GARP - A Hybrid Preprocessing Technique for Semantic Segmentation of Satellite Images with U-Net Architecture

Accurate and intricate semantic segmentation of satellite images holds paramount significance across diverse domains, facilitating precise land cover classification and object identification. This paper introduces a novel hybrid preprocessing technique, GARP (Gamma Correction, Laplacian kernel and Histogram Equalization) for accurate and precise pre-processing. The work also compares the results obtained from GARP with those obtained from existing techniques like Contrast Limited Adaptive Histogram Equalization (CLAHE), gamma correction, histogram equalization, and Laplacian Kernel. Furthermore, the research conducts an evaluation across three distinct loss functions-Jaccard loss, Focal loss, and Binary Cross-Entropy loss. Notably, the U-Net architecture, renowned for its efficacy in medical imaging, is tailored to satellite imagery, further boosting performance. The proposed approach integrates the benefits of Histogram Equalization, Gamma Correction and Laplacian Kernel with the U-Net framework. The outcome substantiate the excellence of this approach, yielding remarkable performance metrics: a mean Intersection over Union (mIoU) of 0.78, a Dice score (DSC) of 0.87, and an Area Error Ratio (AER) of 0.172. These results provide a significant 63.2 % decrease in AER, 8.2 % increase in the Dice score and a 8.6 % improvement in the mIoU when compared with results obtained on original images (without any pre-processing). The paper essentially provides a helpful framework for pursuing precise and effective semantic segmentation in satellite image analysis.