Land use classification using multi-year Sentinel-2 images with deep learning ensemble network

Accurate land use classification is essential for urban planning, environmental monitoring, and agricultural management. Sentinel-2 satellite imagery provides rich spatial and spectral information suitable for this purpose. This study proposes a deep learning ensemble network named IRUNet, which integrates InceptionResNetV2 with a UNet framework for multi-year Sentinel-2 imagery classification over the Katpadi region (2017-2024). Unlike prior works, IRUNet utilizes multi-scale feature fusion and incorporates Test-Time Augmentation (TTA) to enhance prediction robustness. While the data spans multiple years, each year is treated as an independent input without modeling temporal sequences. The proposed method demonstrates superior performance over UNet, ResUNet, and Attention-UNet models, achieving an accuracy of 98.21% and Dice similarity coefficient (DSC) of 88.96%. Additional metrics including precision (94.71%), recall (89.19%), F1-score, and Kappa coefficient have been reported. This research contributes a high-performance, generalizable framework for multi-year land use classification.