To achieve pixel-level classification and identification of thunderstorm activity ranges, research on image segmentation technology of meteorological geostationary satellite images is conducted. Taking Guizhou Province and surrounding areas (24°-30°N, 103°-110°E) as an example, the radiation data of Fengyun geostationary satellite water vapour and long-wave infrared channels (6.25-13.5 μm) are selected as features. By integrating ground-based very low frequency/low frequency (VLF/LF) lightning monitoring and spaceborne Lightning Mapping Imager (LMI) data, labelled data is constructed to establish a deep learning dataset. The improved DeepLabv3+ semantic segmentation network along with added training strategies is used to identify the thunderstorm activity range in geostationary satellite images. The research results show that by adopting deep learning training strategies such as data augmentation, adaptive sampling in active learning, Combo Loss combination loss, and Ranger21 optimiser, the impact of thunderstorm small sample data training on network model performance can be effectively reduced, and the problem of data imbalance can be solved. When further comparing different backbone networks, including MobilenetV2, Xception, ResNet_101, ResNet_50, and HRNetV2-48 for feature extraction, it is found that MobilenetV2 has the fastest running speed while ResNet_101 has the best segmentation performance. In addition, by introducing the convolutional block attention module (CBAM) in the encoder and decoder, the model’s ability to learn target features and fuse information at all levels is greatly enhanced. As a result, both pixel accuracy and average intersection are significantly improved, further enhancing the model’s segmentation accuracy. Through extensive ablation experiments on the test dataset comparing SegNet, UNet, FCN, Lraspp, and the original DeepLabv3+ semantic segmentation network model, it is evident that the improved DeepLabv3+ model is superior to all other models. It achieves a pixel average accuracy of 96.82% and an average intersection over union (MIoU) of 76.93%. This not only showcases its superiority but also to a certain extent addresses the problem of high accuracy but low MIoU of the training model on the test set due to imbalanced sample data. This research extends the RGB three-channel data in image recognition to meteorological multi-dimensional data with more than three channels, aiming to mine thunderstorm characteristic information in satellite images more accurately and efficiently and lay a solid foundation for the next step of applying spatio-temporal recurrent neural networks to thunderstorm activity prediction. This study holds great promise for improving our understanding and prediction of thunderstorm activities.