Automatic Landform Recognition, Extraction, and Classification using Kernel Pattern Modeling

Shirani Kourosh; Solhi Sina; Pasandi Mehrdad

摘要

Detection of terrain features more quickly and accurately is crucial in geosciences for the extracting and classifying landforms. Part of the local relief attributes would be lost in employing the conventional TPI (Topographic Position Index)-based methods to recognize terrain characteristics (landform recognition, extraction, and classification) due to use of the central tendency index (mean index). As a result, the automated recognition, extraction, and classification of terrain features (summit/ridge and pit/drainage) were performed in this research using the kernel pattern modeling based on digital elevation model (DEM) in the raster grid structure. Accordingly, three novel TPI-based algorithms including simple multi-level recognition system (SMRS), complex multi-level recognition system (CMRS), and central position recognition system (CPRS) were developed to recognize the terrain similarity to summit/ridgeline or terrain convex surfaces (and/or deviation from the pit/drainage or terrain concave surfaces). The algorithms were formulated and programmed using Python programming language and integrated in a software package as well. The results show that the SMRS algorithm tends to have more binary gradation (higher contrast) in smaller dimensions of the moving window, being problematic in some geovisual and cartographical applications. Output of the CPRS algorithm shows more continuity and better performance in the extraction of summit points in a vector data model format. All the three algorithms apply greater degree of generalization to the results as size of the moving window becomes larger. The accuracy assessment, sensitivity analysis, and possible sources of the errors were evaluated. Finally feature extraction and landform classification performed based on output results of the algorithms. The accuracy assessment and validation of the models were examined compared to the TPI and DEV (deviation from mean elevation) models by means of the D8 (for drainage path extraction) and inverted D8 (for ridgeline extraction) algorithms. Accordingly, all the developed algorithms perform better than the conventional method of TPI. The SMRS (88.87%) and then CPRS (64.53%) performed better than the other algorithms and TPI (56.83%). The DEV performance (59.52%) is similar to CMRS (59.19%) and better than TPI. Based on the temporal sensitivity analysis, CMRS and SMRS are the most and least sensitive algorithms to the moving window size and spatial resolution variations, respectively. The possible error sources (edge effect, kernel local pattern, and point vectorization) were evaluated for the algorithms. Then, feature extraction (including summit/ridge or pit/drainage) was implemented on outputs of the algorithms. The CMRS and CPRS models displayed better performance in the vector point extraction compared to the other algorithms. A new landform classification system including 25 classes was also designed based on outputs of the developed algorithms combined with the terrain elevation variable. Landform classification maps generated based on the SMRS algorithm is visually different from what provided by the other algorithms and the CPRS landform classification performed better in distinguishing land units. The kernel size variations can modify landform recognition and classification scale in such a way that whatever size of the moving window is larger, the landform unit generalization is greater.

Automatic Landform Recognition, Extraction, and Classification using Kernel Pattern Modeling

摘要

著录项

相关主题

期刊订阅