UMV2: Deep Learning Model for Semantic Segmentation of Satellite Imagery

Semantic segmentation is a computer vision task that categorizes each pixel in an image into a class or object. Although a number of relevant architectures have been proposed in recent years, they incur the problems like computational cost,large amounts of training data, class imbalance, edge uncertainty, varying sizes of objects, shadow and lighting variations. Such a more number of drawbacks degrades the semantic segmentation performance in terms of accuracy, efficiency and generalization capability. In this work, comprehensive architecture UMV2 for satellite image semantic segmentation is proposed. The UMV2 utilizing a fusion of Unet++ architecture and the lightweight MobileNetV2 encoder deep learning model. The Unet++ architecture, an extension of the widely adopted Unet, is employed for its ability to capture hierarchical features and enhance segmentation performance. Integrating MobileNetV2 as the encoder provides computational efficiency, making the model well-suited for resource- onstrained environments, such as satellite image analysis on edge devices. The proposed model leverages the strengths of both architectures, combining the expressive power of Unet++ with the efficiency of MobileNetV2. Extensive experiments are conducted on a diverse satellite image dataset, evaluating the model’s segmentation accuracy of 0.89, mean IOU of 0.52, precision of 0.80, recall of 0.83 and F1-score of 0.82 with the state of art methods. The results demonstrate the effectiveness of the proposed approach in achieving accurate and efficient satellite image segmentation, making it a promising solution for real-world applications in remote sensing and geospatial analysis.