A detection-screening framework for karez (ancient underground irrigation system) using deep learning and geospatial analysis

Abstract

Karez, an ancient engineering marvel, utilizes gravity to transport underground water to the surface without external power. Typically, a karez comprises numerous shafts (vertical wells), and traditional mapping methods are both time-consuming and labor-intensive. To address these challenges, this study developed an integrated detection-screening framework for karez mapping. The karez shafts were detected by using high spatial resolution satellite imagery and deep learning architectures (Faster-RCNN, SSD, YoloV3, and MMDetection). Subsequently, a directed fan-shaped buffering method, combined with hierarchical clustering, was introduced to filter out misidentified shaft-like structures. Results showed that the MMDetection outperformed other models, achieving a mean average precision (mAP50-95) of 0.833. Field validation confirmed that the screening methods eliminated 90.20% of false shaft detections. This study has obtained the largest number of karez shafts to date in the study area, while providing a transferable technical framework for global applications in cultural heritage documentation and arid land water management.