DFDR‐MT: A Spatial FDR–Calibrated Confidence‐Aware Pseudo‐Labeling Model for Semi‐Supervised Change Detection

ABSTRACT

Change detection (CD) in remote sensing plays a critical role in monitoring environmental dynamics, urban expansion, and land‐use changes, yet it remains challenging due to severe class imbalance between changed and unchanged regions and the high cost of obtaining pixel‐level annotations for bi‐temporal imagery. To address these limitations, this paper presents DFDR‐MT, a semi‐supervised change detection framework that effectively leverages unlabeled data through a statistically principled pseudo‐labeling strategy combined with a hybrid Siamese Attention Transformer U‐Net (SATU) architecture. SATU integrates convolutional networks for local feature extraction with Transformers for global context modeling, enhanced by attention‐guided feature fusion. Central to DFDR‐MT is a Dynamic Entropy‐Masked False Discovery Rate (FDR) mechanism, which controls the expected false positive rate while filtering uncertain predictions, ensuring the reliability of pseudo‐labels during iterative training. Within a Mean Teacher paradigm, an entropy‐weighted consistency loss dynamically modulates the influence of uncertain predictions, improving optimization stability, and Hard Negative Mining (HNM) explicitly addresses class imbalance by focusing on challenging samples. Extensive experiments on the LEVIR‐CD and SECOND datasets demonstrate that DFDR‐MT achieves state‐of‐the‐art performance, attaining an F1 score of 91.91% and IoU of 88.41% on LEVIR‐CD, and an F1 of 93.43% and IoU of 84.47% on SECOND, outperforming strong supervised and semi‐supervised baselines. These results validate the effectiveness of the proposed components and highlight DFDR‐MT as a robust, data‐efficient, and statistically grounded framework for high‐accuracy change detection under limited supervision.