Image segmentation for LULC classification using SegFormer and ADE20K mapping

This paper discusses the development of a semantic segmentation pipeline for land use and land cover (LULC) mapping using a Transformer-based deep learning model called SegFormer. The study aims to create efficient and accurate thematic maps from drone images by using a pretrained SegFormer model that was originally trained on the ADE20K dataset. The 150 ADE20K class labels are simplified into four main LULC categories: buildings, vegetation, roads, and water bodies. By choosing these classes, we can avoid the need for manually annotated ground truth data. To improve prediction dependability, we use test-time augmentation (TTA) with horizontal flipping. The output includes a color-coded thematic map overlaid on the original drone image, along with class-wise pixel counts, a confusion matrix, and a coverage-based weighted F1-score. The study evaluated the model under different segmentation scenarios. The model achieved individual class-wise accuracies of 77.79%, 82.85%, 88.25%, and 93.68%, with corresponding F1-scores of 0.2481, 0.4277, 0.2631, and 0.2619, respectively. The average overall accuracy is 88.64%, with a mean F1-score of 0.3214. These results show that the SegFormer-based model provides a practical and scalable solution for near real-time LULC segmentation, especially in cases with little or no ground truth annotations.