Deep-learning-based grassland mapping with Sentinel-2: prioritizing key spectral bands and time periods

Abstract

Accurate grassland mapping is essential for biodiversity conservation and sustainable land management, yet remains challenging due to the spectral and temporal variability of grassland ecosystems. This study presents a Deep Learning approach for grassland classification using multi-temporal Sentinel-2 imagery, incorporating a dynamic feature selection mechanism to prioritize informative spectral bands and time periods. In order to allow the model to adaptively focus on discriminative temporal spectral patterns, we compare a baseline neural network with a modified design that learns to weight input features dynamically. Our findings demonstrate that the feature selection model achieves superior performance (Accuracy: 0.954 ± 0.004, MCC: 0.726 ± 0.027) compared to both the baseline network and single-date models, highlighting the importance of temporal diversity in grassland classification.