| Popis: |
This study aims to draw a scientific framework for plotting soil erosion susceptibility in the Chittagong Hill Tracts of Bangladesh by comparing existing approaches. Data-driven machine learning techniques (including Classification and Regression Tree (CART), Artificial Neural Network (ANN), Support Vector Machine (SVM), and Random Forest (RF)) and a knowledge-based approach (AHP) are used in this study to pinpoint areas of Chittagong that are particularly susceptible to soil erosion while taking into account 18 soil erosion-regulating parameters. Furthermore, the effectiveness of the selected data-driven machine learning models and knowledge-based models was assessed by utilizing soil erosion and non-erosion sites. When evaluating the fidelity of each model using the ROC and AUC, the RF model was shown to be the most accurate and predictive. There is no poor performer among these models; all have AUCs greater than 67 % (RF = 0.86, ANN = 0.73, SVM = 0.67, CART = 0.67, and AHP = 0.82). According to the findings of the Random Forest model, approximately 71.55 percent of the area exhibited a moderate level of susceptibility to soil erosion. In relation to the land area, the high and low zones accounted for 16.91 percent and 11.54 percent, respectively. The specific area shares of 2256.25, 9548.08, and 1539.67 square kilometers were attributed to the high, moderate, and low danger zones, respectively. The best models' results after comparing models of data-driven and knowledge-based approaches can help to estimate soil erosion risk zones and provide insight into establishing appropriate policies to minimize this issue. In addition, the methods used in this research might be applicable to assessing the vulnerability and risk of soil erosion events in other areas. As they begin long-term planning to reduce soil erosion, local authorities and policymakers will find the study's results on practical policies and management options quite helpful. |
Full Text from ScienceDirect