Multi-Temporal Feature Fusion for Spatial-Temporal Dependencies Enhanced Change Detection

Abstract

Change detection (CD) task aims at identifying differences between two images of the same spatial location, captured at different times. Remote sensing change detection (RSCD) faces significant challenges due to the intricate nature of objects within a scene. One of the challenges is that objects that share the same semantic concept can exhibit different spectral characteristics over time. Additionally, variations in illumination and the complex nature of building rooftops that blend seamlessly with their background environments further complicate the task for deep-learning based change detection (DLCD) methods. This is mainly because DLCD methods focus on color and texture. These methods may struggle to effectively capture multiscale changes and may overlook important details at different granularity. Exploring approaches that integrate additional sources of information, such as spatial context, temporal dependencies, semantic segmentation, or domain-specific knowledge, can improve the robustness and effectiveness of DLCD models in CD. Previous methods encoded bi-temporal images independently without leveraging spatial-temporal dependencies. However, this work proposes exploring the relationship among different spatial-temporal features by fusing features from two images, taken at different time points. This work applies a series of convolutional and pooling layers to each input image, and then concatenates the encodings of these layers before further processing, hence combines feature maps from two different time points or inputs. This enables DLCD models to extract richer, more informative representations from the input data, allowing the network to grasp distinct representations that leads to enhanced effectiveness and improved performance in change detection assignments. Our method attains an overall accuracy (OA) of 96.7%, on the EGY BCD dataset.