Spatiotemporal transferability of multi-temporal random forest model for mapping Euryale ferox Salisb. plantations using Sentinel-2 imagery

Euryale ferox Salisb. is a widely cultivated aquatic plant of high edible and medicinal value, but lack of reliable production information hinders effective cultivation planning. Existing studies on Euryale ferox Salisb. mapping remain limited and often fail to fully exploit phenological information, resulting in suboptimal classification accuracy. More importantly, the spatiotemporal transferability of classification models, defined as the ability to maintain accuracy when applied to different regions or years without retraining, has been largely overlooked in crop mapping. To address this issue, we innovatively propose the Spatiotemporal Transferability Index (STI), comprising two novel metrics, STI_F1 and STI_Overall, to systematically quantify model performance retention after cross-regional and cross-year transfer. Using Sentinel-2 imagery acquired over Yugan County in 2024, we optimized features from spectral bands, vegetation indices, texture metrics, and DEM, and then constructed single-temporal (May, July, August, September, and October) and multi-temporal Random Forest (RF) models for Euryale ferox Salisb. planting area extraction. The spatiotemporal transferability of these models was then independently validated in Jinxian County in 2025, and assessed using the proposed STI metrics. The results indicated that Band 1, Band 12, the NDVWI, and DEM were consistently retained as discriminative features throughout the entire growth period. Among the single-temporal RF models, those for July, August, and September achieved high overall accuracies, all exceeding 97%, while the multi-temporal RF model reached 99.87%. In the spatiotemporal transferability validation, the multi-temporal RF model demonstrated the strongest transferability, achieving an STI_F1 of 94.02%, an STI_Overall of 88.86%, an overall accuracy of 90.13%, a Kappa coefficient of 0.8728, and an F1-score of 93.83%. The proposed framework integrates multi-temporal RF classification with an STI-based transferability evaluation, enabling accurate and efficient mapping of Euryale ferox Salisb. planting areas. The framework shows potential for extension to other crops and regions, supporting sustainable agricultural decisions.