Identifying granitic pegmatite-type lithium deposits based on remote sensing technology is a significant method for lithium ore prospecting. To enhance the information extraction accuracy of the deep learning-based semantic segmentation method for granitic pegmatites, this study improved the classic U-Net network. A batch normalization module was added to the convolutional layer of the encoder part, with the ReLU activation function replaced by the ReLU6 activation function. Simultaneously, a composite loss function was constructed to improve operational efficiency and reduce the precision loss in the training process. The domestic GF-2 images of a granitic pegmatite-type lithium deposit were employed to create a dataset for experiments. The results show that the improved U-Net model effectively identified the information on granitic pegmatites in the study area covered by GF-2 images. Compared to the original U-Net network, U-Net model based on VGG backbone network, U-Net model based on MobileNetV3 backbone network, and conventional random forest model, the improved U-Net model has its average intersection over union increased by 14.69, 0.95, 5.08, and 35.34 percentage points, respectively. Moreover, its F1-score increased by 18.38, 1.02, 5.7, and 54.59 percentage points, respectively. Hence, the improved U-Net model achieves the high-precision automatic extraction of ore-bearing granitic pegmatite information from remote sensing images in areas with low vegetation coverage.