Automated ridgeline recognition, using Kernel neighborhood pattern analysis

 Landform refers to any physical feature of the surface with a recognizable structure and shape. Landform elements and structural forms of the terrain surface could, directly and indirectly, drive many other environmental variables. Numerical representation of the surface and uneven pattern of the earth is a common topic in geographical, geomorphological, geological, and geophysical hazard mapping as well as sea-bed exploration. The combination of the earth and computer science with mathematics and geomorphometric engineering interacts with discrete and continuous landforms. Geomorphometry dates back to about 150 years ago and the work of Alexander von Humboldt and geomorphologists, and today with the revolution in computer science and especially digital computer models is developing rapidly. Detection and classification of landforms are of interest to GIS developers, geoscientists, and geomorphometry researchers. In this way, the desired work units are extracted with higher speed and accuracy and used in the form of vector and raster maps. Existing approaches are mainly based on height, terrain derivative, gradient, curvature, flow direction, slope position, morphometric indices, and the like. Also, less attention has been paid to the challenge of matching the diagnostic scale with the Landform scale, and most models have this shortcoming. On the other hand, less attention has been paid to the possibility of vectorization output results and also to the analysis of sensitivity and temporal response algorithms to machine processing. In this research, we attempted to recover and resolve the mentioned shortcoming and problems in the previous works. In this research, using basic algorithms of raster analysis and coding, new methods and algorithms for the automatic detection of landforms have been developed. Focal raster analysis is also emphasized and the moving window technique is used to implement the algorithms. Facing the scale challenge, sensitivity analysis, and the response algorithms to input changes as well as accuracy assessment are other aspects that have been addressed in this research.

 In this study, the Digital Surface Model (DSM) published by the Japan Space Agency in May and October 2015 with a horizontal resolution of about 30 meters was used to work on the topography of the region. These data are obtained from ALOS satellite images. This database is based on DSM data (5m network version) 3D topography, one of the most accurate elevation data on a global scale. The digital elevation model was transformed into a matrix structure using a Python coding environment. Then, raster analysis was implemented using the moving window technique. The moving window algorithm was coded in a way that the dimensions of the moving window could be freely determined and changed. In proportion to the size of the moving window, some adaptive algorithms are implemented to automatically correct and organize the edge effect in proportion to the size of the moving window. In this study, automatic landform detection was performed using spatial analysis of kernel patterns in the raster grid of digital elevation models and the results were presented in the form of three algorithms applied in the detection of topographic peaks and ridges. These algorithms include Multilevel Mean Summit Recognition Algorithm (MLMSR), Complex Multilevel Summit Recognition Algorithm (CMLSR), and Single Point Summit Recognition (SPSR). Each of these three algorithms was first conceptually designed and then coded and executed using the Python programming language. In the next step, the sources of error and specific scenarios of the algorithms were examined. The sensitivity of each algorithm related to the dimensions of the moving window, the resolution, and the size of the raster file, was evaluated, and finally, the accuracy and validation of the three models, using reference layers that were manually prepared and plotted, were assessed. All the procedures were designed in a way that could easily be implemented in an official software and were completely compatible with the structure of machinery processing. Also, being automatic and working on different platforms where one of our priorities.

 In the automatic detection of peaks and ridges using a digital terrain model, kernel spatial pattern analysis was used. In this regard, three proposed algorithms in this field were designed, coded, and executed. The output results of each of the algorithms were presented in the form of a raster and vector data model. Accuracy and sensitivity assessments were performed by considering changes in moving window size, resolution, and raster grid size (row x column) for each of the algorithms. The MLMSR algorithm tends to be in a more binary result in the lower dimensions of the moving window, while the CMLSR and SPSR algorithms do not. In all algorithms, increasing the size of the moving window causes a more generalization ratio. CMLSR and SPSR algorithms are more suitable for cartographic and visual purposes due to the higher degree of grading in the results. Regarding the temporal performance (Runtime) or sensitivity to input changes, the SPSR algorithm performs better. This is especially important when the input file size (number of rows and columns) is large. According to the results of validation and accuracy evaluation, MLMSR and SPSR had better performance than, the CMLSR algorithm. Python programming language has been widely used in the design and implementation of all algorithms, as well as in the field of sensitivity evaluation and validation. Totally more than 500 lines of codes were done for this purpose. All algorithms are automated and are able to execute and store results in raster and vector format using machine processing.

 The results show that the MLMSR algorithm in smaller dimensions of the moving window is tending to more binary results, which is problematic in some graphical and cartographic applications, but the CMLSR and SPSR algorithms showed more gradual trends in their outputs and so, they performed better in this respect. Researchers who intend to study and develop in this field are advised to focus on adaptive algorithms and optimize the dimensions of the moving window in relation to the volume of input information and so, in this way, they increase the flexibility of algorithms in relation to input changes.