Land Cover Classification Using Modified U-net: A Robust Approach for Satellite Image Analysis

Accurate identification of land use and land cover (LULC) is crucial for effective resource management and various geospatial applications. In this research paper, a deep learning-based approach for semantic segmentation of satellite images is proposed to classify LULC. The power of transfer learning algorithms, specifically U-net is leveraged, combined with satellite data to achieve high accuracy in image segmentation. The method utilizes a False Color Composite (FCC) image derived from satellite imagery with a spatial resolution of 10m. Previously classified image is employed as training data in patch format. By modifying the U-net architecture to suit our requirements and employing ensemble classification, we achieve impressive training accuracy of 93.77% and validation accuracy of 93.60%. These results show the model’s precise training, overcoming underfitting or overfitting despite low-resolution satellite images. This approach highlights deep learning and transfer learning’s effectiveness in satellite image classification, particularly for LULC identification.