PNTM-CNN: an approach for saddle-point extraction integrating positive–negative terrain method and multiscale fusion CNN model

Saddle-point extraction is essential for accurately identifying topographic features and landforms and conducting geomorphological mapping. However, the widely used positive–negative terrain method (PNTM) is often plagued by a substantial number of false saddle points, a prevalent issue in many extraction techniques. To address this challenge, this study presents a novel model that combines the PNTM with a convolutional neural network (CNN) called PNTM-CNN. In this approach, candidate saddle points are first identified using the PNTM and then refined using a CNN that integrates multiscale topographic features. The experimental results indicate that the PNTM-CNN model, which leverages four scales of features (elevation, aspect, curvature, slope, and hillshade), effectively reduces the occurrence of false saddle points, achieving a precision of 89%, a recall of 83%, and an F1 score of 85%. This performance significantly exceeds that of the traditional moving window analysis and topological association methods. Although the automation level of the PNTM-CNN model requires improvement, the integration of deep learning methods offers new insights for addressing complex topographic feature extraction challenges and shows a promising application potential.