Convolutional neural network approach for estimating surface soil moisture from ground observation images

抄録

　Soil moisture is a key agricultural variable that influences both crop growth and field workability. This study develops a convolutional neural network (CNN) model capable of estimating surface soil moisture from RGB images captured under actual field conditions. The experiment was conducted on bare soil in Matsudo, Chiba, Japan, where time-series images of a 3 m×2 m field plot were collected over five consecutive days. In total, 98 surface images were captured using a consumer-grade trail camera and cropped into 1,960 subplot images corresponding to the EC-5 soil moisture sensors installed at a depth of 2 cm. Data augmentation techniques, including rotation, flipping, and brightness adjustment, were applied to enhance the variability of the training dataset, resulting in 19,600 images. The CNN model was trained using a mini-batch size of 8 over 20 epochs. Evaluation with 100 test images demonstrated a strong estimation performance for volumetric water content, achieving an R² of 0.84, a root mean square error (RMSE) of 1.83%, and a normalized RMSE of 0.07. This highlights the potential of using low-cost cameras and accessible deep learning techniques for soil moisture estimation without relying on labor-intensive conventional soil sensors or costly spectral sensors.