Comparison of farmland crop classification methods based on visible light images of unmanned aerial vehicles

Abstract: Xinjiang is the main cultivated area of cotton in China. It is very important to obtain data of planting area every year. Access to accurate, large-scale farmland planting information is also the basis for precision agriculture. Many scholars at home and abroad have carried out relevant research. They use different methods to extract information on crop cultivation at different time, but the data are relatively slow to update. Small-scale unmanned aerial vehicle (UAV) remote sensing with low cost, low risk, high temporal and spatial resolution and other characteristics, is very suitable for rapid extraction of crop information and crop classification. Color space conversion, texture analysis and color index and other methods can effectively enhance and tap the potential information of the image, which is helpful to the image classification. In this paper, we used the UAV images acquired in September 2016 to carry out the extraction of crop types in some farmland of the Eighth Division of Xinjiang Corps. Through the conversion of color space and the processing of different texture filtering, the texture features of the objects in the image could be solved satisfactorily, which could better solve the phenomenon of the same spectrum and the heterogeneity of the same kind, and improve the recognition accuracy of the feature. First, the color space conversion and the gray level co-occurrence texture filtering were carried out, and 27 color and texture features were obtained. By comparing the coefficient of variation and the difference coefficients of the 3 color features and 24 texture features, we believed that the brightness, saturation and red second order moment could be used as the optimal classification characteristics. Secondly, due to the lack of near-infrared band data, only the visible light red band and green band were used to build color index to extract vegetation information. This paper calculated the excess green index (EXG) and the visible-band difference vegetation index (VDVI) of the image. By comparing the threshold of the EXG and the VDVI of the gray scale image, it was determined that the EXG could effectively distinguish the different crop types. Finally, the visual interpretation results were compared with the results based on the combination of color texture feature classification and color index classification. The results showed that the measured areas of cotton, maize and grape were 0.490 1, 0.042 1 and 0.143 2 km2, respectively. The areas of the 3 crop types based on color and texture features were 0.454 8, 0.044 1, and 0.139 8 km2, respectively, and the areas of the 3 crop types based on the color index feature were 0.547 7, 0.039 8 and 0.099 4 km2, respectively. The error values for the former method when applied to the classification of cotton, maize and grape were 7.2%, 4.75% and 2.37%, respectively. The results showed that the extraction accuracy of crop type based on color and texture feature was higher than that of color index. However, both of the methods are based on single pixel. For the same kind of crop, due to differences in spectral characteristics, some internal area is included into other crop types, and there is a significant salt and pepper effect. Related data post-processing for the wrongly classified small patch should be performed to improve the classification accuracy. In some researches, the object-oriented classification method on the basis of calculating the correlation index can better solve the problem of discontinuities and incompleteness of the same object based on the pixel classification method. At the same time, the study of farmland information extraction of larger scale UAV data cannot be carried out in this paper, which will be further explored in the follow-up study.