Advanced analysis of agroforestry fractional covers leveraging Sentinel-2 data and deep learning techniques

Agroforestry is a land-use management system where trees or shrubs are grown around or among crops or pastureland. These systems play a crucial role in improving biodiversity, enhancing soil health, and contributing to climate resilience. Accurate mapping and monitoring of agroforestry systems are essential for informed decision-making and effective management due to their diverse benefits and functions. One of the main challenges in monitoring these systems is the presence of Agroforestry Fractional Covers (AFCs), areas where different agricultural and forest components coexist, creating heterogeneous landscapes that are difficult to map and monitor accurately using conventional approaches. With the advent of Earth Observation technologies, particularly Sentinel-2's high-resolution multispectral imagery within the European Union's Copernicus Programme, the capacity to map and monitor these components has significantly advanced. However, traditional machine learning methods often struggle to capture the complexity of agroforestry systems when using Sentinel-2 data. Especially, in identifying and delineating AFCs, due to their inherently mixed nature at the pixel level. This characteristic of AFCs complicates the interpretation and analysis of data, requiring more sophisticated approaches that can address both the heterogeneity and temporal dynamics of these systems. This study employs a three-stage approach to characterize AFCs in Sentinel-2 imagery, leveraging both spectral and temporal information. The first stage involves building a time series from Sentinel-2 Level-2A data, stacking imagery captured over the past five years to analyze seasonal variations, phenological cycles, and long-term trends. Monthly intervals are selected to balance detail and practicality, being frequent enough to capture seasonal changes while keeping data volumes manageable and minimizing cloud coverage issues. Gaps in the data, mainly from cloud cover during the rainy seasons, are filled using a weighted linear interpolation method. For each month, composite images are created using a quality-weighted approach, where individual bands are merged based on factors like cloud probability and viewing geometry. The generated monthly composites are processed using a linear Vertex Component Analysis (VCA) unmixing technique to address the mixed pixels challenge. The VCA algorithm identifies pure spectral signatures (endmembers) of distinct components like trees, grass, crops, and soil, without requiring prior scene information. Through the enforcement of physical constraints (abundance sum-to-one and non-negativity), monthly abundance maps are generated that show the distribution of these components across the landscape. These maps decompose the complex spectral patterns of AFCs into their constituent parts, preparing the data for temporal analysis. The final stage examines the temporal patterns of these unmixed components. The abundance time series is fed into a transformer-based deep learning model, following an encoding step that incorporates both positional and seasonal information. This encoding captures not just the sequence of events but also their cyclical nature. The transformer architecture, with its attention heads and encoder layers, can detect complex patterns in time series data. This approach is particularly relevant for the agroforestry context due to its ability to dynamically weigh different time periods, enabling the understanding of how components like crops and trees develop through their growth cycles. The motivation behind this research is threefold: (1) to develop a methodology for disentangling the spectral signatures of different agroforestry components within mixed pixels, (2) to leverage the temporal dimension of Sentinel-2 data through transformer-based deep learning to capture seasonal and phenological patterns, and (3) to provide stakeholders with accurate, temporally-aware mapping tools that can support sustainable land management decisions and climate resilience strategies. The expected outcome of this research is to achieve a detailed mapping of AFCs, enhancing our ability to monitor and manage complex agroforestry systems. By integrating Sentinel-2 data with advanced unmixing techniques and transformer-based models, we envision producing abundance maps that accurately reflect the spatial and temporal distribution of agroforestry components. This can not only improve classification accuracy but also enable the prediction of future states of these dynamic landscapes.