Automatic identification of soil and water conservation measures from centimeter-resolution unmanned aerial vehicle imagery

Abstract

The emergence of unmanned aerial vehicles (UAVs) has resulted in a new era of remote sensing, especially for applications requiring accurate image classification. This paper describes an automatic method for identifying soil and water conservation (SWC) measures from the centimeter-resolution imagery of UAVs using an object-based image analysis (OBIA) approach and machine learning models, and a support vector machine (SVM) model. The study area is located in Yitong County of Jilin Province, in the black soil region of northeast China. There are four frequently used SWC measures, including ecologically restored forests, ecologically restored grasslands, contour ridges, and ridge belts. A block of red, green, and blue (RGB) images was obtained on May 26, 2018, from the study area, and the images were processed to generate a high-resolution detailed orthomosaic image (5 cm). Several features were derived from the UAV image, including color indices, terrain, texture, shape, and edge information, which were incorporated in the OBIA method. Three color indices were selected to derive vegetation information from the study area, including excess green, normalized green-red difference index, and excess green minus excess red. Then a number of samples were selected to improve the classification results using the SVM model. The results showed that the overall accuracy and kappa coefficient were 91.20% and 0.87, respectively. Thus the OBIA method was effective in identifying, classifying, and describing detailed SWC measures. However, some objects, such as the stages and furrows in the contour ridge measures, were not identified in some regions using the OBIA method, but the machine-learning model SVM resolved this problem. This study shows the substantial benefits of centimeter-scale UAV imagery over satellite and airborne remote sensing and demonstrates the potential of low-cost RGB cameras for the accurate identification of different SWC measures and the detailed derivation of the shape parameters of linear SWC measures.