A study on Random Forest Algorithm in UAV Images for Cultivated Field Classification

• Main Authors: Tzu-Yen Huang, 黃子彥
• Other Authors: 蔡榮得
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/7274kp

碩士 === 國立中興大學 === 土木工程學系所 === 106 === The effectiveness of conventional aerial photography is affected by time and weather, whereas optical satellite imagery can be obstructed by obstacles such as clouds. Image collection using unmanned aerial vehicles (UAVs) has become crucial in recent years. It provides instant results, has high mobility and resolution, and is less affected by weather than conventional aerial photography and satellite imagery. UAVs have thus become a new type of photographic instrument. In this study, a high-resolution UAV camera was employed to capture images of Tuku Township in Yunlin County, Taiwan. Minimum Distance Classifier and the Random forest were used to classify the visible light band respectively. The experimental results show that the classification accuracy of Random forest is obviously better than Minimum Distance Classifier. The subsequent image will be added to the texture image in the visible light band and classified by Random forest. Texture information was added to high resolution UAV orthoimages to enhance the differences in spatial characters among the areas of various agricultural crops, thereby enhancing the accuracy of the high-resolution UAV image classification. Preliminary results suggested that this addition of texture information was indeed discovered to improve the accuracy of agricultural crop classification. Texture analysis was conducted using the grey-level co-occurrence matrix, and the six texture factors (homogeneity, contrast, angular second moment, dissimilarity, entropy, and correlation) were calculated. Various moving window sizes and texture factors were added to the raw images, and training sample areas were selected from the images. The areas were then classified through the use of Random Forest algorithm, which ensured high classification accuracy. According to the results, original bands with a 21×21 moving window achieved the optimal classification accuracy. The overall accuracy and Kappa value of image classification were 88.56% and 0.82, respectively, when only the raw RGB image was employed. After the texture information with a 21×21 moving window size was applied to the image, the accuracy and Kappa value increased to 94.22% and 0.91, respectively. Therefore, implementing the texture information in the image classification process did enhance the classification accuracy.