Remote-Sensing Image Change Detection Based on Adjacent-Level Feature Fusion and Dense Skip Connections

In recent years, the application of deep learning in remote-sensing (RS) image change detection (CD) has been deepening, especially in the challenges faced in lightweight model design, that is, how to effectively balance model complexity and detection accuracy. Although some simplified models have achieved compression in existing research, there are shortcomings in maintaining detailed information and improving detection performance. On the other hand, large network architectures limit their actual deployment in resource-limited environments. This article innovatively proposes a lightweight CD network called Multilevel Feature Cross-Fusing Network (MFCF-Net) to address this issue. MFCF-Net uses a lightweight convolutional neural network for feature extraction and introduces the adjacent layer enhancement module in the feature extraction stage to significantly enhance the network's ability to extract high-dimensional features while only moderately increasing parameters. Furthermore, MFCF-Net integrates dense skip connections and cross-attention mechanisms in the decoder to efficiently propagate contextual information and accurately capture and fuse global spatial features, especially for detecting small changes. Experimental results show that on four public RS CD datasets, MFCF-Net achieves excellent detection performance and has a smaller model size (with parameters as low as 1.35M). Particularly on the WHU-CD dataset, compared with the current state-of-the-art method, MFCF-Net achieves significant improvements, with an F1-score increase in 1.4% and an intersection-over-union improvement of 2.45%.