MultiLayer and ChannelSpatial Feature Interaction Technique for Temporal Variation Detection in Satellite Imagery

Change detection in remote sensing imagery is an important part of Earth observation because it lets us keep an eye on land use, urban growth, environmental changes, disaster response, and infrastructure planning. The main purpose of change detection is to find and separate areas that have changed between two or more images taken at different times, which are often called bitemporal images. Pixel-level change detection looks at each corresponding pixel in temporal images to find even the smallest changes. Conventional techniques frequently depend on basic image differencing, thresholding, or standard convolutional neural networks (CNNs), which predominantly emphasise spatial disparities among pixels. These methods are good at finding big changes, but they often miss small or gradual ones, especially in complicated real-life situations, because they don't use enough feature information. Deep learning techniques have recently become a powerful way to find changes in remote sensing data. Convolutional neural networks and encoder-decoder architectures pull out hierarchical feature representations from bitemporal images, which makes segmentation more accurate than traditional methods. In this case, feature maps made by deep networks have both spatial and channel dimensions, and each one stores different information. The spatial dimension records the structure and position of objects, while the channel dimension records changes in meaning, spectrum, and high-level features. But most current change detection methods only look at differences in the spatial dimension and miss the important information that is stored in the channel dimension. As a result, small changes in the spectrum, the environment, or the structure may not be noticed, which limits the overall performance of traditional models. This study suggests a new deep learning framework called LENet that combines a Channel-Spatial Difference Weighting (CSDW) module and a Layer-Exchange decoding structure to improve pixel-level change detection. The CSDW module is made to find differences not just in the spatial dimension but also in the channel dimension. This module makes the model more sensitive to small, gradual, and large-scale changes by combining and sharing difference information from both sides. The channel-spatial weighting mechanism helps the network tell the difference between important changes and background noise and unimportant features. This makes it better at finding and distinguishing features. Also, bitemporal images naturally have strong correlations because they show the same place on Earth at different times. To be able to find changes, models need to be able to find these correlations between images and use temporal dependencies. LENet uses a Layer-Exchange decoding mechanism to do this. This lets feature maps from the two temporal images interact and share information across decoder layers. This mechanism improves the alignment of temporal features and makes the model better at using complementary information from both images, which helps it find changed areas more accurately. The proposed framework overcomes the limitations of traditional methods that only look at spatial differences or treat temporal features separately by combining channel-spatial difference learning with better inter-image feature interaction. We ran a lot of tests on four well-known benchmark datasets—CLCD, PX-CLCD, LEVIR-CD, and S2Looking—to make sure that the proposed LENet model works. The evaluation metrics encompassed Precision, Recall, F1-score, and Overall Accuracy, guaranteeing a comprehensive appraisal of model performance across various scenarios. Experimental results show that LENet works much better than other methods, such as traditional Siamese networks, encoder-decoder architectures, and attention-based models. The model is better at finding small, gradual, and large changes, while lowering the number of false negatives and making it more sensitive to complex changes. In conclusion, the proposed LENet framework is a strong, accurate, and quick way to find changes at the pixel level in remote sensing images. The system gets both spatial and channel information, models how bitemporal images depend on each other over time, and works better than any other system on a wide range of datasets by combining channel-spatial difference weighting and layer-exchange decoding. This research enhances intelligent Earth observation systems and establishes a robust basis for future investigations in multi-temporal and high-resolution remote sensing change detection.