Cross-Regional Sugarcane Identification via Phenology-Constrained Joint Distribution Adaptation Method

Sugarcane is a vital sugar crop globally, yet its large-scale remote sensing monitoring is often hindered by the high costs of field sampling and the difficulty of reusing historical data across regions.Due to variations in climatic conditions and planting practices, sugarcane exhibits significant spatiotemporal phenological shifts across different regions, causing a sharp decline in the accuracy of traditional supervised classification models when applied cross-regionally. To address this challenge, this study proposes a Phenology-Constrained Joint Distribution Adaptation (PC-JDA) method that integrates biological mechanisms with transfer learning. Building upon the standard JDA algorithm, we innovatively introduce prior phenological knowledge as a constraint mechanism. Specifically, we utilize Dynamic Time Warping (DTW) to quantify phenological similarities between the source and target domains. Furthermore, during the iterative optimization process of JDA, standard NDVI time-series curves of sugarcane are employed to screen and correct the pseudo-labels generated for the target domain, thereby mitigating negative transfer effects. Experimental results transferring from Fusui (source) to Xuwen (target) demonstrate that this method effectively aligns the feature distributions of sugarcane between regions. It significantly improves identification accuracy in the target domain without labeled samples, providing a feasible and cost-effective solution for cross-regional crop mapping.