Semantic Segmentation of Green Areas in Satellite Images Using Deep Learning Models

<jats:p>Accurate identification and segmentation of green areas from satellite imagery are important for environmental management, urban planning, and sustainable development. Accurate determination and calculation of the hectares of green areas are particularly important for monitoring natural events such as forest fires, land-use changes, and climate-related disasters. Such analysis directly expedites the processes of damage assessment, risk evaluation, and response. However, conventional approaches are inadequate for modeling complex spatial structures in high-resolution satellite data that require automated analysis. In this study, a dataset consisting of 8028 satellite images of Kayseri province, manually labelled as polygons on the MakeSense platform, is used. Six classes were identified for green areas: dense vegetation, light vegetation, wooded areas, object surroundings, and similar classes. We converted polygon labels to pixel-level masks and treated the task as a semantic segmentation problem. In this work, a comparative evaluation of the U-Net and DeepLabV3+ models, both based on the convolutional neural network (CNN), was carried out against the transformer-based SegFormer architecture. The performance values obtained from the experimental studies on real satellite images indicate that the reorganized 4-class structure achieved better results than the initial 6-class structure. In the 6-class structure, the best mIoU value was obtained by U-Net and DeepLabV3+ with 0.51, while the highest Dice and F1-score values were achieved by U-Net and SegFormer with 0.66. DeepLabV3+ achieved the best Pixel Accuracy and Precision, with 0.74 and 0.70, respectively. In contrast, the 4-class structure yielded higher, more balanced performance across all models. DeepLabV3+ and SegFormer achieved the highest mIoU value of 0.57, while SegFormer obtained the best Dice, Recall, and F1-score values with 0.72. Pixel Accuracy reached 0.76 for all models in the 4-class structure. These findings show that reducing the class number from six to four improved the segmentation performance by decreasing confusion among visually similar green-area classes. The results indicate that deep learning-based approaches provide an effective solution for automatically detecting and classifying green areas in satellite images. The proposed method holds significant potential, particularly for rapid, accurate analysis of damage after forest fires and for monitoring green area loss. In this respect, the study has the potential to contribute to large-scale environmental monitoring and decision support systems.</jats:p>