The detection of changes with a high degree of accuracy in multi-temporal satellite imagery is still a challenge due to sensor incapability, location problems and misalignment, radiometric noise, and non-linear temporal variations, which have recently motivated registration-aware and semantic-consistency-based change detection frameworks. Traditional CNNs and Transformer-based methods often have difficulty in effectively computing such continuous-time dynamics. This paper proposes an LSNN model, which is a Lightweight Liquid Siamese Neural Network, that has two branches sharing weights for processing the bi-temporal images acquired at times \(t_1\) and \(t_2\) and to capture the long-range dependencies and temporal irregularities through liquid time-constant neurons. A cross-feature differencing module is integrated into the system in order to facilitate the detection of fine-grained structural transitions and to eliminate variations that are not relevant. The experiments carried out on RGB–RGB and synthetic aperture radar (SAR)–multispectral datasets with separate train–test splits and extensive statistical analysis indicate that the LSNN has superior performance with SeK = 87.46%, OA = 96.18%, F1 = 94.38%, IoU = 91.22%, NSE = 0.92, and PBIAS = – 1.82%. The results presented are superior to the current best methods, that is, the state-of-the-art models, which also include Transformer- and CNN-based ones like SCaNet. Furthermore, the model’s robustness to real-world perturbations is showcased by additional evaluations under Gaussian noise, speckle noise, and illumination variations coupled with slight spatial shifts. The LSNN architecture is thus an efficient and trustworthy alternative for satellite-based high-accuracy change detection because of its excellent temporal modeling capacity, low operational cost, and offering steady performance even under diverse sensing conditions.
