Knowledge-constrained deep learning for large-scale landform classification in the Loess Plateau

Landforms are a fundamental component of physical geography and form the foundation of the Earth's surface system. Accurate landform classification is a core task of geomorphological research and has broader implications for related studies. However, the complex morphological characteristics of Earth's surface, shaped by various internal and external forces, make it challenging to delineate landform boundaries on a large scale. In this study, we employed a deep learning method constrained by knowledge to classify landforms in the Loess Plateau, China, within 100.90°–114.55° E and 33.72°–41.27° N. First, the knowledge of loess landforms was extracted based on hydrological and binary characteristic of loess terrain. Then, landform labels, along with Digital Elevation Models (DEMs) and images, were used to train the deep learning model. Third, the trained model was then applied across the study area, where loess hills and ridges were classified. Finally, the classified loess tablelands, ridges, hills and gullies were merged to achieve a comprehensive loess landform classification. The accuracy of the classification was evaluated using 5000 sample points, yielding accuracies of 88.01 % for tablelands, 76.34 % for ridges, 72.00 % for hills, and 91.61 % for gullies, with an overall accuracy of 82.86 %. Using knowledge as a constraint improves classification accuracy by 13 % compared to not using it. A comparative analysis with the Modified U-Net method showed that our proposed method produced more accurate landform boundaries and outperformed the existing approach. In addition, the classification results were statistically analyzed within ecological regions, revealing that landform proportions can partially reflect the geomorphic development characteristics of each region. In future research, this knowledge-constrained classification method is expected to be applicable to a global scale.