CNN-Based Temporal Fusion of UAV and Sentinel-2 Imagery for Crop Monitoring

The use of Sentinel-2 satellite imagery for crop monitoring is constrained by its native spatial resolution. To address this limitation, data fusion-based super-resolution techniques have emerged as a promising approach for generating imagery with enhanced spatial detail. In this study, we propose TVDSR, a Sentinel-2 and UAV image fusion model that integrates temporal and spectral information through a Bidirectional ConvLSTM-based architecture. The images used for super-resolution were constructed from spectral compositions of five vegetation indices. Model performance was evaluated using 15 acquisition dates, including 11 Sentinel-2 images acquired in 2024 and 4 acquired in 2025. Two independent dates were reserved for validation using reconstruction quality metrics. The proposed model was compared against three state-of-the-art reference models: S3-ESRGAN, DeepSent, and UAV-Net. UAV-Net achieved the best performance for the 2024 validation image (RMSE = 0.128, MAE = 0.10, PSNR = 24.06 dB, and SSIM = 0.86). However, TVDSR obtained the best results for the 2025 evaluation image (RMSE = 0.149, MAE = 0.115, PSNR = 22.60 dB, and SSIM = 0.85) and exhibited the lowest performance degradation across acquisition dates, indicating superior temporal generalization capability. Furthermore, TVDSR more consistently preserved the derived vegetation indices, reflecting a better retention of spectral information associated with crop condition, while producing reconstructions that were visually closer to the reference UAV imagery. The joint integration of temporal and spectral information constitutes an effective strategy for improving the robustness and generalization capability of Sentinel-2 image super-resolution for crop monitoring, precision agriculture, crop condition assessment, and agricultural management applications.