Conference Agenda

Climate change is a challenge for all sectors, but especially for agriculture. Rising temperatures in Africa negatively affect agricultural production. In addition, the demand for food is increasing, making it necessary to develop variables that describe agriculture, such as yield forecasts or phenology development stages of the different varieties cultivated. Remote sensing provides spatial and temporal continuous information that allow us to accurately assess the evolution of a particular field. In this study, we use Sentinel-2 optical data to analyse rice and wheat seasons between 2016-2022 in the Gharbia governorate (Egypt) where ground data were collected by the University of Tanta. We applied a yield model for each crop type trained in Spain and study their transferability to this region. In the case of wheat, we also test a second model that aggregates SAR data from Sentinel-1, thus evaluating the fusion of both products. We also studied the integration of the accumulated Growing Degree Days (GDDaccum), since the seasons do not start on the same dates and depending on the temperature accumulation the crops phenology evolves differently. Using the GDDaccum, the models show a better performance that when considering the timeseries evolution against the dates. The optical models provide a RMSE of about 1 T/ha for rice and 0,9 T/ha for wheat. The optical + SAR model manages to reduce the wheat model error to 0,7 T/ha. We also developed a crop type mask to evaluate the yield models performance at governorate level. Preliminary results show a RMSE of 2 T/ha for rice and wheat.

Climate change is threatening food security. To ensure food security, we do not only have to safeguard agricultural production but also optimally distribute the yields between regions. For that, decision-makers need reliable crop yield forecasts so that they can plan which regions are likely to experience crop yield losses and which regions will produce a surplus. Earth observation and machine learning are key tools to calculate such forecasts. Especially Sentinel-1 and 2 data has been used a lot as it provides regular high-resolution information about the state of crops and soil moisture. However, crop yield forecasts based on machine learning are strongly limited by the availability of field-level crop yield data, which farmers often do not like to share publicly. In this study, we evaluated if a model trained with data from a certain region can be applied elsewhere, to use training data more efficiently. Our field-level crop yield forecasts were trained using crop yield data from a farm (846 fields) in the Czech Republic for winter wheat. It was based on Sentinel-1 and 2 data and the machine learning model Extreme Gradient Boosting. The model was then tested for various farms with increasing geographical distance. The baseline was a forecast for fields of the same farm, that were not used for training. Next, the model was applied to another farm in Czechia, one in Ukraine and one in the Netherlands. The model transferability worked well for the other farm in Czechia (R²=0.64 between the forecast 1 month before harvest and observed yield). However, the model performed poorly further away than that (R²<0.13). This was related to very different weather conditions. Adding meteorological predictors or applying the model to more similar areas may help in the future to improve the transferability of the forecasts.

In this work we develop a model to forecast wheat yield at field level from Sentinel-1 (S1) and Sentinel-2 (S2) data. To do this, we calibrate the model with precise yield data acquired between 2018 and 2022 in different fields located in an agricultural region (mostly cereals) in the province of Valladolid and we validate it with data acquired across different provinces of across Spain. To minimize the temporal variability that each season and region may present, we normalized the signal from the two satellites based on the plant temperature accumulation (accumulation of Growing Degree Days, GDD_acum) based on the ERA5 air temperature product. Additionally, we test different Start-Of-Season inputs to start the GDD accumulation: a) the average planting date, b) the WorldCereal crop calendars, or c) Land Surface Phenology (LSP) metrics based on MODIS daily data. Finally, we calibrate the yield model based on a two-parameter linear regression for each GDD_acumrange considering three possible combinations: (1) using only optical parameters of S2, (2) using only microwave parameters from S1, or (3) using one parameter from S2 and one parameter from S1. The calibration results show that with S2 the yield can be estimated with errors of 800 kg/ha around 1400 ºC, S1 provides errors of 1000 kg/ha around 1700 ºC and the integration of S1 and S2 provides the lowest errors of 700 kg/ha around 1400 ºC. We tested the transferability of the model across the major wheat producing Autonomous Communities in Spain (Castilla y Leon, Castilla La Mancha, Aragon, Andalucia and Cataluña). The results show consistent results with the calibration data with minimum seasonal errors ranging from the lowest of 700 kg/ha in Andalucia to the largest 1200 kg/ha in Aragon.

The escalating frequency and intensity of climate shocks underscore the critical necessity of timely and precise data on agricultural conditions to inform responses. This data is often absent in countries where rainfed, low-input and smallholder farming systems are ubiquitous, conditions that heighten vulnerabilities to climate risks. Earth observation (EO) data can fill these gaps and support in-season agricultural monitoring, providing frequent and transparent information to guide government interventions. In smallholder agricultural systems however, there is a need to validate EO data, and build confidence in new data sources. FAO and NASA Harvest collaborated to bridge data and analytical gaps in Malawi and Namibia, binding EO applications, digital tools and field-level assessments to foster an integrated approach to agricultural monitoring. The collaboration focused on three areas: 1) mobile-based tools for the collection in situ data; 2) cropland mapping of small-holder agricultural systems; and 3) yield forecasts based on earth observation data. The creation of a suite of mobile-based survey tools to collect ground truth data was a critical element of this work, with the primary intention to provide actionable information to governments, but with secondary benefits of building an extensive validation and training data set for yield forecast and cropland mapping models. The tools built off existing developments from ArcGIS Survey123, adding new components that sped-up and lowered costs of collecting field level data, including field boundaries and yield estimates, whilst importantly developing features to provide information back to the farmer; a critical aspect to foster trust with farmers. This integrated approach provided government entities with multiple sources of corroborative evidence (field level data, early yield estimates and interpretable EO analysis) to facilitate a more informed and timely response.

Since ground-based statistical methods of yield estimation are currently not available in the Ukraine due to the war, VISTA has provided the Ukrainian Ministry of Agriculture with high resolution information on expected production volumes for winter wheat, rapeseed, winter and spring barley, grain maize and sunflower, i.e. the YPSILON service, in 2022 and 2023. The crop types are classified using Sentinel-2 data. The temporal development of the leaf area is calculated for hundreds of thousands of fields and assimilated into the PROMET crop growth model. The large number of analyzed fields enables an aggregation of yield forecasts in t/ha on different administrative levels. Through combining area under cultivation and simulated yield from PROMET, forecasts can be made for the expected production volume. Coherence information from Sentinel-1 data is used for harvest detection on field level, to determine the share of the harvested area of the total cultivated area. The results for 2023 will be shown, illustrating the resilience of the Ukrainian farmers in terms of food production as well as the consequences of the front line for agricultural production.

In the current workflow as employed for these analyses, forecasts of yield and production are possible up to 6 weeks in advance. Within the Horizon Europe Project STELAR, VISTA aims at making these forecasts possible even earlier, up to 10 weeks in advance. For this, sophisticated tools for data imputation and data fusion are being developed and integrated in STELAR’s Knowledge Lake Management System. A first outlook towards these advances will be given.

The work for Ukraine presented here has been supported by co-funding from ESA’s Network of Resources under contract numbers 3a06VS and 3717ds, as well as from BayWa AG. STELAR is funded by the European Commission’s Horizon Europe program under grant agreement number 101070122.

Precise within-field estimation of cropland productivity is crucial for informed agricultural decision-making, particularly in enabling the optimization of management practices and the pre-harvest anticipation of crop yields. The current availability of high spatiotemporal resolution of optical satellite data offers the opportunity for continuous monitoring and assessment of croplands, allowing for enhanced agricultural productivity and resource utilization. This study presents a workflow for predicting within-field grain yields focusing on winter cereal crops in Switzerland (wheat, barley and triticale). NDVI and the recently introduced kernel NDVI time series were derived from Sentinel-2 data as descriptive indicators of crop status and evolution. To ensure temporal continuity, Gaussian Process Regression (GPR) was used as a curve-fitting function to reconstruct cloud-free time series throughout the growing season. The performance of various machine learning methods (GPR, Kernel Ridge Regression, and Random Forest) to forecast yield at any point in time during the season was compared. The integration of Growing Degree Days (GDD) information as the temporal spacing reference of the time series considerably improved the accuracy and consistency of in-season yield forecasting. Using data from the same year, results indicate that grain cereal yield can be reliably predicted approximately 2-2.5 months before harvest, with an RMSE of up to 0.71 t/ha and a relative RMSE of 7.60%. Although the forecasting accuracy decreases when predicting for unseen years, the results remain satisfactory (RMSE = 0.97 t/ha, relative RMSE = 11.47%). These findings showcase the potential of the proposed workflow for in-field yield monitoring and targeted interventions, potentially reducing yield losses, optimizing farmers' management planning and enhancing food availability.

.