Straw burning is a major source of regional air pollution. However, its reliable remote sensing detection faces problems in distinguishing agricultural fires from non-agricultural thermal anomalies, adequately leveraging burning seasonality, and overcoming the scarcity of pixel-level annotations. To comprehensively address these issues, this study proposes an end-to-end framework for straw burning identification that integrates spatio-temporal weighting and semi-supervised learning. The framework introduces a data-driven spatial weight optimization method to automatically learn discriminative weights for diverse land cover types (e.g., farmland, industry), replacing subjective empirical settings. Furthermore, a temporal weighting model, developed using Kernel Density Estimation, dynamically adjusts classification confidence according to historical burning seasonality, enhancing recall during peak seasons while suppressing off-season false positives. Finally, an adapted Dual-Backbone Dynamic Mutual Training (DB-DMT) strategy collaboratively leverages both limited labeled (24.5%) and abundant unlabeled (75.5%) high-resolution imagery, significantly improving model generalization in label-scarce scenarios. Validation across five representative regions of China demonstrated the framework’s superior performance, achieving a semantic segmentation mean Intersection over Union (mIoU) improvement of 3.33% (to 71.92%) and increasing precision in Henan from 95.21% to 97.71%. Crucially, the framework effectively reduced the off-season false positive rate (FPR) from 5.14% to a mere 0.23% in highly industrialized regions like Tianjin. By systematically mitigating both spatial geolocation bias and seasonal phenology confusion, our approach offers a robust and scalable solution for straw burning monitoring and a transferable paradigm for other environmental remote sensing applications.
Lyu et al. (Sun,) studied this question.