Crop classification mapping in multi-cloud Regions: An integrated approach using GF-SG time series reconstruction and 3D deep convolutional networks

Persistent cloud contamination severely affects optical satellite-based crop classification, creating critical gaps in phenological monitoring essential for food security. This study presents GS-3D FCN, a novel framework integrating Gap Filling-Savitzky-Golay (GF-SG) time series reconstruction with 3D Fully Convolutional Networks (3D FCN) for robust crop mapping under extreme cloud conditions. The method reconstructs cloud-contaminated pixels using GF-SG algorithm and leverages MODIS-Landsat fusion to generate continuous NDVI time series. These reconstructed layers are integrated as spectral-temporal dimensions into the 3D FCN architecture, enabling hierarchical extraction of crop phenological patterns. An enhanced Automatic Cloud Cover Assessment algorithm adaptively adjusts decision criteria to maximize valid pixel retention. To address severe class imbalance from cloud occlusion, we apply Synthetic Minority Over-sampling Technique with Gaussian Noise (SMOGN) combined with Cross-entropy loss Middle Supervision (CE-MidS) and Supervised Contrastive loss Middle Supervision (SupCon-MidS) intermediate supervision strategies, facilitating discriminative feature learning. Applied to the cloud-prone South Bank Yellow River irrigation area in Inner Mongolia, GS-3D FCN achieves 93.80% overall accuracy across four sites. The framework demonstrates exceptional stability, maintaining 85.2% accuracy at 80% cloud coverage compared to 33.6-37.2% accuracy degradation in traditional methods. The GF-SG component achieves RMSE of 0.0853 with edge preservation index of 0.0085. Analysis reveals the critical role of dataset balance in crop mapping under cloud interference. The proposed GS-3D FCN framework provides a simple and effective solution for crop mapping in cloud-prone areas. By integrating near-real-time NDVI data sources, the framework provides a feasible path for accurate crop monitoring in global continuous cloudy areas, and provides a new perspective for future research on the interaction between extreme climate and crop mapping.