Land cover mapping over large Mediterranean basins using Sentinel-2 imagery and a 1D-ResNet deep learning approach

Land cover changes are among the main drivers of global environmental change, making accurate and up-to-date information on land cover essential for environmental monitoring and land management. This study presents a land cover classification of the Guadiana Hydrographic Demarcation (GHD), one of the largest river basins in the Iberian Peninsula and a region subject to diverse environmental pressures. Using Sentinel-2 imagery and a one-dimensional residual neural network (1D-ResNet), land cover maps were produced for the years 2017 and 2025. The classification performance was assessed for 2017 using reference data derived from the Spanish Land Occupation Information System (SIOSE). The proposed approach achieved a high overall accuracy of 0.98 across seven land cover classes over an area of 55,407 km2 . The comparison between the 1D-ResNet classification and SIOSE revealed a moderate spatial agreement of 65.70% with a weighted Intersection over Union (IoU) of 0.52, highlighting notable differences related to seasonal conditions, minimum mapping units and contrasting mapping approaches. Changes between the 2017 and 2025 classifications were particularly evident for the Broadleaf class, which increased by 9.05% of the total study area, reflecting interannual variability in vegetation conditions. Overall, the results demonstrate the potential of combining Sentinel-2 imagery and deep learning models for pixel-based land cover mapping over large and heterogeneous Mediterranean landscapes. Rather than replacing existing land cover products, the proposed approach complements official datasets by providing updated, fine-scale spatial information that can support environmental monitoring, spatial planning and land management under changing environmental conditions.