A Multisource Time-Series Remote Sensing Data Fusion Network for High-Spatial-Resolution Crop Classification

High-spatial-resolution multisource remote sensing image fusion enables precise delineation of crop boundaries and spatial patterns, and distinguishing fine-grained categories that are difficult to separate at medium resolution. This provides a refined and reliable technique for classifying small and fragmented crops with imbalanced category distributions. However, images acquired from different sensors vary in both resolution and semantic characteristics. Effectively mitigating resolution mismatches and leveraging complementary information remains an open challenge. To address this challenge, we propose a multisource and multiresolution image fusion framework. The framework is designed to synergistically integrate remote sensing data of different resolutions, thereby enhancing crop classification accuracy. First, we introduce the multireceptive-field residual block (MRRB). It incorporates multiple receptive field pathways through multiscale dilated-convolution branches, thereby strengthening the model's ability to capture multiscale texture and boundary features of crops. By preserving spatial resolution and enriching contextual information via residual connections, MRRB effectively characterizes crop spatial structure and distribution. Its compact design and moderate parameter volume significantly improve classification accuracy. In addition, we design a multisource multitemporal (MSMT) fusion module to jointly model dynamic temporal variations and spectral differences across multisource time series. This module adaptively balances the contributions of different data sources through parallel temporal and spectral attention branches combined with learnable fusion weights, thereby enhancing the integrated representation of spatiotemporal, spectral, and polarization features. Extensive experiments demonstrate that, while maintaining computational efficiency, the proposed MRRB and MSMT modules together outperform existing state-of-the-art methods in high-resolution crop classification tasks.