High-Resolution (0.9 m) terrace mapping in Low-Latitude hilly regions using deep learning and Area–Slope Denoising: A case study from Guangdong Province, China

Terrace construction serves as a crucial human intervention for improving slope productivity and preventing soil erosion, especially in low-latitude hilly regions with heavy and concentrated rainfall. However, existing terrace mapping methods are largely limited in such areas. In this study, we produced, to the best of our knowledge, the first high-resolution (0.9 m) terrace distribution map for low-latitude hilly regions by integrating deep learning with slope–area threshold denoising, using Guangdong Province—a typical low-latitude hilly region with abundant and concentrated rainfall and extensive terrace distribution—as a case study. The results showed that the optimal 0.9 m resolution achieved 93.34% overall accuracy (OA), 79.18% F1-score (harmonic mean of precision and recall), and 65.53% intersection over union (IoU) on our Guangdong terrace dataset, processing 400 km² in 89 min, showing an excellent accuracy–speed trade-off. Statistical analysis revealed that terraces in Guangdong are predominantly small (4725 m² on average) and gently sloped (8.85°). Provincial validation confirmed superior performance, with a user's accuracy of 80.38%. Spatially, terraces are mainly clustered in inland hilly areas, especially in northern Guangdong, with additional clusters in the peripheral zones of the Pearl River Delta and in the western and eastern regions. The cities with the largest terrace coverage are Qingyuan (801.98 km²), Zhaoqing (733.11 km²), and Shaoguan (701.53 km²). This high-precision dataset supports agricultural management, soil erosion risk assessment, ecological conservation, and rural revitalisation, while also offering a transferable framework for mapping terrain features in complex landscapes worldwide using remote sensing imagery.