Mapping of soil management zones by unsupervised classification of drone aerial images

Abstract

This study aimed to provide knowledge for the use of images from unmanned aerial vehicles (UAVs) for soil sampling. Typically, the diagnosis of agricultural fields requires soil samples from several locations to ensure representativeness. However, mostly the soil properties are spatially heterogeneous. Therefore, depending on the sampling point selection, the results obtained may not accurately reflect the field productivity. To overcome this sampling problem, this study used UAV images as supporting tools for sampling in soil diagnosis and investigated the relationship between the unsupervised classification results of UAV images and soil physicochemical properties in a 1 ha soybean field. The UAV observations were performed shortly after the soybean harvest. The pixel values of the multispectral and thermal infra-red images and a digital surface model were used as features for unsupervised classification by k-means++ clustering. After the smoothing process, the distribution of each cluster in the target crop field was mapped on a geographic information system. Moreover, this was compared to the analytical values of total carbon, total nitrogen, moisture content, particle size distribution, pH (H₂O), and electrical conductivity in 43 soil samples collected from the field to evaluate the feasibility of using aerial images in the selection of sampling points. The distribution of the obtained clusters corresponded to the trends in the soil physicochemical properties within the target field, even though soil analytical values were not learned as features. These results were consistent with the normalized difference vegetation index distribution during soybean growth. Overall, it suggests that unsupervised classification of UAV images can provide useful information for sampling in soil diagnosis.