Shuffle-CDNet: A Lightweight Network for Change Detection of Bitemporal Remote-Sensing Images

Abstract

Change detection is an important task in remote-sensing image analysis. With the widespread development of deep learning in change detection, most of the current methods improve detection performance by making the network deeper and wider, but ignore the inference time and computational costs of the network. Therefore, this paper proposes a lightweight change-detection network called Shuffle-CDNet. It accepts the six-channel image that concatenates the bitemporal images by channel as the input, and it adopts the backbone network with channel shuffle operation and depthwise separable convolution layers. The classifier uses a lightweight atrous spatial pyramid pooling (Light-ASPP) module to reduce computational costs. The edge-information feature extracted by a lightweight branch is integrated with the shallow and deep features extracted by the backbone network, and the spatial and channel attention mechanisms are introduced to enhance the expression of features. At the same time, logit knowledge distillation and data augmentation techniques are used in the training phase to improve detection performance. Experimental results showed that the proposed method achieves a better balance in computational efficiency and detection performance compared with other advanced methods.

BSCDNet: A Building Change Detection Network With Category Differentiation Using a Graph Attention Mechanism and Multitask Learning

“…Due to its characteristics and advantages, multitask learning-based SCD is a popular research field. The most recent topics in this field include enhancing change features based on semantic constraints [34], using Siamese semantic-aware encoders to extract multiscale features [14], and guaranteeing temporal symmetry via a temporal-symmetric transformer model [31].…”

“…2) Method Implementation: Several pixel-level and objectlevel semantic change detection methods were selected and implemented for comparison. The pixel-level change detection methods include HRSCD.str1 [15], HRSCD.str4 [15], SCDNet [28], Bi-SRNet [30] and MTSCD [14]. The semantic segmentation and change detection branches adopted crossentropy loss and binary cross-entropy loss with loss weights of 0.25 and 0.5, respectively.…”

“…SCD, which can obtain the semantic types of changed buildings by certain image classification processes, offers a reasonable solution for the proposed building change detection issue. SCD can be categorized into four implementation modes [14][15][16]: training a land cover mapping network to compare the results for pixels in the image pair (Mode 1); utilizing two-branched Siamese networks for bitemporal image change analysis (Mode 2); decoupling the SCD task into independent BCD subtasks and semantic segmentation subtasks (Mode 3); and integrating the BCD and semantic segmentation into a single multitask network so that land cover information can be used for change detection (Mode 4). In these schemes, multitask networks can simultaneously extract building semantic types and morphological changes in an end-to-end manner.…”

Method of Building Detection in Optical Remote Sensing Images Based on SegFormer

“…The semantic fusion module is used in the output part to aggregate rough information in the shallow layer and fine information in the deep layer. The interaction of feature information from different layers is achieved by using the shuffle [ 27 ]. The shuffle is applied to address issues such as the partial loss of building information and the lack of long-distance information.…”