CDasXORNet: Change detection of buildings from bi-temporal remote sensing images as an XOR problem

The up-to-date building information is significant to urban planning and economic assessment. Automatic building change detection (BCD) from bi-temporal remote sensing images is essential for updating building status efficiently. Nevertheless, BCD remains challenging due to the complex building appearance, the diverse imaging conditions, and the building’s positional inconsistencies between the bi-temporal images. Recent convolutional neural network-based BCD methods have achieved impressive performance. However, most existing methods employed subtraction or concatenation to identify building changes. Such simple change-deciding operations ignore the spatial–temporal correlation between the bi-temporal features and cannot model the building changes effectively, resulting in overmuch misclassifications. This paper proposes a hierarchical XOR approximating network CDasXORNet to model building changes robustly. An XOR approximation operation is proposed to produce discriminative building differential features from the bi-temporal inputs. We assume that BCD and the logical XOR function have the same nature (i.e., when the two inputs are identical, the output is unchanged/False; otherwise, it is changed/True). This applies to the building change and unaltered pixels simultaneously. Thus, by approximating XOR operation, CDasXORNet can simultaneously exploit the spatial–temporal correlation and the changed and changeless information of buildings. Hierarchical XOR approximation operations are subsequently designed, which process only high-level features to mitigate the influence of substantial irrelevant spectral differences. In addition, the residual linear attention mechanism is introduced to refine the building change features further. Experiments on three publicly challenging datasets demonstrate that our method achieves promising BCD results with fewer commission errors and higher overall performance than the comparative approaches. •A change-deciding module is created to model both change and changeless information.•A hierarchical XOR approximation operator is designed to reduce false change areas.•A residual linear attention module is introduced to enhance building change features.