Three-Season Spectral—Phenological Modeling for Source-Domain Classification and Cross-Region Candidate Mapping of Populus euphratica Using Sentinel-2 Imagery

Accurate regional mapping of Populus euphratica remains challenging because of its spectral similarity to co-occurring riparian vegetation and the limited availability of species-level reference data across geographic regions. To address these challenges, we developed a three-season Temporal Transformer that integrates Sentinel-2 spectral bands and vegetation indices from spring, summer, and autumn. In the lower Tarim River source region, the model achieved an overall accuracy (OA) of 92.06% and a Macro-F1 score of 92.25% under stratified random pixel-based evaluation. Under spatial-block evaluation, the model achieved an overall accuracy of 87.16%±0.60% and a Macro-F1 score of 86.39%±0.74% across three training runs. Ablation experiments further demonstrated that combining all three seasons with the complete set of spectral bands and vegetation indices yielded the highest mean classification performance among the evaluated configurations. The source-trained model was subsequently applied to Sentinel-2 imagery over Kuqa City without retraining or fine-tuning, identifying 1173.76 km2 (8.79% of the valid mapping area) as candidate P. euphratica areas. Of the predicted candidate area, 68.36% fell within the high-confidence category, with a mean normalized entropy of 0.213. These findings highlight the value of multi-season spectral–phenological integration for source-region species classification and demonstrate its potential for uncertainty-aware candidate mapping and field-survey prioritization in regions with limited reference data.