Performance Evaluation of Support Vector Machine and Random Forest Techniques for Land Use-Land Cover Classification—A Case Study on a Mili Scale Agricultural Watershed, Tadepalligudem, India

Land Use-Land Cover Mapping, obtained at a specific time, plays a crucial role in planning and monitoring both regional and global surfaces of the earth. For monitoring and analysis of small-scale agricultural watersheds with an area of less than 10,000 ha, a Land Use-Land Cover (LULC) classification that is nearly accurate is required. With the availability of Sentinel-2 datasets with spatial and temporal resolutions of 10 m and 5 days, LULC mapping and monitoring can be improved. However, conventional methods of classification are unable to demonstrate real features. In recent years, the introduction of machine learning techniques has made it possible to classify LULC effectively and efficiently, particularly for watersheds on the milli-scale. Therefore, in the present study, LULC maps are obtained for an agricultural watershed in Tadepalligudem, West Godavari district, Andhra Pradesh, using supervised Machine Learning methods. To highlight the importance of machine learning in LULC classification for a milli-scale watershed, Support Vector Machines (SVM) and Random Forests (RF) are compared with conventional Maximum Likelihood Classification methods. LULC was calculated using Sentinel-2 data acquired from the United States Geological Survey for three months in 2020. The field observations detected a total of four classes, and the training datasets were obtained using both visual inspection and field observations. Using Overall Accuracy, Kappa Coefficient, and R-Squared values, the accuracy of the LULC maps relative to the ground truth is determined. The study reveals that the overall accuracy of SVM and RF is 87.5 2.00 and 85.5 2.00, respectively, whereas it is 70.87 1.80 for Maximum Likelihood classification. The mean Kappa coefficient values for SVM and RF are 0.86 and 0.85, respectively, while for maximum likelihood classification, it is 0.67. The observed average R-squared value for the support vector machine is 0.67, whereas it is 0.69 for the RF. In addition, it is observed that SVM performs admirably among the three classification methods.