Lightweight liquid siamese neural network for robust multimodal satellite image change detection

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Accurate change detection in multi-temporal satellite imagery is challenging due to sensor limitations, noise, and complex temporal variations.
• Existing Convolutional Neural Network (CNN) and Transformer models struggle with continuous-time dynamics in satellite data.
• Registration-aware and semantic-consistency frameworks are emerging to address these limitations.

Purpose of the Study:

• To propose a novel Lightweight Liquid Siamese Neural Network (LSNN) for high-accuracy change detection in bi-temporal satellite imagery.
• To enhance the model's ability to capture long-range dependencies and temporal irregularities.
• To improve the detection of fine-grained structural changes while mitigating irrelevant variations.

Main Methods:

• Developed a Lightweight Liquid Siamese Neural Network (LSNN) with shared-weight branches for processing bi-temporal images.
• Integrated liquid time-constant neurons to model temporal irregularities and long-range dependencies.
• Incorporated a cross-feature differencing module for precise structural transition detection and noise reduction.

Main Results:

• LSNN achieved superior performance on RGB-RGB and SAR-multispectral datasets, outperforming state-of-the-art CNN and Transformer models.
• Key performance metrics include SeK = 87.46%, Overall Accuracy (OA) = 96.18%, F1-score = 94.38%, Intersection over Union (IoU) = 91.22%.
• Demonstrated robustness against Gaussian noise, speckle noise, illumination variations, and spatial shifts.

Conclusions:

• The LSNN model offers an efficient and reliable solution for high-accuracy satellite-based change detection.
• Its strong temporal modeling capacity and low operational cost make it a valuable tool.
• The model provides consistent performance across diverse sensing conditions and perturbations.