Lightweight Heterogeneous Change Detection Based on Backpropagation-Free Online Fine-Tuning

Change detection (CD) based on heterogeneous remote sensing images is crucial for disaster response, including floods and earthquakes. In this article, a lightweight backpropagation-free heterogeneous CD method based on online local learning (OLL) is proposed to minimize the computational requirements of the online fine-tuning of the CD network. During the OLL-based fine-tuning, weight gradients are directly calculated using the current layer output features, eliminating the need for the computationally intensive backpropagation process. To address the challenge of unavailable feature labels, heterogeneous CD is innovatively modeled as a contrastive learning process. The contrastive learning enables the network to increase/decrease differences of changed/unchanged features, thereby eliminating the dependence on feature labels. Furthermore, the weight gradient computation process is newly integrated into the forward propagation process on a field-programmable gate array (FPGA) based on OLL fine-tuning. This integration reduces the memory required for storing intermediate activations compared to traditional backpropagation-based methods, thereby improving the viability of implementation in resource-constrained environments. The experimental results on real heterogeneous remote sensing images show that our OLL-based CD method achieves performance comparable to state-of-the-art methods while significantly reducing computational load, improving processing speed, and lowering energy consumption.