3M-CDNet-V2: An Efficient Medium-Weight Neural Network for Remote Sensing Image Change Detection

Abstract

Remote sensing-based change detection (CD) is a critical technique of detecting land surface changes in earth observation. Inspired by recent success of a lightweight CD network -3M-CDNet, we implemented 9 meaningful modifications to it, for example, the incorporation of MHSA (Multi-Head Self-Attention). This elaborately designed model is termed as 3M-CDNet-V2. Its effectiveness and advantages were demonstrated on three engineering CD datasets, and experimental results indicated that:(1) relative to other state-of-the-art algorithms, the proposed model obtained very competitive or slight better performance on both visual comparison and quantitative metrics evaluation. (2) By applying a novel transfer learning strategy, 3M-CDNet-V2 could perform well on the small dataset. (3) The incorporation of MHSA brings a substantial accuracy improvement while moderately increasing the computational complexity. (4) The late fusion framework and deep supervision contribute most to the performance gain of 3M-CDNet-V2, while the introduction of low-level features to the classifier by skip connection could guide the model to focus on detailed spatial information such as small changes, narrow shaped objects, or accurate boundaries. We hope that our 3M-CDNet-V2 model helps in improving the understanding of network architecture design for CD.

Transformers for Remote Sensing: A Systematic Review and Analysis

“…Additionally, incorporating semantic and pixel-level information into the loss function is another way to address this problem. To tackle the issue of small object change detection, Zhao et al proposed a novel method that integrates multi-head self-attention for computational efficiency optimization [ 49 ]. The method incorporates skip connections to enhance the classifier’s ability to detect small objects.…”