ABSTRACT In the cool highlands agricultural area in the Kanto region in Japan, large-scale vegetable cultivation is taking place in sloping fields where Andosols are distributed. In some steeply sloping fields in the area, soil erosion has resulted in the loss of surface soil and its redeposition, causing heterogeneity of soil productivity. In this study, a high-resolution soil map (1 m resolution) was delineated using drone images and machine learning to understand the status of soil productivity in sloping vegetable fields. A digital elevation model (DEM) and orthoimages were created from the analysis of images taken by a drone. Then, 13 topographic index maps, such as slopes, were created from the DEM. The orthoimages were then converted to black and white images to quantify surface soil color. Based on the black and white images and topographic indices of the field, the distribution map of 1) organic carbon content of surface soil and 2) layer thickness of A horizon in the study area were delineated by the Regression-Kriging method. The Empirical Bayesian Kriging method was used to delineate maps of 3) gravel content in the soil profile (0–60 cm) and 4) depth to the gravel layer. Using the 13 topographic index maps and the maps from 1) to 4) as features, a predicted soil map was delineated using the random forest method with eight soil series groups as the map unit. Nine features were selected by the best-predicted model. High-Humic Cumulic Allophanic Andosols and Skeletal Cumulic Allophanic Andosols were generally covered on gentle slopes with low LS-Factor, which was calculated by slope and specific catchment area. Skeletal Low-humic Allophanic Andosols were mainly distributed on steep slopes with high LS-Factor and susceptible to soil erosion. A high-resolution soil map reflecting soil erosion was able to delineate using drone images and machine learning in Andosol's sloping upland field.