Temporal generalization in evergreen leaf type classification using tailored Sentinel-2 composites

Large-scale forest ecosystem mapping relies critically on distinguishing deciduous and non-deciduous tree cover through advanced remote sensing technologies. Existing mapping approaches frequently suffer from spatial resolution limitations and temporal constraints. However, precise, high-fidelity forest cover characterizations are essential for forest management, ecological monitoring, and conservation planning. In this study we applied a novel methodology for classifying leaf types – evergreen versus deciduous – using Sentinel-2 multispectral satellite imagery at 10 meter resolution and machine learning, with the aim of strengthening the robustness of predictions and eliminating the need for retraining for unseen years when training on multi-year data. Key contributions include recursive feature elimination to identify the most relevant spectral bands and indices, and optimizing compositing methods to boost classification accuracy, balancing robustness and temporal detail. Eight machine learning models were tuned and trained on 16,162 tree crowns across 48 areas in Bavarian strict forest reserves (2019 to 2023) and validated with ForestGEO Traunstein Forest Dynamics Plot ground truth data (2018). We achieved an F1 score of 0.863 and an accuracy of 0.839 on the test area. Importantly, we found that model performance improved markedly with tree height, leading us to recommend our methodology for trees taller than 20 m. Results were benchmarked against the Copernicus High Resolution Layer Dominant Leaf Type product, with our top-performing model surpassing the Copernicus product in both metrics. This data-driven approach provides a scalable solution with temporal generalization, leveraging freely available satellite imagery and cloud-compute, aiding more effective forest management and environmental monitoring.