Prediction and Spatial Mapping of Soil Properties Using UAV-based Multispectral Imagery and Random Forest Regression

Modern agriculture faces the critical challenge of enhancing crop productivity due to the decline in the farming population and the increasing demand for food. In particular, the precise monitoring of soil properties, which significantly influence crop yield and quality, plays a crucial role in achieving sustainable agriculture. However, traditional methods based on field soil sampling and laboratory analysis are time-consuming and costly. To address these limitations, this study developed a model that predicts and spatially maps key soil properties (Total Nitrogen, Soil Organic Matter, and Silicon Dioxide) by integrating UAV-based multispectral imagery with machine learning techniques, specifically Random Forest regression. Field experiments were conducted in four salt-affected paddy fields located in Hwaseong, Gyeonggi, South Korea. Thirty soil samples were collected from each field for laboratory analysis, and high-resolution multispectral images were simultaneously acquired using UAV. Reflectance bands and a range of soil- and vegetation-related spectral indices were extracted from the imagery and used as input variables for model development. The model achieved the highest prediction accuracy when both reflectance bands and spectral indices were combined, with validation R² values reaching up to 0.513 for SOM and 0.449 for TN (Field 2). Furthermore, comparison between Kriging-interpolated laboratory measurements and Random Forest-predicted maps demonstrated a strong similarity in spatial distribution patterns, validating the feasibility of UAV-based, non-destructive soil property prediction. The results of this study provide a practical foundation for the development of site-specific fertilization strategies and improved soil management in precision agriculture. Moreover, the predictive performance of the model is expected to be further enhanced by incorporating additional auxiliary variables such as soil physical properties and topographic information, as well as by applying deep learning-based regression techniques in future research.