Conference Agenda

Ecosystem resilience represents the capacity to withstand and recover from perturbations. Resilience is a fundamental functional property of forests, especially in view of increasing anthropogenic and climate pressures. The focus of recent large-scale resilience studies has been on the relationship between resilience and climate, with little exploration on the relationship between forest resilience and its diversity, upon which management practices can have an impact. In this study, the sensitivity of European forest resilience to structural diversity is quantified using remotely-sensed data. Two established resilience indicators are extracted from MODIS derived kNDVI time series, and forest structural diversity is accounted for by horizontal, vertical and combined horizontal and vertical metrics derived from NASA’s GEDI instrument. A Random Forest model is leveraged to isolate the interplay between resilience and structural diversity and to disentangle confounding environmental variables such as background climatic conditions. The study finds that European forests with a higher level of structural diversity are systematically associated with higher resilience levels. Importantly, diversity in canopy complexity is more important for resilient forests than variability in canopy height, and this relationship is consistent under increasing temperature patterns. This suggests that forest management promoting forest heterogeneity and especially canopy complexity has the potential to offset the decline in forest resilience associated with climate warming.

Monitoring Biodiversity Change is essential for planning and tracking the effectiveness of conservation initiatives aligned with international agreements.

Satellite remote sensing enables monitoring Biodiversity at scales relevant to conservation by providing continuous and repeatable ecosystem observations. Current frameworks primarily focus on mapping binary metrics, like changes in forest extent, which alone misses the impact of global (climate, air pollution) and small-scale degradation on the integrity and resilience of forests’ biodiversity.

This work introduces a monitoring framework to directly track Biodiversity Integrity Change, where Biodiversity Integrity is defined by the degree to which a forest composition, structure and function fall within a dynamic range of reference states that account for seasonal phenology and multi-annual resilience to past stressors such as drought.

Our approach uses Artificial Intelligence (AI) models that integrate multi-sensor satellite time series, including Imaging Spectroscopy, RADAR and LiDAR, capturing changes in composition, structure and function. These AI models analyze spatiotemporal patterns to understand seasonal and multi-annual variability at multiple spatial and temporal scales, and it pinpoints to forests that are deviating from expected phenological, structural or functional reference states.

Our AI-driven analyses enhance existing forest extent monitoring systems, by directly observing Biodiversity Change (structure, composition and function) within and between ecosystems, which is essential to plan and monitor progress towards achieving area-based conservation targets, as well as to understand the main threads to Biodiversity Integrity (e.g. land use conversion, climate change and extremes, air pollution).

We present preliminary findings from tracking changes in forests across the US Pacific region (California, Oregon, and Washington) from 2015 to 2024. We evaluate the effectiveness of various AI models, including novel models developed by our team that integrate Landsat and Sentinel-1 data, as well as Foundation Models (NASA’s Prithvi and PRESTO) which leverage multi-sensor satellite time-series to analyze spatiotemporal patterns.

Tropical rainforests face significant challenges, with nearly 40% of the remaining areas considered disturbed or degraded. This degradation has profound implications for both climate change mitigation and biodiversity conservation efforts. The carbon emissions resulting from tropical forest degradation are substantial, sometimes even surpassing those from deforestation in certain regions, though estimates vary widely. Beyond carbon concerns, degradation-induced changes in forest structure can significantly impact ecosystem function and integrity. These alterations can lead to reduced biodiversity and diminished ecological services provided by healthy rainforests. Recent advancements in satellite remote sensing technology have made it possible to detect even minor disturbance events in tropical forests. However, the full impact of these disturbances remains poorly understood, highlighting a critical gap in our knowledge of forest ecosystem dynamics. The Kunming-Montreal biodiversity framework, while ambitious, is hampered by a lack of high-quality indicators for ecosystem integrity. This deficiency makes it challenging to effectively monitor and assess the impacts of human-induced degradation on forest ecosystems. To address this issue, a novel approach has been developed to evaluate the impacts of disturbance events on tropical forests. This method compares near-coincident GEDI (Global Ecosystem Dynamics Investigation) shots that happen to sample forest carbon and structure before and after disturbance events detected by other remote sensing systems, providing valuable insights into changes. We show how this technique can be applied across the wet tropics to assess the impacts of various types of disturbances on forest ecosystems. By quantifying these effects, we aim to better understand the consequences of degradation and inform more effective conservation and restoration strategies for tropical rainforests

Conservation of forest ecosystems is essential for maintaining biodiversity and ecosystem services. This study leverages Earth Observation (EO) data to address global biodiversity monitoring challenges in the face of increasing natural and anthropogenic disturbances. Focusing on the Ticino Park, a temperate mixed forest in northern Italy, we investigated the impact of drought—an escalating stressor on Earth system functioning—by analysing Sentinel-2 imagery from 2017 to 2022, particularly during the severe drought of 2022.

To enhance the detection of plant water stress, we conducted direct and continuous monitoring of functional traits that indicate tree health and structural status in relation to drought conditions. Specifically, we derived high-temporal-resolution time series of leaf area index (LAI), canopy chlorophyll content (CCC), and canopy water content (CWC) from Sentinel 2. We also analysed forest environmental characteristics and species composition to assess their influence on physiological responses and corresponding spectral changes observed by EO satellites.

Our results showed strong correlations between Sentinel 2 -derived plant traits and ground measurements, with CCC having the highest correlation with ground data (r² = 0.82, nRMSE = 13.56%) and LAI closely following (r² = 0.75, nRMSE = 11.49%). Daily standardized anomaly (DSA) analysis highlighted significant variations linked to forest types, showing that pine and black cherry experienced the greatest stress, while hygrophilic species such as black alder and chestnut were less affected. The DSA maps provided spatial and temporal patterns related to drought-induced vegetation stress in the Ticino forests in 2022.

Our workflow provides quantitative insights into forest functional states at high spatial resolution (10 m), crucial for effective management and conservation measures. These findings highlight the importance of understanding species-specific responses to drought for improved forest monitoring and management strategies in response to climate-induced challenges.

Foliar functional traits are dynamic plant properties that vary across space and time, serving as principal tools for monitoring plant physiology and terrestrial ecosystem processes. Phenolics are the most crucial secondary metabolites that play key roles in plant defence against biotic and abiotic stressors, leaf decomposition, as well as consequent influence on nutrient cycling and soil microbial composition. However, spatially continuous information on canopy phenolic remains poorly characterized at the landscape level. Current and proposed spaceborne imaging spectrometers offer unique opportunities to map foliar phenolics quantitatively through space and time. Our recent work (Xie et al, 2024) demonstrated that foliar phenolics can be accurately estimated across temperate tree species using leaf spectroscopy. In this study, we leveraged imaging spectroscopy data from PRecursore IperSpettrale della Missione Applicativa (PRISMA) mission to predict and map foliar phenolic variations at canopy scale in a mixed European temperate forest. Two data-driven approaches, namely partial least square regression and Gaussian processes regression, were applied to link lab-measured phenolic concentration with PRISMA plot-level spectra (400–2400 nm). The performance statistics indicated reasonable precision and accuracy of the model results. Maps derived from the best-performing model (based on cross-validated nRMSE) provided a wall-to-wall assessment of canopy phenolics, capturing both inter and intra-species variations across the landscape. Further, we compared the phenol map with the distribution of leaf mass per area and canopy nitrogen. The results indicated that the synergy patterns across the three functional traits were consistent with the known leaf economic spectrum. These findings highlight the potential of spaceborne imaging spectroscopy to characterize spatial and temporal dynamics of ecologically important plant phenolics. Our study also paves the way for improved global monitoring of ecosystem integrity and plant responses to environmental stress and climate change, particularly with the anticipated launch of hyperspectral missions like ESA’s CHIME and NASA’s SBG.

Accurate in-situ biodiversity estimates are crucial for effective conservation strategies and rely on the integration of satellite remote sensing (SRS) data with on-the-ground measurements. Recent advancements in SRS technology enable high-resolution, near real-time observations of land use-land cover (LULC) changes. Model-based indicators, such as the Biodiversity Intactness Index (BII) and GLOBIO, are designed to translate such dynamics into estimated changes in biodiversity intactness of ecosystems. However, existing indicators are subject to some important limitations, including lack of evaluation of predictive performance against observed data, reliance on a relatively small fraction of available biodiversity data, and not integrating potentially important SRS data products.

In this project, we particularly address the lack of model performance testing, deep diving into different evaluation strategies for large-scale intactness models. We explore several cross-validation approaches, from standard random sampling to spatial, environmental, and cross-study alternatives. Using these approaches, we estimate the predictive performance of the BII model for relative species abundance. While the in-sample accuracy is high, predictive capabilities do not generalize to unseen, out-of-sample data, which is driven by the structure of the model. To improve generalization, we develop a Bayesian hierarchical model pipeline, with a hierarchical structure based on biogeographical entities. This model also includes a richer set of environmental predictors. While the Bayesian model performs significantly better in standard cross-validation, it struggles considerably when train-test splits are done across spatial, environmental and study dimensions.

These results highlight that the prospect of building good broad-scale predictive models is currently very challenging due to data limitations. This especially concerns the lack of at-scale, representative biodiversity inventories for many parts of the world and many taxonomic groups. We outline some potential paths forward to improve predictive models in this space, including environmental DNA for large-scale sampling and the need for more historical, high-resolution SRS products.

Here, we pioneer the use of multi-sensor Earth Observation (EO) data and insect in situ data collated from various "big data" platforms (iNaturalist, GBIF, and GenBank) to develop a framework for measuring insect-based biodiversity intactness patterns across Africa. The insect taxa used in this framework are sensitive to ecological changes stemming from unsustainable farming practices, urbanization, and logging. Insect diversity patterns have proven to be valuable indicators of overall ecosystem biodiversity intactness. Compared to megafauna, insects occur in all climate zones and occupy diverse micro-habitats, making them excellent predictors of ecosystem intactness at spatially explicit scales, even over larger regions.

The UN Convention on Biological Diversity (CBD) and its technical working group for the post-2020 framework have called for unbiased (i.e., accurate), measurable, and scalable frameworks and indicators for biodiversity. These frameworks should ideally account for localized drivers of biodiversity loss, support the estimation of planetary boundaries, and assess ecosystems' capacity to deliver services (such as pollination by insects). The UN Kunming-Montreal Global Biodiversity Framework likewise emphasizes the need to connect biodiversity loss with ecosystem services and focuses specifically on the integrity of agro-ecological landscapes.

We estimated biodiversity intactness as the ratio between the actual (or currently observed) insect diversity (o) and the historic or potential estimated insect diversity (p). Predictors included spectral features from 10-20m Sentinel-2 satellite data, 1-km WorldClim climate variables, 25-m tree heights from the Global Ecosystem Dynamics Investigation sensor, and 1-km human footprint data. Pixel-based biodiversity intactness predictions could be aggregated at the country level or across conservation priority corridors. Across Africa, high insect-based biodiversity intactness was observed in natural tropical forests, montane "sky islands," wetlands, islands in Lake Victoria, and arid countries such as Namibia. The framework can be adapted to focus on locally threatened or endemic insect species by analyzing individual species within the assemblage. The indicator values remain stable across diverse climate zones, and pixel-level data can be spatially aggregated to support country-level reporting mechanisms.

Landscape change and habitat fragmentation are recognised drivers of biodiversity change, but properly isolating and assessing their impacts can be challenging without appropriate data and statistical techniques. Indeed, the spatial scales at which they occur are often confounded by other stressors such as anthropogenic pressures or climate change. Here, we combine remotely sensed land cover time series with a recent technique from causal inference, synthetic controls [1,2], to test the impact of landscape modification on French breeding bird diversity metrics (Temporal Monitoring of Common Birds, STOC programme, 2001-2019). This method requires time series of treated and untreated units and a date of treatment assignment. This date is detected from annual land cover products [3]. By constructing appropriate controls, variations in STOC metrics can finally be attributed to landscape changes.

In parallel, foundation models trained on remotely sensed imagery (RSFMs) offer unprecedented predictive accuracy and generalisation power for downstream tasks such as biodiversity metric estimation, without requiring tailored training and precise understanding of the ecological processes at play. Therefore, a second objective is to analyse RSFM predictions based on annual Landsat imagery mosaics (RGB, NIR, SWIR bands) centered on altered landscape plots identified by synthetic controls: Do the deep learning models detect and rely on the same structural changes that have disentangled effects on biodiversity metrics, or do they miss these elements? The results of this cross-analysis between causal effect estimation on the one hand, and interpretation of deep learning predictions on the other, has the potential to increase understanding, confidence, or possibly caution in the adoption of foundation models for biodiversity modelling.

[1] Abadie, Alberto. Journal of economic literature 59.2 (2021): 391-425.

[2] Fick, Stephen E., et al. Ecological Applications 31.3 (2021): e02264.

[3] Zhang, Xiao, et al. Earth System Science Data 16.3 (2024): 1353-1381.