It is important for the environmental monitoring of the Earth’s surface because it provides rapid and accurate information required for efficient decision-making in matters of disaster relief, urban planning, and ecosystem preservation. This paper focuses on accurately classifying, segmenting, and identifying environmental components from high-resolution satellite images using modern CNNs. It employs pixel-level semantic segmentation for classification into land-cover classes of rangeland, urban land, agricultural land, forest land, water, and barren land, with the help of the encoder-decoder framework from the architecture known as U-Net. The architecture features extraction is used in this study to allow a better focused segmentation at the edges and borders of objects. It includes prior processing which incorporates data augmentation and also geometric corrections along with image normalization. Also, one-hot encoding is applied for converting mask images, that use specific colors to represent various land cover categories, into an acceptable form for the neural network. It allows every pixel to be represented using a binary vector where all elements of the vector are assigned to a specific class of land cover. This makes it possible for the model to capture categorical properties of the data and for the network to produce correct predictions. The encouraging results show that the system is able to automate land cover mapping, anomaly detection, and sustainable resource management. This research fosters ecological monitoring, early threat identification, and practical insights into environmental and urban development through a scalable and effective framework for geospatial analysis. One-hot encoding is also integrated into the model, allowing the model to handle the categorical nature of the land cover classes better for the segmentation task.