Comparative Evaluation of Machine Learning Algorithms for Rice Terrace Extraction from RapidEye, Sentinel-2 and Landsat-8 Images

Abstract

Remote sensing images has been reported as valuable data to extract the rice terrace. However, most of these studies have been focused on high and very high spatial resolution remotely sensed images. In this paper, we investigate the capability of three medium resolution remote sensing data, namely, RapidEye, Sentinel-2, and Landsat-8 for rice terrace extraction. Moreover, both Pixel Based Image Analysis (PBIA) and Object Based Image Analysis (OBIA) are utilized to classify rice terrace using robust machine learning classifiers, namely, Multilayer Perceptron, Random Forest, and Support Vector Machine algorithms. All three remote sensing data provide high accuracies of rice terrace classification with PBIA, with accuracies ranging from 90.3% to 92%. OBIA does not perform as well as PBIA. In general, the accuracy of OBIA decreases when the threshold of segmentation increases. OBIA applied RapidEye provides accuracy higher than 85%. Sentinel-2 shows lower accuracy at above 80%. Landsat-8 image shows the least accuracy below 75% at higher threshold levels. Although the classification accuracy for OBIA shows dependence on spatial resolution of remote sensing images, the output results for the three classifiers do no show significant difference except in the ability to distinguish smaller patches of rice terrace in images of higher resolution. Based on the results, PBIA is considered to offer a simple and more accurate method for rice terrace extraction in the study area.