Multi-Class Semantic Segmentation of Green Areas from Satellite Images Using Transformer-Based Models
7. Uluslararası Uygulamalı İstatistik Kongresi (UYİK 2026), İstanbul, Türkiye, 11 - 13 Mayıs 2026, ss.261, (Özet Bildiri)
- Yayın Türü: Bildiri / Özet Bildiri
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.261
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Kayseri Üniversitesi Adresli: Evet
Özet
Accurate identification and segmentation of green areas from satellite imagery is critical for environmental management, urban planning, and sustainable development. Accurate determination and hectare calculations of green areas are particularly important in monitoring natural events such as forest fires, land-use changes, and climate-related disasters. Such analyses directly accelerate damage assessment, risk evaluation, and response processes. However, traditional methods are insufficient for modeling complex spatial structures in high-resolution satellite data, highlighting the need for automated analysis. This study uses a dataset of 8028 satellite images of Kayseri province, manually labeled as polygons using the MakeSense platform. Green areas were divided into six classes: dense vegetation, sparse vegetation, wooded areas, object surroundings, and similar categories. Polygon labels were converted into pixel-level masks and treated as a semantic segmentation problem. In this study, the convolutional neural network (CNN)-based U-Net and DeepLabV3+ models were comparatively evaluated with the transformer-based SegFormer architecture. Experimental studies conducted on real satellite images resulted in performance values of mIoU: 0.508, Dice: 0.661, Pixel Accuracy: 0.732, Precision: 0.675, Recall: 0.665, and F1-score: 0.661 obtained with the U-Net model. Class-based evaluations showed higher accuracy in dense green areas and wooded regions, while performance was relatively lower in sparse and complex classes. This highlights the difficulty of multi-class green area segmentation and the impact of inter-class similarity on model performance. The findings demonstrate that deep learning-based approaches offer an effective solution for automatically detecting and classifying green areas in satellite images. The proposed method holds significant potential, particularly for the rapid and accurate analysis of damage after forest fires, monitoring of green area loss, and calculation of hectare-based area. In this respect, the study has the potential to contribute to large-scale environmental monitoring and decision support systems. Ongoing studies aim to improve class-based performance, address data imbalances, and compare different model architectures.