Adaptive Patch Reshaping for Edge-Based Semantic Segmentation in Remote Sensing

Accurate boundary detection between distinct classes is a major challenge in the semantic segmentation of remote sensing images. A widely used strategy involves dividing large images into patches for classification, which requires carefully prepared training, test, and validation datasets. When the resolution of the input image exceeds that of the available labels, upsampling the labels to match the image resolution often leads to suboptimal results. To address this issue, we propose an alternative approach, Adaptive Patch Reshaping (APR), that subdivides both the input image and label patches, then reassigns them to classes according to an information-theoretic criterion. By reshaping patches based on their similarity to representative class characteristics, this method constructs a new training set from the original one. Experimental results using Sentinel-2 imagery demonstrate that the proposed technique, implemented with a lightweight CNN, achieved a classification precision of 89% for the training set, 86% for the test set, and 81% for the validation set. Furthermore, comparative analysis confirms that APR improves the delineation of boundaries between regions, yielding higher F1-scores and Structural Similarity Index values compared to standard patch-based baselines.