Fusion of Transformers and Unet for Advanced Multi-Scale Aerial Image Segmentation

This paper introduces a sophisticated machine learning method for segmenting satellite images utilising proprietary image data generators and trustworthy data augmentation. A range of augmentations, including random trimming, rotation, contrast changes, and the CLAHE technique, are applied using the Albumentations package to improve model generalisation and lessen overfitting. In order to handle big datasets, new data processors are created that produce frame bits and match mask labels for effective training. The segmentation skills of U-Net are combined with the feature extraction strength of Inception-Res in the InceptionResNetV2-UNet architecture. The model's predictions are visualised using a unique decoding technique, while mask labels are encoded using one-hot encoding. The approach generalises effectively to unseen satellite photos, as seen by the findings, which provide extremely accurate segmentation maps. This method is appropriate for large-scale sensor applications as data augmentation and proprietary generators greatly increase segmentation accuracy and scalability. The model outperformed baseline techniques in segmentation, demonstrating great accuracy and generalisation on unknown data.