Deep Learning-Based Water Area Segmentation and Water Coverage Ratio Analysis in Satellite Images

Creative Commons License

Kılıçkaya F. N., Taşyürek M., Öztürk C.

BRITISH CONGRESS-VI, London, İngiltere, 18 - 21 Haziran 2026, ss.128, (Özet Bildiri)

  • Yayın Türü: Bildiri / Özet Bildiri
  • Basıldığı Şehir: London
  • Basıldığı Ülke: İngiltere
  • Sayfa Sayıları: ss.128
  • Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
  • Kayseri Üniversitesi Adresli: Evet

Özet

Remote sensing-based monitoring of water resources is critical for drought management, flood risk analysis, land-use planning, and sustainable environmental decision-making. This study aims to automatically extract water surfaces from satellite images obtained from the Kayseri Metropolitan Municipality and to quantitatively calculate the water coverage rate for each image. In this context, water areas were manually labeled using polygon-based annotations on the MakeSense.ai platform, and the labels were exported in COCO format. Binary water masks were generated, and the dataset was split into training, validation, and test subsets. Instead of an image-level classification approach, this study adopted a pixel-level semantic segmentation strategy. Accordingly, the CNN-based U-Net and DeepLabV3+ models and the transformer-based SegFormer model were comparatively evaluated for water surface segmentation. Model performance was assessed using Intersection over Union (IoU), Dice coefficient, precision, recall, F1-score, pixel accuracy, Mean Absolute Error (MAE), and Root Mean Square Error (RMSE). Experimental results demonstrated that SegFormer achieved the highest segmentation performance on the test set, with an IoU of 0.8740, Dice coefficient of 0.8964, precision of 0.8987, recall of 0.8983, F1-score of 0.8964, and pixel accuracy of 0.9740. U-Net achieved results close to SegFormer, with 0.8662 IoU and 0.8932 F1-score, whereas DeepLabV3+ showed lower performance, with 0.8591 IoU and 0.8862 F1-score. In estimating water surface coverage, SegFormer also achieved the lowest error values, with an MAE of 0.0159 and an RMSE of 0.0374. Based on the water ratios calculated from segmentation outputs, images were categorized as “no water”, “partial water”, and “dense water”. In this stage, U-Net achieved the highest overall accuracy of 0.9476, while SegFormer provided the best class-balanced performance with a macro F1-score of 0.8811. Overall, the findings indicate that deep learning-based semantic segmentation methods can effectively extract water surfaces from local-scale satellite images and support municipal-scale water resource monitoring.