A 3D convolutional neural network approach for monitoring land cover change using Sentinel-2 satellite imagery

Effective monitoring of land cover dynamics is essential for sustainable ecosystem and resource management, particularly in rapidly changing landscapes. This study presents an enhanced 3D Convolutional Neural Network (3D-CNN) architecture with skip connections for detecting land use/land cover (LULC) changes using high-resolution Sentinel-2 imagery. The proposed framework jointly exploits spatial and temporal patterns through volumetric kernels and architectural innovations such as skip connections to improve learning depth and accuracy. Applied to Fars Province, Iran, the model achieved superior classification performance, with an overall accuracy of 98.94% using fused NDVI and Sentinel-1 radar bands (VV, VH). The model’s performance is expressed as a Spatial Consistency Index (CI) rather than simple accuracy, achieving 98.94% consistency against the FRWMO operational baseline and 97.8% validation accuracy via independent PlanetScope (2023) samples. The model achieved a spatial consistency index of 98.94% over stable land-cover zones, with cross-split mean accuracy of 80.6% confirming its generalization capability. Unlike conventional methods, our model captures subtle and large-scale transitions without relying solely on static classification comparisons. While map-based reference data was used for consistency evaluation, future research will incorporate independent ground truth for robust validation. Our results highlight the potential of deep, spatiotemporally aware models in national-scale environmental monitoring applications.