Identifying reservoirs in northwestern Iran using high-resolution satellite images and deep learning

Abstract

Reservoirs play a critical role in terrestrial hydrological systems, but the contribution of small and medium-sized ones is rarely considered and recorded. Particularly in developing countries, there is a rapid increase of such reservoirs due to their quick construction. Accurately identifying these reservoirs is important for understanding social and economic development, but distinguishing them from other natural water bodies poses a significant challenge. Thus, we propose a method to identify reservoirs using high-resolution satellite images and deep learning algorithms. We trained models with various parameters and network structures, and You Only Look Once version 7 (YOLOv7) outperformed other algorithms and was selected to build the final model. The method was applied to a region in northwestern Iran, characterized by an abundance of reservoirs of various sizes. Evaluation results indicated that our method was highly accurate (mAP: 0.79, Recall: 0.76, Precision: 0.82). The YOLOv7 model was able to automatically identify 650 reservoirs in the entire study region, indicating that the proposed method can accurately detect reservoirs and has the potential for broader-scale surveys, even global applications.

Keywords

• Reservoir
• deep learning
• object detection
• Iran