Deep learning with geographical post-processing optimization: an integrated framework for detecting qanat activity states

Abstract

As ancient underground water systems that sustained civilizations in arid regions for millennia, qanats represent both remarkable hydraulic heritage and vital water sources, with the Persian Qanat (inscribed on the World Heritage List in 2016) requiring dynamic monitoring for effective protection and management. This study overcomes limitations of prior spatial-distribution-focused research by constructing the first multi-region annotated dataset from very high-resolution resolution Google Earth satellite imagery across Iran, Afghanistan, Morocco and China, classifying 8,587 active and 17,383 inactive qanat samples. Our YOLO11-based model (enhanced with C3k2 backbone and C2PSA attention) integrates a novel post-processing framework where DBSCAN clustering removed 90.8% of outliers – collectively achieving 97.16% precision (9.5% improvement over baseline) and 76.56% recall. Applied to 11 Persian Qanat World Heritage Sites, the system identified 41,781 shafts in 889 qanats, including 15,742 active and 26,039 inactive qanats, revealing key patterns: 6/km² density, 169 m (SD = 46.3 m) spacing, and 95% occurrence in bare/sparsely vegetated areas on gentle slopes (mean 2.5°). This high-precision dataset enables prioritized conservation of inactive qanats as cultural relics and sustainable management of active systems, demonstrating how AI-geospatial integration can revolutionize archaeological monitoring in arid landscapes.