Fig 1.
Experimental setup for wild-simulated ginseng cultivation.
The left image shows the pot specifications used for cultivation, measuring 25 cm in height and 9 cm in diameter, which provides an optimal environment for root development. The right image illustrates the cultivation system of using a Genesis FarmBot (bed size: 1.5 m × 3 m, height: 0.5 m), equipped with an X-Y-Z gantry.
Fig 2.
Sample soil surface images at different moisture levels.
These images are provided solely for qualitative illustration, and no spatial or physical scale is implied (not to scale).
Table 1.
Hardware and Software Development Environment.
Fig 3.
Soil moisture sensing system and sensor placement.
Table 2.
Camera Setting for in situ image acquisition.
Fig 4.
Camera Setting for in situ image acquisition.
Fig 5.
Input and processing of soil images.
The original soil image on the left is converted into a three-channel RGB matrix. This process uses IMRED to decompose pixel values into red, green, and blue channels. The resulting matrix is input data for feature learning in deep learning models.
Table 3.
(Top) Comparison of pre-trained CNN architectures, including input/output shapes, number of features, and total parameters. All models take 224 × 224 × 3 input images and generate architecture-specific feature maps. (Bottom) Architectural overview of the soil moisture prediction model, consisting of a pre-trained CNN base followed by modified top layers.
Table 4.
Performance comparison of deep learning architectures for soil moisture prediction. The upper section summarizes R², RMSE, learning rate, and epoch for six models. DenseNet121 showed the highest prediction accuracy, while NASNetMobile performed the poorest.
Fig 6.
Comparison of model complexity, MAE, and generalization gap.
Bubble size denotes generalization gap and labels indicate R². DenseNet121 achieved the best trade-off between accuracy and efficiency, while NASNetMobile showed poor generalization despite low complexity.
Fig 7.
Comparison of actual and predicted values across the six DNN regression models: (a) EfficientNetB0, (b) MobileNetV2, (c) NASNetMobile, (d) ResNet50, (e) DenseNet121, (f) InceptionV3. Each plot demonstrates the alignment between actual and predicted values, with the red dashed line representing perfect prediction.
Models with tighter clustering around the perfect prediction line, such as DenseNet121 and InceptionV3, exhibit higher prediction accuracy, whereas models like NASNetMobile show greater deviations.
Table 5.
Descriptive Statistics of Soil Moisture by Depth.
Table 6.
Performance comparison of six machine learning regression models for soil moisture prediction. Reported R² and RMSE correspond to the best cross-validated hyperparameters. Random forest and bagging decision tree regression achieved the best predictive performance, followed by gradient boosting. Polynomial regression and KNN regression showed moderate performance, while SVR with an RBF kernel yielded the lowest. All models were evaluated on the same dataset containing multi-depth soil moisture measurements.
Fig 8.
Performance evaluation and comparison of six machine learning models for soil moisture prediction.
Left panels display R² scores across different hyperparameter values. Red shaded areas represent the standard deviation from k-fold cross-validation on the test set, and blue lines indicate training performance. Right panel is a scatter plot of actual vs. predicted soil moisture on the testing data using the optimized hyperparameter. The black dashed line indicates the 1:1 reference, and tighter clustering of points along this line reflects better predictive performance.