Landsat-8 and 9, and Sentinel-2 as the baseline. The FAO WaPOR v3 actual evapotranspiration and interception dataset is combined with precipitation in a Water Deficit Index to differentiate irrigated fields from rainfed fields. Preliminary results from an unsupervised approach show a significant reduction in 2023 of the types of fields identified as ‘irrigated’ in the pre-collapse years. Given that much of the summer crops would have been planted by June, this change in 2023 is hypothesized to be due to lack of irrigation in already-planted fields, rather than farmers changing crop types in response to irrigation loss.

References:

[1] Naddaf, M. (2023). Ukraine dam collapse: what scientists are watching. Nature, 618(7965), 440-441.

[2] Vyshnevskyi, V., Shevchuk, S., Komorin, V., Oleynik, Y., & Gleick, P. (2023). The destruction of the Kakhovka dam and its consequences. Water international, 48(5), 631-647.

[3] Deines, J. M., Kendall, A. D., Butler, J. J., & Hyndman, D. W. (2019). Quantifying irrigation adaptation strategies in response to stakeholder-driven groundwater management in the US High Plains Aquifer. Environmental Research Letters, 14(4), 044014.

Potato is the fourth staple food crop, and its planting area is constantly expanding. Accurate acquisition of potato distribution is of great significance for planting area detection, yield estimation and the planting structure adjustment. To this end, this study developed a field-parcel-based methodology for mapping Potato distribution by integrating the edge detection, image segmentation and machine learning algorithms based on multi-temporal Sentinel-2 data. The Canny edge detection results from single 10 m resolution bands of Sentinel-2 data (blue, green, red, and near-infrared bands) are aggregated by different weights can detect richer edge information and provide sufficient details for field parcel extraction. Then the watershed segmentation is used to extract the field parcel with accuracy of 85%. the random forest machine learning classifier was used by combing the spectral and index features to identify potatoes at the parcel scale and the mapping accuracy achieved 84%, which can provide technical support for accurate potato management, accurate area and yield estimation.

Timely information of pasture condition is essential for the management of cattle rotation. Farmers have to decide when and where to move cattle herds, based off the grazing activities within active pastures and the recovery rate of vacant pastures. Traditionally, this information is gathered with time consuming manual surveying techniques using Plate Meters. However, Earth Observation can observe the grass condition across a wide management area in a timely and consistent manner, helping farmers to survey reduce and acting as in invaluable, scalable management tool.

The SmartFarm system utilizes both Sentinel 1 & 2 in near-real time to estimate average pasture grass biomass (kg ha^-1). Users register fields to include in their management plan, where each field’s biomass will be comparatively displayed against other fields in a ‘Grass Wedge’ management graphic.

This system works by firstly generating biophysical variables from Sentinel 2 data, namely LAI and FAPAR, to understand the photosynthetic health of the grass. However, the timely utilization of these variables can be limited by cloud cover. Therefore the SAR backscatter response of the field, measured by Sentinel 1, is compared to the average backscatter signal of the wider landscape in a 3km radius. Using the difference between these signals reduces the impact of atmospheric noise on the Sentinel 1 signal and yields a reliable, comprehensive measure of the vegetation structure. A deep learning model was trained using these EO signals and 10,000 measurements of grass biomass gathered using Plate Meters. The nature of the data was challenging to model due to its fast, often sub-monthly cyclical patterns, but the model performed well and achieved an RMSE of 708 kg ha^-1 and a r² of 0.27. The system has been operational for over 8 months and routinely estimates pasture biomass for 523 fields across the UK.

The green area index (GAI) is a key biophysical variable for crop monitoring. The most accurate methods for its large-scale estimation rely on optical remote sensing data. However, these can be hampered by frequent cloud cover. In this context, synthetic aperture radar (SAR) offers the advantage of being able to provide dense time series that can be used to complement the sparse GAI series derived from optical data. In this study, SAR-to-optical GAI regression is performed using a transformer neural network with past and current values of SAR backscatter and interferometric coherence, as well as past values of GAI when available. Sentine-1 and -2 data acquired from 2018 to 2021 over the Hesbaye region of Belgium are used for cross-validation. The model is trained on three growing seasons and tested on the fourth for each fold. The results show that the model can successfully predict Sentinel-2-derived GAI with an average R2=0.88 and RMSE=0.74, outperforming methods relying on radiative transfer model (e.g., Water Cloud model) inversion. The method is also validated with data collected in situ in eight maize fields in Belgium (R2=0.87 and RMSE=0.75). These promising results pave the way for the generation of accurate, dense LAI time series throughout the growing season, allowing for timely crop monitoring in cloud-prone regions.

Crop insurance is a key element to increase the resilience of farmers, particularly small-scale farmers in countries exposed to the impact of climate change. Because Multi-Peril Crop Insurance based on individual loss assessment is too costly to operate for millions of small-scale farmers, Weather Index Insurance (WII) has been introduced decades ago. WII uses weather parameters, such as rainfall to determine payouts. More recently, Area-Yield Index Insurance (AYII) – a crop yield loss policy – has been introduced: it is particularly suited to the needs of small-scale farmers by providing a more comprehensive loss of yield protection for natural, climatic, and biological perils compared to WII.

Based on elements from both WII and AYII, a hybrid solution, Weather Area Index Insurance (WAII) is proposed: it is an index that on one hand significantly improves WII, and on the other hand streamlines AYII. The reasons behind this new index are:

1. WII is exclusively based on rainfall data, hence not considering when, where and how much crop area have been effectively planted and the seasonal phenological crop development. This solution creates uncertainties in payouts.
2. AYII, as applied in RIICE, is an advanced and complete insurance index solution. However, yield estimation requires a plant growth simulation model specific to each crop. Insurance companies tend to shy away from this solution because a time-consuming calibration period is needed. That is, an AYII solution should rather be understood as a long-term objective when stepping into a new geography.

WAII Rainfall and Flood indexes are obtained by combining satellite rainfall data, seasonal cultivated area and associated crop growth trends derived from high resolution remote sensing time-series, and crop-specific triggers at key phenological stages. In collaboration with stakeholders looking at low-cost insurance solutions meeting farmers' actual needs, WAII is piloted on maize, cassava and rice in Cambodia.

Independent, high-quality, and comprehensive harvest statistics are an indispensable information basis for the government, economy, and the general public. So far, regional harvest statistics in Germany have relied on labor-intensive estimates from knowledgeable experts, which are becoming increasingly difficult to recruit.

Sentinel data combined with machine learning offers a promising and objective method to not only ensure harvest statistics at the regional level but also to increase spatial granularity even further. In a pilot project, the Hessisches Statistisches Landesamt (Hessian Statistical Office) has developed an automated procedure to model crop yields for four important crops (winter wheat, winter barley, winter rapeseed, winter rye) comprehensively at the field scale. This procedure has been applied to six German states for the years 2022 and 2023 and will be scaled up nationwide in 2024 for future integration into operational statistical production.

To achieve this, two periods in May and June of each year with the least cloud-cover were identified, and Sentinel-2 L2A images were assembled into mosaics per period. Sentinel-2 data were combined with field geometry and crop type information from the Integrated Administration and Control System, in-situ yield measurements from official statistical surveys, as well as additional meteorological and soil variables to train an ensemble of various machine learning algorithms. These models were then used to model harvest yields for all fields of each respective crop, which can be aggregated to any desired units.

The cross-validation results reveal relative errors ranging from 10.6% for winter rapeseed to 19.8% for winter rye. These errors seem to be primarily influenced by the size and variability of the available training data. Hence, by incorporating data from additional states and years, we anticipate a further reduction in the prediction error associated with harvest yield.

Tillage is a fundamental agricultural practice aimed at preparing soil for planting by loosening it and integrating organic matter. The choice and frequency of tillage methods significantly influence agricultural field productivity and health. Tillage practices are categorized based on the percentage of the soil surface with crop residue cover (CRC). Conventional tillage involves extensive soil disturbance with CRC <15%, while conservation tillage aims to minimize soil disturbance with CRC >15%. Conservation tillage contributes to soil quality improvement by reducing erosion and greenhouse gas (GHG) emissions.

Sentinel-1 (S1) and Sentinel-2 (S2) play a vital role in detecting and monitoring tillage practices. S1 and S2 provide information on tillage patterns and field characteristics, aiding in the differentiation between conservation and conventional tillage practices. Vegetation indices, e.g., Normalized Difference Vegetation Index (NDVI) and Normalized Difference Tillage Index (NDTI), derived from S2, aid in detecting CRC. S1 offers all-weather monitoring capabilities for detecting tillage by capturing changes in surface roughness and moisture content associated with tillage. Moreover, the Integration of satellite data with Artificial Intelligence (AI) algorithms enhances tillage detection accuracy and efficiency.

A challenge exists in distinguishing between crop residues and soil, especially in smallholder farming systems. However, capabilities in satellite technology, such as high spatial and spectral resolution of S2 and S1 imagery, offer solutions to these challenges.

In “AI-based tillage detection for improved agricultural and climate policies” project funded by ESA, we aim to develop AI models, i.e.,1D ResNet, for detecting conservation and conventional tillage using S1, S2, and Landsat data. The resulting models will be validated in collaboration with the Baltic Sea region’s agricultural paying agencies. Currently, we are in the process of assessing the feasibility of detecting conservation tillage and developing the conventional tillage detection model with the following achieved preliminary results: F1 score: 0.80, TNR: 0.63, TPR: 0.78.

The spatial distribution of global cropland extent serves as the foundation for various analysis in agricultural applications, including crop growth monitor and water management, natural resource utilization, environmental assessment, public health initiatives, and sustainability evaluations. With the advent of cloud computing and the better availability of high-resolution satellite imagery, numerous cropland extent products have emerged, typically offering resolutions ranging from 10 to 30 meters. However, these existing datasets exhibit significant disagreement in cropland extent delineation. In the present work, we addressed the disagreement among six existing cropland extent datasets (WorldCover, ESRI LC, GLAD, FROMGLC, Globeland30, and GLC-FCS) for the year 2020. Our analysis revealed that 51.27% of the potential global cropland areas (defined as regions classified as cropland by at least one product) were consistently labeled as cropland by all six datasets considered. We identified two primary factors contributing to the observed disagreement: (1) variations in the definition of cropland among different products and (2) misclassification during the cropland identification process based on satellite data. To further investigate the impact of misclassification, we selected three products with cropland definition similar to the "temporal crop" category in FAOSTAT (WorldCereal, WorldCover, and annually composited Dynamic World). Despite using similar cropland definition, disagreement maps still revealed that 34.76% of potential "temporal cropland" areas exhibited discrepancies among the three selected products. Verification using visually interpreted validation samples conducted in North Europe and North Asia, which are regions of large disagreement, demonstrated that WorldCereal achieved the highest classification accuracy, with an overall accuracy of 84.19%, and WorldCover exhibited the strongest correlation with statistical data from FAOSTAT (R² = 0.92). This work addresses the importance of conducting global cropland extent data following an unform applicable definition, thereby enhancing the utility of satellite-derived datasets for decision-makers and stakeholders across both public and private sectors.

The food system presents a huge challenge for the planet in terms of producing sufficient nutritious and affordable food, but also in reducing the destructive planetary impact that food is having. New technologies can make a significant contribution to this goal, especially when supported by the secure and permission-enabled sharing of data.

SPADE EU Project (HE 101060778) is acting within the research and innovation front of the agri-food and environment sectors. Technologies employed involve Unmanned Aerial Vehicles (UAVs) used for Earth Observation, also catering for Agriculture and Environmental monitoring (livestock, cropping, forestry) and aiming to support evidence-based decisions for improving food security at national to global scales.

The SPADE ecosystem involves deployment of UAV formations (single UAV, collaborating UAVs, UAV swarms) equipped with edge-computing devices (AI/ML) that detect focused risks (e.g. livestock grazing health care risks).

This work will present the first results of SPADE in livestock use cases while also focusing on how a trustworthiness framework enables connectivity among the topics of edge computing AI tools, ML modelling & data assimilation, livestock assets digital twinning and needs arising or changing due to the environment.

The aim is to establish a collaborative ecosystem for securely sharing and exchanging data amongst participating actors such as farmers, drone providers, regulators and others in order to maximise the benefits of new and aggregated data whilst minimising the risks associate with data sharing. A trustworthiness data framework will enable a collaborative governance model that considers all participant stakeholders interests as well as optimising environmental factors.

The broader vision is to integrate farm and supply chain level data sharing with regional, national, and global scale exchanges. This in turn will enable other entities such as regulators, policymakers, meteorologists and other satellite services to collectively address food challenges through benefitting from evidence-based knowledge and insights.

Over the last decade, food security has become one of the world’s greatest challenges. Reliable, timely and legitime information on food production is required to inform decision-making process. The main expectations about the Earth Observation (EO) contribution to the agriculture statistics are cost-efficiency, better granularity and timeliness improvement.

The ESA Sen4Stat project demonstrates validated open source tools and best practices for national agricultural statistics with Sentinel data and facilitates the EO uptake by the National Statistical Offices.

The Sen4Stat toolbox processes Sentinel-1 and Sentinel-2 according to the state-of-the-art and delivers 5 types of products: (i) 10-m optical and SAR temporal syntheses, (ii) spectral indices and biophysical variables time series, (iii) 10-m crop type maps, (iv) crop growth conditions metrics and (v) crop yield estimations at various aggregation levels. The project is working with pilot countries such as Spain, Senegal, Pakistan, etc. to address a wide diversity of cropping systems and agricultural data collection protocols and sampling frames.

In Spain, a 10-m crop type map was generated (F-Scores for the main crops higher than 80%) and coupled with national statistical survey to allow reducing the confidence interval around the acreage estimates by more than 50%. An irrigation map was also produced to update the sampling frame. In Senegal and Mali, the data collection protocols were adjusted to facilitate the integration of EO data and improve the acreage and production estimates. In the Sindh province of Pakistan, a pilot activity is ongoing with an ad-hoc field campaign co-organized with the Ministry of Agriculture to estimate the irrigated wheat area.

The Sen4Stat toolbox is available for download and the next 18 months will be dedicated to capacity building activities for the growing community. FAO and World Bank are also actively contributing to the Sen4Stat uptake through the EOStat programme and pilot activities.

The conservation and creation of landscape features is recognised as a key conservation tool to halt the loss of agricultural biodiversity in European farmland.

This study introduces a new indicator to quantify landscape features in EU agricultural land, based on the LUCAS Landscape Feature survey. We developed a comprehensive methodology to measure and categorise landscape features, distinguishing Woody, Grassy, Wet, and Stony LF types. Our approach gives a robust and reproducible estimate of the indicator at the EU Member State and possibly regional levels, based on a reliable and statistically representative sample of landscape features.

The methodology combines office-based photo-interpretation with field surveys collecting 3.8 millions field points, ensuring accuracy in determining the presence and type of landscape features within agricultural contexts. Together with information on biodiversity and ecosystem services, it will play a crucial role in evaluating the performance of major policies related to biodiversity conservation in agricultural lands, aligning with the Common Agricultural Policy and the EU Biodiversity Strategy for 2030. Besides, it will play a role in the assessment of natural based solutions for mitigating climate change effects, biodiversity loss and crop production (food) security.

Our findings reveal that, in 2022, landscape features covered 5.6% of EU agricultural land. Woody features were the most prevalent, followed by Grassy, Wet, and Stony features. The percentages of landscape features varied across EU Member States, with Malta and Cyprus exhibiting higher values.

The novel indicator developed is based on a comprehensive and reproducible method for quantifying these features, providing essential insights for policy and decision-making in sustainable agriculture.

With the adopted amendment of the EU regulation 2018/841 on the inclusion of greenhouse gas (GHG) emissions and removals from land use, land use change and forestry (LULUCF), the EU member states agreed that – starting with the report in 2028 – the calculation of emission pools at national level should make use of geographically-explicit data. Earth Observation (EO) can support the implementation of the regulation by providing timely, seamless and high-resolution information for monitoring land use activities and land management practices related to GHG emissions and removals.

Frequently, EU-wide mapping initiatives that make use of data from the Copernicus program and derived products provided by the Copernicus Land Monitoring Service (CLMS) are complemented by national-level approaches that usually aim at generating more tailored datasets for specific monitoring requirements. In this context, the project KlimaFern funded by the German Federal Ministry of Food and Agriculture evaluates the potential of new EO-based national datasets on agricultural land use to enhance climate reporting in the LULUCF sector for Germany. We will present preliminary results of mapping area-wide GHG-related land use activities such as crop rotations, grassland conversion or planting of hedgerows and coppices. These products are derived on a national scale using state-of-the art machine and deep learning algorithms and multi-modal satellite image time series (e.g., Sentinel-1 and 2, Landsat, PlanetScope). All products are compared against available data at national level to assess their potential for improving climate reporting and are evaluated in terms of quality, accuracy and consistency against existing and foreseen products of the CLMS.

The presentation will summarize the preliminary project results and highlight challenges for a successful implementation of EO data for monitoring obligations. Finally, it will point out synergies and relationships of climate related land use monitoring efforts with other policy initiatives at national and EU level.

The escalating incidence of food fraud on a global scale, driven by economic incentives and increased consumer demand, poses significant challenges to food integrity and safety. Food adulteration, including the deliberate substitution or addition of inferior materials to enhance appearance and profitability, has become more prevalent, particularly with the expansion of specialized food products like organic and protected geographical indication (GI) foods. Traditional detection methods entail high costs and time demands, necessitating the pursuit of alternative approaches.

The THEROS project addresses these challenges by developing an integrated toolbox to modernize the verification process for organic and GI food products, enhancing traceability, security, and transparency in the supply chain. Leveraging technologies such as Earth Observation, photonics, IoT, DNA metabarcoding, blockchain, and artificial intelligence, the project aims to detect and prevent adulteration effectively. Key features of the THEROS toolbox include advanced analytics, machine learning algorithms, and digital interfaces to streamline verification processes. Interoperability with existing control systems ensures seamless integration, while harmonized data sharing facilitates improved decision-making among stakeholders.

The project's Earth Observation-based monitoring approach, initially developed for CAP compliance monitoring, will be expanded within THEROS, to enable large-scale verification of organic farming practices. Machine learning algorithms will customize detection algorithms to identify distinct patterns indicative of organic, under-conversion, and conventional crops, enhancing monitoring accuracy throughout the growing season. Statistical approaches will facilitate intra-field analysis, verifying practices like crop rotation and biodiversity maintenance. Additionally, will extract biophysical parameters to monitor organic crop conditions and phenological stages, providing recommendations for optimal cultivation practices. Methods to assess carbon footprint will ensure organic agriculture acts as a carbon sink.

Summarizing, THEROS project embodies a multi-actor approach to combat food fraud, offering high-quality labeled food products to consumers while improving trust, traceability, and sustainability across the EU food supply chain.

Sustainable agriculture is pivotal in achieving the 2030 United Nations Agenda, ensuring zero hunger and environmental preservation. Meeting global food demand requires producing more with less, necessitating effective agricultural monitoring. Timely crop mapping is crucial for various activities, including water management, supply chain control, and crop risk assessment. Satellite remote sensing offers a reliable means for crop mapping, particularly utilizing Sentinel mission data for its temporal frequency and multi-spectral capabilities.

However, current operational products often lack the capability for operational in-season monitoring due to limited training data and complete time series observations. To address this gap, we propose an innovative "in-season mapping approach" utilizing Sentinel-2 satellite data and dynamic crop presence probability maps generated through a Deep Neural Model. Our approach provides both long-term in-season mapping and short-term dynamic mapping, offering spatially explicit information on crop sequence, management and phenological development. Our model, a fully convolutional 3D CNN incorporating Feature Pyramid Network (FPN), simultaneously extracts spectral-spatio-temporal features for identifying crop dynamics over time.

The model produces dynamic segmentation maps at each satellite passage (short-term mapping) and aggregates them to generate annual in-season maps, capturing crop presence and sequence (long-term mapping). Our contributions include developing a novel approach for in-season mapping, providing insights into when and where crops are cultivated, and offering a technique for training and evaluating short-term segmentation maps. Evaluation on multi-site, multi-season datasets in the north of Italy demonstrates the reliability and accuracy of our model, with satisfactory overall accuracy and kappa measures for both short-term and long-term predictions. By making our dataset public, we aim to facilitate reproducibility and comparison with other models, fostering advancements in agricultural monitoring and sustainable food production.

The use of plant protection products near urban areas has raised concerns about potential pesticide exposure to nearby residents and the environment. In response, this study aims to estimate the impact of pesticide bans near sensitive areas, particularly urban regions, across the European Union. Using available EO EU data on urban settlements and crops, we estimate the agricultural areas affected by the ban and analyze the relative agricultural production loss. The study utilizes the EU Crop Map 2018, Corine Land Cover (CLC) 2018, and the Global Human Settlement layer to gain a comprehensive understanding of land distribution and characteristics in proximity to sensitive areas. Results include the percentage of agricultural land within 10 meters of urban areas, highlighting country and regional differences, and identifying crops with economic significance closer to urban areas. Furthermore, the study quantifies the hypothetical impact of the ban, explores the potential reduction in sales, and estimates the theoretical amount of pesticides used in these zones. This comprehensive EU-wide analysis aims to provide valuable insights for policymakers and stakeholders, addressing the need for evidence regarding the impact of pesticide bans in sensitive areas.

The Horn of Africa is currently grappling with its worst drought in 40 years, resulting in significantly reduced agricultural productivity and food security after five consecutive failed rainfall seasons. Nearly 1.2 million people have been internally displaced due to the impact of drought on pastoral and farming livelihoods, exacerbating hunger in 2022 [1,2]. Earth Observation (EO) plays a crucial role in gathering and analyzing information related to climate drivers.

We adopt a data-driven modern approach to understand the complex processes involved (climate-migration-conflict-food insecurity) that help evaluate and adopt mitigation policies [3]. We collect climate and socio-economical data from sources and introduce a causal inference framework for Somalia to evaluate the impact of cash-based interventions on food security. Our contributions lie in leveraging the abundance of EO data to understand the dynamic system of food insecurity, where climate and socio-economic factors interact in a complex manner, and in estimating the effectiveness of humanitarian interventions in mitigating food insecurity levels. In particular, we assess the Average Treatment Effect [4] of humanitarian aid as cash transfers, on the IPC index, considering different spatial and temporal resolutions. This work ties directly with SDG-2. Assessing the causal effects of humanitarian interventions would facilitate a quantitative estimation of policies and a way to improve them in future emergencies. Results suggest that surpassing certain thresholds of cash aid has a positive effect on the level of IPC: as the number of beneficiaries increases, the level of IPC decreases.

[1] WMO. Africa suffers disproportionately from climate change, 2023.

[2] WFP. Impacts of the Cost of Inaction on WFP Food Assistance in Eastern Africa (2021 & 2022), 2023.

[3] Fyles, H., & Madramootoo, C. Key drivers of food insecurity. In Emerging technologies for promoting food security. Woodhead Publishing. 2016.

[4] Pearl, J., Causality, Cambridge University Press, 2 edition, 2009.

Mapping and characterizing cocoa planted areas with Earth Observation data and accurately disentangling them from other land cover is not only paramount for effectively monitoring and reporting on sustainability goals related with cocoa production but also for the EU Deforestation Regulation. However, accurately representing the complexity of the cocoa planted area is a challenging task. Cocoa is grown mostly on smallholder plantations with various agricultural practices, ranging from mono-cultural plantations to agroforestry systems with cocoa shaded by other trees with varying densities and spatial distribution. Here we combine a curated dataset of cocoa plot location and very high resolution (VHR; 0.5m) multispectral satellite imagery covering ∼33% of Ivory Coast area, in a deep learning framework to map cocoa. The selected deep learning model is based on a U-net architecture with efficient-netb5 encoder. To train the model, batches of tiles of 512x512 pixels were used and two sample sizes were tested: i) 221,158 and ii) 2,069,855 (full dataset) tiles. Both samples were split into 70% training and 30% validation. An independent and randomly selected VHR image (66,244ha) served as a test set. Despite the heterogeneity of cocoa plantations, our model was able to generalize well and to differentiate between cocoa and non cocoa areas accurately at this unprecedented spatial resolution. Results show that the improvement related to the use of a larger sample was limited (F1: +2.3%) and not proportionate considering the increase in training time (22h to 153h). The best performance metrics on the test set with the first (smaller) sample size gave a F1 score of 0.92 with Precision and Recall of 0.93 and 0.91 respectively. Building on the results of this study, future work will focus on the discrimination of mono-cultural cocoa from cocoa grown areas with different shade tree densities.

Sentinel-1 data enables continuous environmental monitoring regardless of weather conditions and time of day. This capability is crucial for agriculture, where timely information is essential for decision-making. The goal of this study was to explore Sentinel-1 VH and VV backscatter coefficients (σ⁰) data in providing detailed information on maize growth stages, estimation of soil moisture content (SM), and soil clay content (SC) in an agricultural area of Casalina, Italy. Specifically, correlation coefficients (r) were calculated to quantify the strength and direction of the relationship between σ⁰VH, σ⁰VV and maize height (MH), SM and SC, respectively. Satellite images (N=8) were acquired between 09^th July and 4^th September 2023. For correlating SM and SC with σ⁰VH and σ⁰VV data only images with bare soil pixels were selected (N=2) to eliminate the backscatter influence of vegetation. In-situ MH was measured from 17^th July to 04^th September, i.e. from ~20 to ~350 cm of crop height, while in-situ SM was measured on 09^th and 16^th July and varied from ~17 to ~45% (average ~26%) on both dates. In-situ SC (considered relatively stable over time) was measured in late 2018 and averaged 27-32%. The r values between MH-σ⁰VH and MH-σ⁰VV were 0.54 and 0.18, between SM-σ⁰VH and SM-σ⁰VV 0.01 and 0.08, and between SC-σ⁰VH and SC-σ⁰VV -0.26 and 0.39, respectively. MH-σ⁰VH and MH-σ⁰VV were more correlated in the period from 16^th July to 02^nd August, i.e. from ~20 to ~150 cm (r= 0.7 and 0.65, respectively), indicating a stronger association during this period. Although the relationship between SM and SC with backscatter coefficients was weaker, Sentinel-1 data could provide valuable insights into agricultural dynamics, offering farmers timely information for informed decision-making and resource management practices.

Agriculture can make a significant contribution to the reduction of atmospheric CO2 through adapted cultivation methods and actively binding CO2 through the targeted build-up of humus in the soil. In addition, more resilient farming methods such as crop rotation and the planting of catch crops contribute significantly to more sustainable food production.

To exploit this potential of carbon farming, a solution for monitoring, reporting and verification (MRV) of these measures is required. Valuable information on vegetation and soil can be derived from high-resolution, globally available Earth Observation data from the Copernicus program. These data are therefore ideally suited to enable this solution, both for past and future periods.

Within the ESA Business Application project "EO4CarbonFarming", such an MRV tool is being developed based on Copernicus data. It can monitor the growth of catch crops and verify farming measures taken to assure CO2 uptake and humus build-up in the soil. To facilitate this, high-resolution optical and radar data (Sentinel-1, Sentinel-2) are used in combination with high-quality pre-processing methods, a radiation transfer model and newly developed algorithms specifically for the derivation of humus in the soil.

The sound determination of carbon stocks in fields is an essential component of assessing the effectiveness of carbon farming for the climate and to develop business models on this basis.

For a pilot region in Austria, carbon storage in the soil has been calculated in 2017 and 2022. From these analyses, carbon sequestration is computed and evaluated on an aggregation level. Additionally, patterns of SOC content are analysed, both for the individual timesteps and the changes between them.

This project is funded within ESA Contract Number 4000134139/21/NL/MM/mr.

The primary challenge in crop mapping and monitoring lies in the absence of field-level crop labels especially in developing world, impeding the advancement of training supervised classification models. The objective of this study is to investigate the capabilities of the C-band Sentinel-1 (S1) Synthetic Aperture Radar (SAR) for developing generalized crop type models, particularly targeting the identification and monitoring of sunflower crops in Ukraine on a national scale without requring in-season field labels. Globally, the sunflower is the fourth most essential oilseed crop with Ukraine dominating as the largest producer and exporter. This study examines the interaction between S1 signal and sunflower, aiming to identify and monitor the phenological stages of sunflower. The analysis encompasses SAR backscattering coefficients and polarizations in VH, VV, and VH/VV ratio, emphasizing disparities between ascending and descending orbits attributable to sunflower's directional behavior. Utilizing the distinctive SAR-based signature of sunflower, the study presents a generalized model for the identification and mapping of sunflower fields. The model utilizing features from S1-based descending orbits demonstrated superior performance compared to that based on ascending orbits due to sunflower's directional behavior, achieving F-score of 97%, in contrast to F-score of 90% for ascending orbits. The generalized approach to map sunflower was applied to assess the impact of the Russian full-scale invasion of Ukraine. The national sunflower planted areas and corresponding changes in 2021, 2022 and 2023 were estimated using a sample-based approach for area estimation. Sunflower area was estimated at 7.10±0.45 million hectares (Mha) in 2021 which was further reduced to 6.72±0.45 Mha in 2022 representing a 5% decrease. In 2023, sunflower acreage remained relatively stable at 6.63 Mha with no significant variation. Our findings bolstering initiatives such as the collaboration between the NASA Harvest and the Ukrainian government, aimed at providing timely information for mapping major crops.

The 'European Monitoring of Biodiversity in Agricultural Landscapes' (EMBAL) is a monitoring initiative initiated by the European Commission that gathers information on the state of biodiversity in agricultural landscapes across all 27 EU member states. Developed within the EU Pollinator Monitoring Framework, EMBAL is a standardized and sample-based in-situ survey of 500x500m landscape sections (plots).

EMBAL provides comprehensive data, including general information on land use and land cover, information about landscape elements and pollination potential at parcel level as well as specific vegetation data on a transect level in grassland and arable habitats. Both the methodology and the sampling frame are harmonized with LUCAS (Land Use and Coverage Area frame Survey).

Following a successful pilot in 2020, EMBAL was applied in all 27 EU member states in 2022 and 2023, surveying a total of 3,000 selected plots in both years. This extensive rollout served to gather harmonised baseline data on biodiversity across EU27 and provided a comprehensive field test of the EMBAL methodology across different European landscapes.

In this contribution, we offer an overview of the EMBAL 2022 and 2023 rollout, the EMBAL survey methods and parameters and provide an outlook on the results.

Carbon farming is one of agriculture's answer to climate change and includes agricultural practices able to capture and store C in soils as soil organic carbon (SOC). Unlocking the potential for Carbon farming to scale relies on the establishment of robust protocols to monitor, report, and verify (MRV) changes in SOC stocks. Different MRV protocols are available in the voluntary C market, exploiting different approaches to quantify C removal. Capturing spatial and temporal variability of SOC can be challenging since SOC values varies substantially over space and changes occur slowly through time. Some of these issues can be tackled indeed by “hybrid approaches”, i.e. by combining remote sensing (RS) and process-based models with direct field measurements to verify model predictions.

In this context, project Remote-C, funded by Italian Research Ministry, aims at developing an approach to estimate changes in SOC thanks to a spatialized version of Roth-C model fed by RS products and spectroscopic readings from proximal soil sensors. The main goal is to understand if remote and proximal sensing are added values in delivering timely and spatially accurate inputs to reduce uncertainty of soil C model. Overall project scheme is given in an attached file.

To develop and test the operating MRV tools addressed in Remote-C the consortium will exploit existing test site made available by 2 EU-funded projects (PRIMA - Farms4Climate and H2020 - ClieNFarms). These farms are pioneers in testing carbon farming solutions eligible for payment schemes. RS data will be exploited to i) map biophysical variables of crops from multispectral data; ii) characterise crop residues and iii) detect tillage operations with SAR data. These variables will be ingested by a light use efficiency model (e.g. SAFY) and outputs from crop model and RS data will be exploited by a spatial modelling toolbox based on the Roth-C model.

The project is in its early stages and the workshop could be a suitable arena to discuss the proposed approach.

There is currently a lack of high-resolution pan-European information on land use management, especially in terms of how intensively and extensively grasslands are managed. This type of information is needed for economic land use modelling and for assessing policy impacts, such as the latest reforms from the Common Agricultural Policy (CAP) and other European Union Green Deal targets. Here we present the results of a grassland management map for Europe that uses the Copernicus Corine land cover for 2018 as the basis for allocating grazing livestock (obtained from statistical sources and informed by expert knowledge on national grazing) to relevant land cover types. Using different densities of livestock (calculated from statistical sources), we use a rule-based system to map ten grassland management types to the Corine land cover map at a resolution of 100 m. These include classes such as high-, moderate- and low-density pasture systems, high-, moderate- and low-density managed grassland systems, rough grazing, silva-pastural agroforestry, and managed and unmanaged semi-natural systems. The map is currently being validated using experts as well as remotely sensing products such as the frequency of mowing events, which are currently available for individual countries such as Germany. Once the Copernicus very high-resolution layer on grassland mowing event data becomes available in the latter half of 2024, these grassland management classes will be further refined using this pan-European data set.

Australian broadacre crop production covers more than 20 million hectares across various agroecological environments, from Mediterranean-type climate and poor sandy soils in the low-rainfall zones of Western Australia to tropical climate and fertile deep soils in Queensland, for example. The main crops grown in Australia are wheat, canola, and barley. Here, we will focus on wheat with an average yield gap of > 1.7 t/ha or 50% of the water-limited yield. Changes in agricultural practices, such as a shift toward earlier dry sowing and improved control of fallow weeds, have improved water use efficiency and maintained productivity despite climate change-imposed reduction in wheat potential yields in Australia. Improvements in genetic material with better and broader adapted cultivars further contribute to improved grain production. Albeit existing opportunities for further enhancements, such as more skilful and timelier management (e.g., nitrogen), some farmers are closing the yield gap, achieving potential production in their environments. However, climate change, particularly with shifts in rainfall patterns and extreme events, continues to put pressure on Australian farmers and increases their risk exposure. Thus, evaluating the impact of farmer practice on changes in the yield gap, as well as on risk and sustainability outcomes, is yet to be fully realised. Using a combination of remote sensing data, bioeconomic modelling, and grower surveys we propose a three-component framework to systematically assess the performance and impact of technology adoption at the farm level. We discuss our developments, challenges, and opportunities to align the various dimensions of yield-gap time series and its drivers at a continental scale.

In the context of the Paris Agreement and the Green Deal targets, the agricultural sector needs to be converted from a source of more than 10% of EU's greenhouse gas emissions in to a significant sink. All agricultural practices that contribute to the conservation and sustainable increase of carbon stocks in the soil are grouped together under the term carbon farming. Among these agricultural practices, cover crops appear to be one of the most promising ways of sequestering more carbon in the soil. A cover crop is a non-commercial crop planted during the fallow period when the soil is typically bare between the cash crop harvest and the following season's planting.

Agricultural practices transformation requires an operational monitoring system to assess the impact of these practices on the carbon balance of agricultural parcels. Previous studies have shown that it is possible to detect the presence or absence of cover crops using Sentinel-1 (S1) and Sentinel-2 (S2) images. The Sen4CAP system first designed for the CAP paying agency run on calendar year while the agro-ecological activities take place on a long-term basis beyond the crop rotation. This study highlights the potential of the Sen4CAP system to assess the expected impact of consecutive cover crops along with other farming practices. In order to quantify the impact of these covers on the potential carbon storage in agricultural soils, it is necessary to obtain other characteristics about the cover, such as the duration of the cover, the periods of bare soil, the green biomass and the composition of the cover. This study presents the results obtained from 6-year time series analysis of Sentinel-1 and Sentinel-2 dataset at parcel level and assessed the spatial distribution and the evolution of these farming practices at regional level.

The EOX AgriApp is a flexible dashboard service for continuous remote sensing analysis of agricultural parcels or, more generally, any monitoring areas of interest. It visually presents satellite information such as vegetation profiles of different satellite spectral signals interactively linked to image time series down to the parcel level—even allowing for comparison between parcels.

Value-added derived information such as machine learning- or threshold-based markers (crop type, harvest events, mowing events, vegetation cover changes, …) are available as chart annotations and thematic map layers. A compliance rule engine evaluates specific parcel-level requirements and provides customised thematic map visualisations.

The service is rounded off by a high-performance full-dataset search engine with dynamic filter combinations and interactive charts allowing for detailed analysis of machine learning results and compliance rule engine outcomes.

Usage areas for the EOX AgriApp include—but are not restricted to—Common Agricultural Policy implementation, drought- and irrigation monitoring, CO2 flux visualisation and quantification, nature capital and biodiversity monitoring. In the context of the CAP Area Monitoring System the EOX AgriApp has successfully been in operative use within the Austrian and Irish Paying Agencies since the beginning of 2023, serving near real-time insights for millions of parcels.

During 2024, the Copernicus Land Monitoring Service will start releasing the products of the new High Resolution Layer Vegetation Land Cover Characteristics (HRL-VLCC) as part of the Copernicus Land Monitoring Service (CLMS). The HRL-VLCC groups the annual mapping of vegetated land cover characteristics, jointly producing for the first time the former HRL Forest and HRL Grassland together with the new HRL Crops. Within the latter, mapping of major crop types (HRL Crop Type) and crop management practices relevant for agricultural monitoring (that contain layers depicting date of emergence and harvest for main and secondary crops; bare soil duration; fallow land; and cover crops) are included, while giving continuity and enhancing the already existing layers.

With access to the pre-release of the 2017 to 2021 HRL Crop Type products, we propose an evaluation through a comparison with annual farmers’ declarations. For this purpose, we utilize publicly available GeoSpatial Application (GSA) datasets from eight countries and four years with harmonized crop types merged into a single database named CHEAP (Common Harmonized European Agricultural Parcels). Only parcels and points not used in the model's training set were included in the analysis. The validation process employs classical statistical measures to quantify user and producer accuracies, as well as f1-scores per crop, region, and year. Additionally, insights into the spatial consistency of the products are provided.

Preliminary results assessing the major crop types (including wheat, maize, rapeseed, barley, sunflower, sugar beet, and potatoes) underscore a suitable quality of the products. Independent and thorough quality assessments such as those provided here stand to benefit the uptake and impact of Copernicus Earth Observation products. This quality assessment underpins the future developments of applications using the HRL Crop Types in the domain of crop production assessments, indicator development, and land use assessments.

The high resolution (10m) imagery of the Copernicus Sentinel-2 (S2) sensor leads to a consistent, continuous, high quality and near-instantaneous Earth Observation (EO) information flow. This information can be leveraged to predict and map crop masks and crop type cultivations over a vast AOI within a growing season. In order for such maps to be suitable for crop statistics and food security assessments they need to be of high quality and supported with accuracy assessments. In this contribution we present the steps and workflow to derive crop type maps that are at the core of the Copernicus4GEOGLAM Service activated in Tanzania and Kenya under challenging circumstances with small fields and frequent mixed-cropping.

Firstly a grid based stratified systematic random sampling is applied which takes into account variability in the AOI like agro-ecological zones , elevation and landcover types including irrigation use . VHR imagery and S2 timeseries are used to digitize these samples to be visited by enumerators in the growing season. Field information like cropping pattern, crop type and crop stage is stored in digital forms and send to a database. After a quality assessment the digitized samples are used to train (75%) and validate (25%) classifiers in the IOTA2-toolbox . The classifier is run on timeseries of S2 L3A data (synthesis data of unclouded/undisturbed pixels closest to a centre-date of a month) covering the growing season. The Random Forest classifier yielded highest accuracy with up to 89% overall accuracies for crop masks and croptype maps during different seasons in both countries AOIs.

Utilizing Earth Observation (EO) and Machine Learning (ML) to automate Soil Organic Carbon (SOC) monitoring marks a significant advancement for food security and Sustainable agriculture aligning with the United Nations’ Sustainable Development Goal 2 for zero hunger. In Flanders Belgium, a comprehensive methodology is developed to explore the spatial and temporal content of an extensive collection of over 8000 sentinel 2 images on 680000 hectares of farmland. The scope of the current study is to develop valuable soil health indicators, in support of The Common Agricultural Policy (CAP). This developed methodology combines satellite products from the Copernicus services with precise soil measurements to deploy EO-based ML models for predicting SOC levels through time. The large-scale soil quality data products developed, covered all Flanders, facilitate monitoring that underpins the CAP providing detailed insights at both pixel and parcel level. This approach simplifies the creation of soil quality maps showcasing SOC values relative to average conditions and taking into consideration soil-pedoclimatic factors thus enabling targeted soil health interventions. Such detailed classifications are crucial for the effective management of soil health in Flemish croplands. Current research is focused on improving soil health EO-based evaluation using advanced technologies like sensor data analysis edge computing and Federated AI while ensuring semantic interoperability for improvement. Current efforts are trying to tackle data-sharing challenges and, the ability to integrate IoT sensors and hyperspectral satellite images.
The presented methodological framework addresses the requirements and complexities inherent in soil health and agricultural sustainability and investigates how those research priorities can be aligned with the United Nations’ Sustainable Development Goal 2.

High resolution satellite data and advanced machine learning are used to map key commodity crops, namely cacao, oil palm, rubber, coffee, pasture, and soy, with a primary goal of understanding their spatial distribution and quantifying associated carbon emissions. Such information is relevant for policies related deforestation-free supply chains (i.e. EUDR) and climate change mitigation efforts (i.e. REDD+). Our study aims to develop a robust and accurate mapping framework for major commodity crops across tropical regions, leveraging the power of machine learning algorithms. We seek to provide a comprehensive assessment of land-use changes associated with these crops and estimate their carbon emissions footprint. Utilizing high-resolution Sentinel and Planet satellite imagery and ground truth data, we employ state-of-the-art machine learning algorithms, including convolutional neural networks (CNNs) and location encoding methods. The algorithms are trained on a diverse dataset encompassing various environmental conditions and cropping systems. This allows us to achieve a nuanced understanding of the spatial and temporal patterns associated with each commodity crop. The presentation will cover the following components:

1. Cacao, Oil Palm, Rubber, Coffee, Pasture, and Soy Mapping: The machine learning models are tailored to accurately classify land cover types, emphasizing the identified commodity crops.
2. Carbon Emissions Estimation: We integrate ancillary data, including climate variables, and land-use change, to estimate carbon emissions associated with land-use changes driven by the cultivation of commodity crops.
3. Spatial and Temporal Dynamics: Our analysis explores the spatial and temporal dynamics of commodity crop expansion, providing insights into patterns of deforestation, and land-use transitions.

We anticipate that our study will yield high-resolution maps depicting the spatial distribution of commodity crops across the tropics and associated carbon emissions linked to these commodity crops. The works aligns with recent European Union(EU) regulations to curb the EU market’s impact on global deforestation and provides valuable information for monitoring land use following deforestation, crucial for environmental initiatives and carbon neutrality goals.

References:

Masolele, R. N., Marcos, D., de Sy, N., Abu, I.-O., Verbesselt, J., Reiche, J., & Herold, M. (2024). Mapping the diversity of land uses following deforestation across Africa. Scientific Reports, 14, Article 1681. https://doi.org/10.1038/s41598-024-52138-9

Masolele, R. N., De Sy, V., Marcos, D., Verbesselt, J., Gieseke, F., Mulatu, K. A., Moges, Y., Sebrala, H., Martius, C., & Herold, M. (2022). Using high-resolution imagery and deep learning to classify land-use following deforestation: a case study in Ethiopia. GIScience & Remote Sensing, 59(1), 1446-1472. https://doi.org/10.1080/15481603.2022.2115619

Masolele, R. N., De Sy, V., Herold, M., Marcos, D., Verbesselt, J., Gieseke, F., Mullissa, A. G., & Martius, C. (2021). Spatial and temporal deep learning methods for deriving land-use following deforestation: A pan-tropical case study using Landsat time series. Remote Sensing of Environment, 264, Article 112600. https://doi.org/10.1016/j.rse.2021.112600

The Copernicus4GEOGLAM service was established in 2021 as part of the Copernicus Global Land Service (CGLS). It aims at supporting agricultural monitoring at national/sub-national level in countries with high food insecurity. The service can be activated upon request by a country to the GEOGLAM¹ secretariat and to the EC-Joint Research Centre (JRC). The service products consist of crop type maps and crop area statistics which complement other CGLS. Products are derived over an Area Of Interest (AOI) in the country during (in-season) and at the end of the growing season. The service allows to timely detect anomalies in the crop areas and supports agricultural and food security decision making. Crop type mapping is based on Sentinel-2 imagery using a random forest classifier that is trained and validated with data from dedicated field campaigns. Stratified random sampling is applied to identify the sample areas and ensure data are statistically sound and reflect the different crop classes in the AOI.

Since 2021 the service has been activated in Uganda, Kenya and Tanzania with AOIs ranging from about 100k km² to almost 250k km². Crop maps accuracy can reach 80%, but crop specific accuracy tends to vary, with omission errors usually higher than commission errors.

The Ministries of Agriculture relies on Copernicus4GEOGLAM products to complement existing crop statistics that are often obsolete or only available for few administrative areas. Besides, knowledge on crops spatial distribution is needed to implement fertilizers programmes at national level, crop insurance mechanisms and yield forecasts.

EC-JRC supports the use of these products in the countries and distributes the in-situ data to the community working on other LULC mapping programmes, like WorldCereal and the upcoming Copernicus Global Land Cover and Tropical Forest Mapping and Monitoring. Feedback from users helped JRC improve the quality of the in-situ datasets.

The Combat Against Climate Change on Cotton Communities (C5) feasibility study created a prototype agroclimatic advisory solution. This ESA EO Science for Society activity supports farmers and the cotton industry in Bangladesh by providing information on climatic issues affecting cotton-grower health.

The health-related output products tackle the physiological stress of heat exposure, and were split into forecast and historic components based on user needs. The forecast system was designed for farmers. It downloads forecast data, calculates the Heat Index, and generates district-level summary statistics and visualisations. Tailored health advisories are prepared based on this data, and all information is disseminated fortnightly in bulletins via agricultural extension workers. The infrastructure is capable of warning about dangerously hot and cold conditions.

The historic component was designed for industry officials and used custom-generated wet bulb globe temperature (WBGT) data. The prevalence of hazardous temperature events were calculated from long-term WBGTs. A “lost labour” dataset was also derived: work-rest guidelines recommend increased rest at higher WBGTs to reduce overheating risk. Applying work rest guidelines to WBGTs calculates the amount of time outdoor workers cannot work and stay safe – a metric with economic and social implications.

C5 datasets make use of I) up-to-date cotton crop area mapping to ensure they are linked to current production areas; and II) vegetation health derived from standardise LAI anomalies to assess current crop condition. The C5 cotton crop map utilises Sentinel-1 and Sentinel-2 data. Multi-temporal surface reflectance cloud-free composites for the dry and wet seasons in Bangladesh capture phenological changes in the crop associated with changes in canopy chlorophyll content. Sentinel-1 monthly VH backscatter composites characterise changes in canopy structure. These multitemporal composites are use as features to generate a Machine Learning model using cotton farm locations as training data, and result in an annual Bangladesh cotton map.

Intensification of agriculture is causing the decline of segetal flora species, with resulting negative ecological impacts, such as increased soil erosion and food and habitat loss for animals. One way to promote the conservation of such plant species is through result-based payment schemes that reward farmers based on observed biodiversity outcomes in their fields, but cost and time required for the monitoring hampers a more widespread implementation of such schemes. Automated monitoring of segetal flora species is particularly challenging due to their small sizes and partly overlapping spectral signatures. Using the latest advances in deep learning, we investigated the potential of UAVs for segetal flora species monitoring by focusing on an arable area in a UNESCO biosphere reserve in Saxony, Germany, and evaluated the usage of different UAV sensors to disentangle the different plant species. The presentation will focus on opportunities and challenges in segetal flora species monitoring via UAV with particular emphasis on: i) species for which training data can easily be developed from RGB images, ii) sensor and flight height maximizing the classification accuracy, iii) difficult to map species, and iv) potential for result-based payment schemes for other species that were not observed in our study area, but that are of interest for the implementation of such schemes in Europe.­

Accurate and spatially explicit information on irrigation is crucial for sustainable water resource management, crop condition monitoring, and modeling. Although progress has been made in irrigation mapping using remotely sensed data, field-level irrigation mapping in temperate regions remains challenging, as most studies have focused on semiarid regions. In this study, we assessed the applicability of different time series for irrigation mapping, utilizing Sentinel-2, Sentinel-1 time series, and Landsat-based Land Surface Temperature (LST) data over northern Germany. This area is characterized by heterogeneous field sizes, crop patterns, irrigation systems, and management practices. An extensive amount of field-scale irrigation data reported by farmers was collected and used as a reference for model training and validation. The derived Vegetation Indices (VIs), Tasselled Cap components, and LST were aggregated over the growing season and specific key phenological stages. Subsequently, irrigated areas were classified using Random Forest (RF) and gradient boosting (XGBoost)-based classifiers. The overall accuracy achieved satisfactory levels (approximately 80%). The performance varied across different regions and showed the significance of the availability of observations during the growing season, with the most important variables listed as LST, optical-based VIs, and Sentinel-1 based metrics for specific crops such as maize. The synergistic use of optical, radar, and LST data significantly enhanced the classification accuracy, demonstrating the potential of integrating these data sources for improved irrigation mapping in temperate regions. In combination with process-based agro-ecosystem models, such as the simulation model for nitrogen and carbon dynamics in agro-ecosystems (MONICA), a map of irrigated/non-irrigated fields not only allows for more accurate seasonal crop yield predictions for the respective area but also opens a path towards the quantification of water use for irrigation. This integration showcases the enhanced utility of the mapping for sustainable agriculture and water resource management.

With the outbreak of conflict in Ukraine, uncertainty in the amount of commodities available to the international market has greatly increased. Ukraine itself is a major exporter of grain to the world, accounting for 12% of wheat and 16% of maize sold on the commodity market in 2019. Additional uncertainty comes from potential changes in global trade relations with Russia, who also provide significant exports to the global markets. Understanding changes in cropped area in Ukraine due to ongoing active conflict, as well as changes in overall productivity due to supply chain constraints resulting from cessation of commercial shipping and transportation can be done with EO data. Amid continuing conflict, Ukraine’s agricultural sector could produce significantly less than previous years with global consequences due to these constraints. This project focuses on using EO tools to pair very large agricultural units with nearby small farms of 10 to 100 hectares. Targeting soy, sunflower, barley, wheat and maize, we generated in-season crop classification models and production estimates based on Sentinel imagery. We used these maps to target small farms with signficant yield gaps. Identifying willing large farming institutions, we will work with input supply and financing support from companies such as Syngenta to offer favorable terms to large and surrounding small farms to facilitate accelerated yield increases. Quatitative baselines from production maps will enable appropriate targeting of these inputs and monitor the success of the program.

In this study, we demonstrate a novel remote sensing approach for early- and in-season crop type mapping across diverse geographical contexts, utilizing multi-temporal Sentinel-2 imagery. The method used the analysis of time series Normalized Difference Vegetation Index (NDVI) data, integrated with a rule-based classification system that aligns with the specific phenological stages of various crops as per their respective agricultural calendars. This approach has been effectively applied by NSL Geospatial unit of FAO, in distinct regions such as Libya, for mapping crops like dates, olives, barley, and wheat, and in Myanmar, for identifying major agricultural categories including rice, maize, pulses, oilseed, sesame, sorghum, and others. By leveraging NDVI's sensitivity to phenological changes and synchronizing it with established crop calendars, our methodology eliminates the reliance on current-year ground truth data, enabling accurate, cost-effective, and timely crop classification. This approach not only demonstrates significant potential for enhancing agricultural monitoring and management globally but also exemplifies the adaptability of NDVI-based analysis in varied agricultural settings.

In this study, we demonstrate a novel remote sensing approach for early- and in-season crop type mapping across diverse geographical contexts, utilizing multi-temporal Sentinel-2 imagery. The method used the analysis of time series Normalized Difference Vegetation Index (NDVI) data, integrated with a rule-based classification system that aligns with the specific phenological stages of various crops as per their respective agricultural calendars. This approach has been effectively applied NSL Geospatial unit of FAO, in distinct regions such as Libya, for mapping crops like dates, olives, barley, and wheat, and in Myanmar, for identifying major agricultural categories including rice, maize, pulses, oilseed, sesame, sorghum, and others. By leveraging NDVI's sensitivity to phenological changes and synchronizing it with established crop calendars, our methodology eliminates the reliance on current-year ground truth data, enabling accurate, cost-effective, and timely crop classification. This approach not only demonstrates significant potential for enhancing agricultural monitoring and management globally but also exemplifies the adaptability of NDVI-based analysis in varied agricultural settings.

West Africa is facing two major challenges of the 21^st century: climate change and population growth. Both are closely linked to food security in the region. Rising temperatures and increasingly variable precipitation threaten traditional rain-fed agriculture relying on the rainy season. Furthermore, West Africa has one of the highest population growth rates in the world, its population will increase to 1.2 billion people by 2050. To guarantee sufficient food supply and to achieve the Sustainable Development Goals (SDG), a sustainable intensification of agriculture is needed (i.e., increasing yields without additional land consumption and without adverse effects on climate change) and mitigation and adaption strategies against the negative effects of climate change are required. Sustainable intensification (SI) practices offer the opportunity to stabilize/increase yields and to operate more resource-efficiently. The implementation of SI practices are tasks of the farmers but incentives must be created to accompany the process to enable investments or bridge short term losses. Therefore, a monitoring system is needed. The monitoring of the implemented SI practices is time consuming, costly and over large areas not feasible. Remote sensing has proven to be a suitable instrument to monitor agriculture area and the management strategies in a reliable and cost-effective way. This study in the Senegal River Valley shows the potential of remote sensing to feed a monitoring system on field scale. To address the lack of field data, we trained a convolutional neural network (CNN) to delineate field boundaries from Planet data and monitor agricultural management on field level. The agricultural management like sowing and harvesting dates or irrigation and flooding events are identified using change detection in Sentinel-1 time series. Those data can be used in Manobi Africa’s agCelerant platform rewarding farmers implementation of SI practices and to link farmers with financial institutes like banks and insurance companies.

Satellites capture an extensive amount of data daily, resulting in an ever-growing collection of Earth Observation (EO) data. Despite the availability of this data, there are still challenges when it comes to extracting relevant information from long time-series data streams. The recently finished ESA’s AI4FOOD project aimed to address these challenges, particularly in data fusion and advanced time series analytics within cropland monitoring applications.

AI4FOOD focused on advanced Machine Learning (ML) techniques to develop new algorithms for the creation of continuous optical and radar data streams. Specifically, the project focused on the fusion of Sentinel-2 and Sentinel-1 data and evaluating aspects such as the predictability of time series in dynamic land environments. This collaborative effort was achieved with partners from VITO, Sinergise, and the University of Valencia. The undertaken activities led to the development of the FuseTS toolbox – an open-source toolbox supporting users in complex data fusion and time series analytics tasks.

FuseTS is created based on the close collaboration between project stakeholders and partners. Using their requirements, state-of-the-art ML algorithms for data fusion and time series analytics were integrated into the final toolbox. The resulting Python library, available on GitHub, provides a solid foundation for data fusion and time series analytics. It offers essential data fusion, gap-filling, and smoothing services, such as Whittaker and Multi-Output Gaussian Process Regression (MOGPR). Additionally, FuseTS provides functions to extract valuable insights from the data fusion pipeline by detecting peaks and valleys and extracting vegetation phenology metrics. The FuseTS library enables the seamless execution of the offered algorithms on both local xarray data structures and through openEO, a community standard for EO processing. In this presentation, we will present the main functioning of FuseTS, as well as provide examples of cropland monitoring applications, such as fusion, gap-filling and start- and end-of-season detection.

Agriculture and food security face multiple challenges, at global scale, but also in Europe. Population growth and changing dietary habits lead to increasing demands for food production, and at the same time, agricultural products are required by the bioenergy market. Furthermore, climate change and the decline natural assets ask for sustainability and adaptation measures in agriculture. For supporting knowledge-based decisions in this context, Earth observation can contribute data and information products.

In this poster contribution, we present data products and workflows on crop type and rotation mapping as well as early crop type detection for Germany. Time series of Sentinel-2 and Sentinel-1 data are used in combination with LPIS (Land Parcel Identification System) reference data and machine learning techniques to retrospectively map seventeen crop types over Germany at an annual basis. From these multi-annual datasets, we delineate maps of crop rotation including pixel-based reliability information. Further, the potential of early within-season detection of individual crops and groups of crop types is systematically tested based on the classification chain and additional Sentinel time series analyses. The results show good potential for early detection of key crops (e.g. rapeseed and maize), but with a high variability in detection accuracies between the full range of the seventeen considered crop types. The presented data products will be used within mobile and web Apps developed by our partners (DLR-DW and University of Wuerzburg) for farmers and agricultural consultants within the project Agrisens DEMMIN 4.0 (funded by the German Ministry of Food and Agriculture).

Ukraine is a breadbasket cereals and oilseeds producer and exporter. In 2021, Ukrainian farmers planted winter crops, ignoring that a full-scale invasion by Russian forces would start a few months later. Consecutive to the invasion on February 24th, 2022, the occupant took control of about one third of the country, including some of the most productive agricultural regions. Immediate concern around how much winter and summer crops would be planted arose. This work details how the NASA Harvest team for Rapid Agricultural Assessments for Policy Support (RAAPS), (a) leveraged bi-weekly three-meter spatial resolution Planet composites for building in-season crop type maps for 2022 and 2023 and (b) estimated unbiased planted areas both for free and occupied territories.

Based on clustering approaches and domain knowledge, cropland was split into winter crops and potential summer crops. Then, winter cereals and rapeseed were separated. Finally, residual cropland was either classified as summer crop or as fallow land.

Unbiased area and accuracy estimates resulted in overall accuracies of 90+-3% in free territories and 81 +- 4% in occupied territories in 2022. In 2023, overall accuracies were 88+-2% for free territories and 77+-3% for occupied territories. As of mid-July 2022, 20.5% of Ukraine's cropland was under occupation against 16.03% at the same time in 2023. A detailed assessment of changes in planted areas per crop type, between 2022 and 2023, will be presented at the workshop. This work demonstrated the importance of coupling remote sensing and domain knowledge for mapping major crops and deriving statistical information, in situations where no ground data is available.

The Democratic Republic of Congo (DRC) is the fourth poorest country globally and has a rapidly increasing population. Congolese livelihood depends on small-holder farmers production. The historically unstable Masisi and Rutshuru regions in Eastern DRC, have seen a resurgence of conflict since the end of 2021. Civilian people displacements have been observed since then. The NASA Harvest team for Rapid Agricultural Assessments for Policy Support has investigated whether people displacement leads to cropland abandonment, which might in turn lead to increased food insecurity.

Given cropland abandonment is defined by previously active cropland being left non-cultivated, yearly active cropland maps had to be produced and inter-compared. Leveraging Planet three-meter spatial resolution and bi-weekly temporal resolution composites, yearly season B (February to July) active cropland maps were generated from 2021 to 2023. For each year, training data was collected in a pseudo random manner. First, a 10*10 km grid covering the Masisi and Rutshuru regions was generated. Then, a random point was placed within each grid cell. Cropland and non-cropland samples were collected in the vicinity of each grid point, through satellite image time series annotation. Active cropland was mapped using a random forest classifier. Time series of Normalized Difference Vegetation Index, Enhanced Vegetation Index and Soil Brightness Index were used as independent variables. In order to avoid confusions between cropland abandonment and common fallow practices, active cropland proportion was computed on a 500*500 meter grid. The assumption was that, at this level, the proportion of active cropland should be stable from year to year, regardless of fallow practices.

Preliminary results seem to indicate no significant trend towards cropland abandonment in the vicinity of conflict hotspots. However, by the time of submitting this abstract, this is still a work in progress. Statistically conclusive results will be shared at the workshop.

Understanding the effects of different compost-derived organomineral fertilizers

enriched with nitrogen (N) and phosphorus (P) treatments on wheat yield and

protein content is essential for optimizing agricultural productivity while ensuring

nutrient-rich products with low environmental impact. The emergence of

commercial satellite constellations, due to their high spatial and temporal

resolution, provides new opportunities for monitoring and forecasting crop yields,

supporting informed and timely decisions to improve food security.

Our study aims to explore the use of PlanetScope imagery to retrieve wheat yield

and protein content in response to pelletized organomineral fertilizers applications.

Eleven field plots (24 m2) were treated with different P (at sowing) and N (at

tillering) fertilizations strategies, from conventional to organic, with three

replications (n=33) in the EEA Aula Dei CSIC (Zaragoza, Spain) in 2023. Field

measurements of N uptake efficiency were made during the crop cycle. At harvest,

yield and protein were measured for each plot.

Using Planet Scope, we detected treatment differences in the spectral bands and

obtained phenological metrics with error in less than 8 days. Distinct separability

was found after the N application in the NIR band and using NDVI. Plots treated with

organomineral fertilization showed higher yields (p < 0.001) and protein content.

We calibrated the Agriculture Remotely-sensed Yield Algorithm (ARYA) model to

examine its capabilities of yield predictions at high resolution, and a generalized

linear model to forecast protein content. The results showed a linear correlation of

(X, Y R2) and A, B RMSE per yield and protein content, respectively.

This study demonstrates the utility of Planet Scope for monitoring wheat yield under

different treatments and improving nutrient management strategies for sustainable

agriculture aligned with the European Green Deal objectives.

According to the Food and Agriculture Organization (FAO), agriculture accounts for 70 percent of global freshwater withdrawals. As agricultural water management is key for sustaining food security, accurate soil moisture monitoring is crucial for mitigating the impacts of drought on crop production.

The satellite Sentinel-1 with the active microwave remote sensing Synthetic Aperture Radar (SAR) imaging has emerged as an effective tool to estimate surface soil moisture. This EO technology shows great potential for high temporal and spatial resolution soil moisture monitoring across agricultural landscapes. Therefore, the is an amazing opportunity for mapping the spatial and temporal soil moisture dynamics.

However, for data calibration and validation, ground-truth measurement of soil moisture is required ensuring the accuracy and reliability of remote sensing-derived soil moisture estimates.

There are different techniques for in-situ soil moisture estimation, but they are mostly at the point scale, which can make it challenging to use them for remote sensing calibration and validation.

Recently, the Cosmic Ray Neutron Sensor (CRNS) which is an in situ nuclear technology shows the capability to estimate field-scale soil moisture in large areas of up to 20 ha and has demonstrated its ability to support agricultural water management. The need for in-situ data, such as that provided by CRNS lies in its ability to offer ground-truth measurements of soil moisture at landscape level.

The integration of in-situ CRNS soil moisture and EO data facilitates not only the validation and calibration of remote sensing imagery but also provides real-time soil moisture information valuable for implementing climate-smart irrigation strategies, contributing to a sustainable food production initiative.

This approach represents a significant advancement in soil moisture monitoring by combining remote sensing with nuclear technology.

Advances in field-based plant phenotyping, ranging from low-cost handheld devices to extensive satellite imagery, are opening new avenues for understanding and optimising plant responses to environmental factors. The scientific challenge is to scale-up such observational capabilities from in-orbit systems. The present study aims to characterise the within-field spatial and temporal variability of environmental conditions affecting crop growth. After pre-processing the Sentinel-2 data with the open source system Sen4Stat, the Green Area Index (GAI) was retrieved by inversion of a radiative transfer model, i.e. a locally tuned BV-Net algorithm. For each year of Sentinel-2 acquisition (2017-2023) and each parcel of the study area (Wallonia), the GAI time series was used to (1) infer the growing season of the main crop and (2) produce a spatial indicator of vegetation growth heterogeneity based on this in-season time series. The resulting maps effectively capture the homogeneity or heterogeneity observed in the GAI profiles. The evaluation of the seasonal maps included an analysis of agronomic factors such as crop rotation, agrometeorological data, soil maps and slope, which helped to interpret the sources of heterogeneity within fields. Despite significant variation in the seasonal maps between years, discernible patterns emerged, highlighting similarities in conditions or crops between years. An important finding is that certain fields, identified as spatially homogeneous on the basis of soil characteristics, exhibit heterogeneity in vegetation growth. Conversely, fields that appear to have strong spatial heterogeneity based on the soil map may either have a fair degree of homogeneity in vegetation growth, or a pattern of heterogeneity that differs from the soil map. The versatility of the method extends its applicability to different agricultural settings and to any type of crop. The resulting maps could guide dynamic agricultural practices towards greater sustainability, including irrigation, fertilisation and spraying management, and could also provide opportunities for new soil sampling designs and targeted in-field phenotyping.

Monitoring and preserving biodiversity in mountain agroecosystems are critically important because these systems provide several services such as carbon sequestration, food and wood supply, as well as offering habitat diversity for different species. For example, extensively managed alpine pastures in South Tyrol, Italy, host rare and endemic plant species and represent one of the few examples of traditional management in the area. Remote sensing techniques can provide useful information on biodiversity over large areas. In this context, we aim to explore the promising potential of the hyperspectral sensor on the Italian Space Agency's PRISMA mission, providing data in 239 spectral bands.

In this study, we exploit PRISMA images to estimate species diversity in the grasslands of the Sciliar Natural Park in South Tyrol (Italy). Specifically, we verify the existence of a direct relationship between remotely sensed reflectance values and species diversity derived from field data by using the Spectral Variation Hypothesis. According to this hypothesis, areas showing pronounced spectral variation in an image are often indicative of high environmental heterogeneity, thus serving as a powerful indicator to estimate species diversity. A field data collection campaign was carried out during summer 2023 to quantify species diversity indices, e.g., the Shannon Diversity Index and the Pielou´s Evenness Index. From two PRISMA images, acquired during the summer and autumn of 2023, spectral diversity indices like the Rao's Q Index were calculated. To verify the relationship between in situ and remotely sensed data, we compare the satellite-based spectral variation with the ground-based species diversity by regression analysis. Results provide interesting insights into the strengths of PRISMA hyperspectral data, such as spectral resolution, and their limitations, such as low spatial resolution and availability of images.

Companies are increasingly under pressure to address their environmental and social impacts, including in their supply chain. Many stakeholders, including customers, investors and regulators are demanding greater transparency and accountability in regards to environmental performance. In this context, monitoring the environmental impacts of a company’s supply chain is essential for ensuring compliance, reducing risk and enhancing sustainability. Given the urgent need for companies to take action to evaluate, plan and mitigate environmental impacts, LandGriffon fills an essential gap enabling companies to act in environments of limited information.

LandGriffon, developed under a Horizon Europe project, is inspired by the need to move beyond life-cycle assessment approaches and provide robust spatially explicit information on agricultural supply chain impacts. It addresses the challenge of a lack of traceability by providing a framework for companies to spatialize agricultural supply chain knowledge and evaluate impacts across a range of indicators, including, land footprint, greenhouse gas emissions, land conversion, water use and pollution and biodiversity, as accurately as possible.

In this presentation, we will introduce the Landgriffon methodology (https://landgriffon.com/) and demonstrate its application to company supply chain data to align with forthcoming regulations and commitments, including EU deforestation regulation, Science Based Targets Network and the Taskforce for Nature Related Financial Disclosures. However, given the scale and complexity of agricultural supply chains, there are considerable uncertainties and limitations associated with data and methods. We will also highlight these gaps and aim to stimulate a community of practice to improve our capabilities in a coordinated way, because collaboration and openness will be critical to achieving real improvements in the sustainability of agricultural supply chains. We hope that LandGriffon can be a tool around which this could be achieved and so help drive more positive futures for society and nature.

Consistent Rice Information for Sustainable Policy (CRISP) is a 2-years ESA funded project that aims to address Indicator SDG 2.4.1, which measures the proportion of agricultural land area under productive and sustainable agriculture. This initiative, in collaboration with FAO and a selected group of Early Adopters (EA), endeavors to contribute in the achievement of sustainable food production systems and resilient agricultural practices by 2030.

CRISP focuses on scaling up advanced and cost-effective Earth Observation (EO) solutions to provide crucial information on seasonal rice planted area, growing conditions, yield forecast, and production at harvest. The project adopts a user-oriented approach, recognizing the importance of active users' involvement in introducing and understanding the proposed solutions.

At this stage, needs have already been collected and translated into the most valuable products for EAs. Similarly, most relevant Test Sites were identified and the focus is in addressing the best practices in EO technology. CRISP leverages existing operational rice area-yield services, such as RIICE (Remote Sensing based Information and Insurance for Crops in emerging Economies) to serve as a foundation for the solution development that once operational, will offer a comprehensive suite of tools designed to facilitate the generation of products on a global scale.

This approach of active involvement not only educates Early Adopters on the capabilities and limitations of the proposed solutions but also ensures to meet realistic and feasible requirements, leading to successful service endorsement. The CRISP Project leverages EO Platform as a Service cloud native technologies provided by the CGI Insula platform to address large processing in a cost effective manner.

CRISP's methodology involves a thorough review of EO best practices, experimental algorithm evaluation, and the use of multi-mission EO systems, including Sentinel-1, Sentinel-2, PlanetScope, and forthcoming NISAR, to ensure the provision of a robust and scalable EO solution.

-

<p>Dragonfly is an innovative farming application that offers farmers a comprehensive set of tools to manage their farms effectively, starting from seeding to harvesting. Through the processing and analysis of satellite imagery, Dragonfly provides farmers with insights and information that allow them to optimize their use of nitrogen fertilizer in plots. With Dragonfly, farmers can reduce costs, increase yields, and promote sustainability by making the best use of their resources.</p>

The world is changing due to global warming and the need for accurate worldwide temperature monitoring has never been more important.

LST (Land Surface Temperature) plays a decisive role not only when it comes to an optimized irrigation management within the agricultural industry, as the world faces increasing water scarcity and the need for food security. Implementing sustainable water management and monitoring plant stress and health is also crucial within the forest sector as the demand for wood and paper continues to grow. Not only are both agriculture and forestry challenged by water scarcity, but the risk of wildfires is rising with global warming, posing a fundamental threat to both. By measuring LST and analysing further processed data such as ET (Evapotranspiration), more insight can be gained with regard to employing predictive analytics for irrigation scheduling, yield estimation, frost and heat event prediction.

To provide the necessary information and therefore be able to act and react, OroraTech has already 2 working sensors in orbit, delivering data reliably (FOREST-I and II). It is equipped with a single Mid-Wave Infrared (MWIR) and two Long-Wave Infrared (LWIR) sensors that scan our planet every day. To increase the revisit rate and overcome the difficulties of cloud cover, OroraTech's ambitious schedule foresees 8 satellites operating in orbit by 2025 (12 hour revisit) and up to 100 sensors by 2027 (30 minute revisit) in lower earth orbit.

From 2025, OroraTech's constellation will be able to monitor the diurnal cycle of temperature, providing valuable data on day and night temperature fluctuations that are crucial for assessing plant stress. The satellite's on-board processing capabilities, which allow for real-time data analysis, will enable even faster responses to detected anomalies.

With a swath of 410 km, large areas can be covered at the same time. This and a GSD of 200m (super resolution 70m) will provide the capability to detect timely critical land surface temperature changes, complementing larger missions with data that fill the gap between Trishna's or LSTMs overpasses.