Overcoming the uncertainty challenges in detecting building changes from remote sensing images

Detecting building changes with multi-temporal remote sensing (RS) imagery at a very high resolution can help us understand urbanization and human activities, making informed decisions in urban planning, resource allocation, and infrastructure development. However, existing methods for building change detection (BCD) generally overlook critical uncertainty phenomena presented in RS imagery. Specifically, these uncertainties arise from two main sources: First, current building change detection datasets are designed primarily to detect changes in buildings, while changes in other land-cover classes are often classified as an unchanged background. Because of the manual labeling process, background elements that resemble buildings, such as roads and bridges, are at significant risk of being misclassified as building changes, introducing aleatoric uncertainty at the data level. Second, changes in parts of buildings that affect appearance, texture, or style without altering their semantic meaning, known as pseudo-changes, along with the imbalance between changed and unchanged samples, together lead to epistemic uncertainty at the model level. To address these challenges, we present an Uncertainty-Aware BCD (UA-BCD) framework. In detail, we employ a Siamese pyramid vision transformer to extract multi-level features from bi-temporal images, which are then decoded via a general decoder to obtain a coarse change map with inherent uncertainty. Subsequently, we introduce the aleatoric uncertainty estimation module to estimate the aleatoric uncertainty and embed it into the feature space. Then, a knowledge-guided feature enhancement module is developed to leverage the knowledge encoded in the coarse map to enhance the multi-level features and generate a refined change map. Finally, we propose an epistemic uncertainty estimator that takes the bi-temporal images and the refined change map as input and outputs an estimate of epistemic uncertainty. This estimation is supervised by the entropy calculated from the refined map, ensuring that the UA-BCD framework can produce a change map with lower epistemic uncertainty. To comprehensively validate the efficacy of the UA-BCD framework, we adopt a dual-perspective verification approach. Extensive experiments on five public building change datasets demonstrate the significant advantages of the proposed method over current state-of-the-art methods. Additionally, an application in Dongxihu District, Wuhan, China, confirms the outstanding performance of the proposed method in large-scale BCD. The source code of the project is available at https://github.com/Henryjiepanli/UA-BCD.