This study presents a deep learning approach for detecting pixel-level changes in hyperspectral remote sensing images. The aim is to create a reliable method for identifying and classifying land cover changes over time, which plays an important role in environmental monitoring, urban development, and agriculture. The task is treated as semantic segmentation using paired hyperspectral images captured at different periods of the same location. The model is based on a U-Net structure that effectively handles the high dimensionality of hyperspectral data. Its encoder–decoder design with skip connections helps capture both spatial and spectral details. The images are normalized and divided into smaller patches before training. A multi-class cross-entropy loss function is used, and parameters are optimized with Adam. Experiments were conducted on the Wetland Area Hyperspectral Change Detection Dataset and the Hermiston Dataset. The model achieved 87.13% overall accuracy and a mean Intersection over Union (IoU) of 59.69% on the Wetland data, and 94.13% accuracy on the Hermiston data. Despite class imbalance, it performed well across multiple land cover categories. These results show that the proposed approach offers an efficient and scalable way to detect detailed changes in hyperspectral imagery.