Abstract
To adapt to precision agriculture, the current agricultural soil map’s soil information must be updated. Digital Soil Mapping (DSM) updated soil maps using field data and environmental covariates, as well as data mining and classification techniques, and then generated spatially high-resolution maps. The goal of this study was to create a high-resolution, high accuracy agricultural soil map in paddy fields where land improvement had previously occurred, and to predicted soil group distribution using machine learning with the Random Forest.
The model accuracy and Kappa coefficient of test data obtained from machine learning were 62.5％ and 0.46, respectively. These values were nearly identical to previous studies that used relatively small point data, as did this study. While the Gley Lowland soil group predicted with high accuracy, the Lowland Paddy soil group and the Gray Lowland soil group were not clearly separated. The Confusion Index indicated that the predicted soil map was uncertain at 0.64, and four out of the five soil groups had high values. This high level uncertainty was attributed to a small sample size of minor classes and the difficulty in distinguishing between the Lowland Paddy soil group and the Gley Lowland soil group. In predicted soil group map, the distribution area of the Gley Lowland soil group was reduced while that of Gray Lowland soil group was increased. Despite differences in accuracy between soil groups, the predicted agricultural soil map for paddy fields could reflect the current situation and had a high resolution by using DSM.
