For the management of the variability of soil fertility in agricultural fields, the estimation of soil physical and chemical distribution using aerial images taken by UAV (Unmanned Aerial Vehicle) has been attracting attention. As a related study, Morishita and Ishitsuka (2020) proposed a method to accurately estimate soil moisture distribution from UAV images by constructing a machine learning model based on a data augmentation method for dozens of soil moisture data measured in the field. This method can be applied to other soil indices for use in field management and fertilizer design. In this study, we attempted to apply machine learning (random forest regression) in estimating the distribution of soil physical and chemical properties using UAV images by augmenting the analysis data of real samples, and verified the effectiveness of the method by comparing it with the conventional method (multiple regression). Specifically, a random forest regression model was constructed for soil physicochemical analysis (soil texture, total carbon, total nitrogen, pH (H₂O), and electrical conductivity) with 36 soil samples collected within a single field, using the pixel values of multispectral, thermal infrared, and Digital Surface Model (DSM) images observed by UAV as explanatory variables. The effectiveness of the machine learning application was tested by comparing the estimation accuracy of random forest regression models and multiple regression models.　The results showed that the spatial estimation by random forest regression (Coefficient of determination for test data: r-R² = 0.54 ~ 0.85) attained clearly higher accuracy than that of the multiple regression model (m-R² = 0.19 ~ 0.49) for each soil analysis value. This suggests that the application of the machine learning algorithm with ground-truth data augmentation is effective in the spatial estimation of soil physicochemical properties using UAV images. Furthermore, the correlation between the growth conditions during the cropping period and the estimated distribution of each soil characteristic also allowed us to discuss the factors of uneven crop growth. This suggested that the method for the spatial estimation of soil properties used in this study can contribute to more efficient field management.

To understand the labor-saving method of developing a high-accuracy three-dimensional model with reduced doming effect using an aerial photograph from a real time kinematic-unmanned aerial vehicle (RTK-UAV) without using ground control point (GCP), we examined the relationship between the camera angle, data processing method, degree of doming, and vertical accuracy of the model. Consequently, as the camera angle was tilted more, the effect of reduced doming effect and vertical error increased. Moreover, by optimizing the camera position accuracy according to the RTK, reducing the occurrence of the doming effect was almost completely possible. Furthermore, by optimizing the lens distortion model according to the RTK, reducing the vertical error within 2 cm was possible. To ensure operational efficiency, a combination of -70 ° of the camera angle, using “load camera location accuracy from XMP meta data" function of the photogrammetry software, and using "fit additional correction" function during lens calibration processing in the software was optimal. Furthermore, extremely high doming and vertical error reduction effects can be obtained even with a conventional method, which combines GCPs with non-RTK-UAV, by arranging GCP at the center of the measurement range.