Robust Multi-Scale Semantic Segmentation of Aerial Images Using Hybrid Coupled Neuro Trans-Classifier

The aim of semantic segmentation of digital remote sensing image files is to attain semantic uncoupling. The identification of changes at multiple time points is central to the use of remote sensing applications. Supervised deep learning techniques have been used in recent years to achieve high accuracy for change detection in RS images. However, these methods demand large number of labeled samples for training which is again time consuming and tiresome. Based on the proposed lightweight complication module as the encoder and an enhanced Swin Transformer as the decoder, we suggest an Edge Transformer network. The encoder uses a more efficient complication approach than depth-wise divisible complication based on depth-wise complication because it uses our novel Lightweight Feature Extraction Block (LFEB). Additionally, a cross-stage partial structure is integrated into our LFEB for improved point birth. Regarding the decoder, which is based on the Swin Transformer, we creatively suggest a region-based classification for the first time. A benchmark RS dataset application evaluates our model to demonstrate 91.3% accuracy and MSE measurement of 0.009 while achieving RMSE at 0.08 and Standard Deviation at 0.005 which surpasses performance of SSDT and CATNet as current best models. The method proves efficient and robust for segmenting complex RS images which makes it appropriate for urban planning applications and environmental monitoring and disaster assessment work. This directly leverages the Transformer's commemorative sequence to focus on semantic information using the Convolutional Neural Network subcaste, improving the reconstruction delicacy and the model's point expression capability.