Image-based machine learning models for customized soil moisture management.

Crop growth can vary even under the same cultivation conditions, highlighting the limitations of conventional smart farming systems that apply uniform treatments to all crops. These average-based approaches often overlook individual plant needs shaped by microenvironments and physiological differences, resulting in inefficient resource use and reduced yields. While crop-specific management is important for improving productivity, there is a lack of non-invasive methods to monitor soil conditions at the individual plant level. This study presents an AI-based system that combines soil sensors and image analysis to support customized moisture management. Transplanted wild-simulated ginseng was used as a model crop. RGB images of the soil surface were collected with sensor data from different depths (3 cm, 10 cm, and 15 cm) to capture vertical moisture distribution. Several deep learning models were evaluated for predicting surface moisture, with DenseNet121 showing the highest accuracy (R² = 97.3%, RMSE = 4.14). For deeper soil layers, the random forest regression model achieved the best performance (R² = 90.6%, RMSE = 4.97), effectively capturing nonlinear moisture dynamics. These results demonstrate that surface image data can be used to estimate soil moisture non-invasively and enable data-driven, plant-specific crop management systems. This research provides a foundation for data-driven, customized, soil moisture management in smart farming. Future studies should focus on validating the model across diverse crops and soil types, and integrate additional spectral data to enhance its robustness and scalability.