A Region-Adaptive Phenology-Aware Network for Perennial Cash Crop Mapping Using Multi-Source Time-Series Remote Sensing

Abstract

Highlights: What are the main findings? We proposed a Region-Aware Module (RAM) that encodes topographic and climatic factors to estimate phenological phase shifts and adaptively recalibrate phenological windows across different regions. We proposed a Multi-Head Phenology-Aware Module (MHA-PAM) that captures short-, mid-, and long-term phenological rhythms through region-modulated attention, enhancing temporal feature learning and interpretability. What are the implications of the main findings? The proposed RAM-PAMNet framework not only enhances cross-regional crop mapping accuracy but also maintains strong generalization under complex climatic and topographic variations. This approach demonstrates significant robustness and interpretability across heterogeneous agricultural regions, offering a promising solution for large-scale, phenology-driven monitoring of perennial crops in remote sensing applications. Monitoring and identifying perennial cash crops is essential for optimizing agricultural resource allocation and supporting sustainable rural development. However, cross-regional recognition remains challenging due to cloud contamination, irregular mountainous topography, and climatic-driven phenological shifts. To address these issues, we propose a Region-Adaptive Multi-Head Phenology-Aware Network (RAM-PAMNet) that incorporates three key innovations. First, a Multi-source Temporal Attention Fusion (MTAF) module dynamically fuses Sentinel-1 SAR and Sentinel-2 optical time series to enhance temporal consistency and cloud robustness. Second, a Region-Aware Module (RAM) encodes topographic and climatic factors to adaptively adjust phenological windows across regions. Third, a Multi-Head Phenology-Aware Module (MHA-PAM) captures short-, mid-, and long-term phenological rhythms while integrating region-modulated attention for adaptive feature learning. The model was trained and validated in Changde, Hunan (694 patches; augmented to 2776; 70%/15%/15% split) and independently tested in Yaan, Sichuan (574 patches), two regions with contrasting elevation, terrain complexity, and hydrothermal regimes. RAM-PAMNet achieved an OA of 83.3%, mean F1 of 78.8%, and mIoU of 65.4% in Changde, and maintained strong generalization in Yaan with an mIoU of 59.2% and a DecayRate of 9.5, outperforming all baseline models. These results demonstrate that RAM-PAMNet effectively mitigates regional phenological mismatches and improves perennial crop mapping across heterogeneous environments. The proposed framework provides an interpretable and region-adaptive solution for large-scale monitoring of tea, citrus, and grape.