Deep learning (DL) models for change detection (CD) are affected by the changed/unchanged and hard/easy sample imbalance during the training process. Most of the loss functions for solving the sample imbalance problem are a static loss which is difficult to adapt to the variation of data distribution. In this article, we propose a dynamic method termed dynamic adaptive focal loss (DAFL) function. Specifically, we first statistically count the number of changed/unchanged samples in different batches of training data and a dynamic weighting factor is constructed to dynamically and adaptively balance their proportions. Furthermore, a dynamic modulation factor is proposed to suppress the hard/easy sample imbalance. In addition, we employ a CD model based on a progressive scale expansion network (PSENet), which is trained using DAFL for remote sensing images. Experimental results on three CD datasets (CDDs, Sun Yat-sen University (SYSU)-CD, and LEarning VIsion Remote Sensing (LEVIR)-CD) indicate that DAFL outperforms all baseline approaches. Our proposed method achieves the maximum improvement, with an F1 -score of 0.33%, 0.8%, and 0.94% for sufficient sample size, and 2.15%, 2.61%, and 3.89% for small-sample size, respectively. This advancement is crucial for the application of CD, which provides an alternative method for solving sample imbalance in the condition of varying sample sizes, especially small sample conditions that are common in the real scenario.

Article link: https://doi.org/10.1109/TGRS.2024.3392696