Forest Assembly and Species Composition with AI and Earth Observation Data, a scalable approach

An accurate spatial distribution of forest species composition is essential for biodiversity monitoring, management and protection. Combining this information with field structural metrics (e.g., basal area, species co-occurrences, canopy height) significantly improves our ability to estimate ecosystem functions and understand their relationship with biodiversity. These insights are crucial for regional biodiversity assessments, territorial planning, and forest management, contributing directly to nature conservation. A geospatial approach is particularly valuable when studying forest biodiversity dynamics, as it allows for the analysis of species composition and interactions within a community-based framework.

Recent advancements in high spatial resolution remote sensing have shown the effectiveness of machine learning algorithms, with rapid progress driven by developments in artificial intelligence. The integration of remote sensing data with AI-based methods has proven useful. In our study, we leverage Earth Observation (EO) data derived from Sentinel-2 satellite imagery, including maximum, minimum, median, and near-extremes percentiles of the NDVI and its standard deviation as a key index of forest canopy, density, health and irregularity. We also incorporate Sentinel-2 canopy height derived data, which is essential for understanding forest structure and vertical stratification. These combined metrics provide a comprehensive understanding of vegetation phenology and heterogeneity, supporting more accurate assessments of forest composition and structure.

Field data for this study are derived from forest inventory datasets, serving as the foundation for calibrating and validating the models, enabling precise estimations of species composition, basal area, and other structural parameters critical to biodiversity monitoring.

Mechanistic species distribution models (SDMs) and community assembly (JSDMs) models have driven substantial advancements in biodiversity research, offering insights into environmental filtering and competitive dynamics within ecosystems. In this study, we present a hybrid geospatial approach that combines SDM — specifically, Hierarchical Modeling of Species Communities — with AI algorithms to map forest species composition, relative abundance, and basal area across Italy. This approach is crucial for applications in biodiversity conservation and forest management, enabling more informed decision-making for land and forest management.

Our hybrid framework integrates EO-derived features from Sentinel-2 (e.g., canopy height, NDVI metrics) with pedological and bioclimatic variables, functional traits, Community Weighted Means, Functional Dispersion Index, and phylogenetic distances. By modeling these variables, we aim to capture the complex interrelations between forest species and their environment. To further enhance interpretability, we employ a Machine Learning algorithm based on association rule learning,

The integration of remote sensing data and AI methodologies, combined with field inventory datasets, can provide a powerful tool for biodiversity research and forest management. The incorporation of field data ensures the accurate calibration and validation of models, improving the reliability of predictions. This geospatial approach, leveraging Sentinel-2 EO data, not only advances our capacity to monitor species distributions but also contributes to understanding forest ecosystem dynamics in the context of nature conservation. By bridging remote sensing, AI, and field data, we offer a comprehensive framework to address the multifaceted challenges in biodiversity, ecosystem services, and sustainable land management.