The study describes a Pyramid Scene Parsing Network based deep learning method for semantic segmentation and classification of satellite images. PSPNet’s unique pyramid pooling module captures both local and global contextual information that helps in segmentation. For better model evaluation, predicted mask images are divided into fine, mixed and blurred patches, ensuring detailed assessment. Specialized accuracy metrices are used for each type of patch for fine patches IoU, Boundary IoU, Dice Coefficient, and Pixel Accuracy is used for accurate boundary. For mixed patches Mean IoU, Weighted IoU, and Mean Per-Class Accuracy is used for balanced classification. For blurred patches SSIM and MAE is used that focuses on texture consistency and reconstruction quality. This customized metric evaluation considers spatial complexity, making accuracy measurement more robust and realistic. The PSPNet-based method reaches an overall metric value of 5.3163 (Original Image) and 5.3088 (Masked Image), proving its potential in detecting complex spatial relationships and enhancing semantic segmentation for satellite imagery.