Image-based machine learning models for customized soil moisture management

Area of Science:

• Agricultural Engineering
• Artificial Intelligence in Agriculture
• Precision Agriculture

Background:

• Conventional smart farming applies uniform treatments, failing to address individual crop needs due to microenvironmental variations.
• Inefficient resource use and reduced yields result from overlooking plant-specific requirements.
• Non-invasive methods for individual plant soil monitoring are lacking, hindering customized crop management.

Purpose of the Study:

• To develop an AI-based system for customized soil moisture management.
• To enable non-invasive monitoring of soil conditions at the individual plant level.
• To improve smart farming efficiency and crop yields through plant-specific data.

Main Methods:

• Utilized RGB images and soil sensor data (3cm, 10cm, 15cm depths) for wild-simulated ginseng.
• Employed deep learning models (DenseNet121) for surface moisture prediction.
• Applied random forest regression for deeper soil layer moisture dynamics.

Main Results:

• DenseNet121 achieved high accuracy (R²=97.3%) for surface moisture prediction.
• Random forest regression demonstrated strong performance (R²=90.6%) for deeper soil layers.
• Validated the non-invasive estimation of soil moisture using surface image data.

Conclusions:

• AI-driven image and sensor analysis enables customized soil moisture management.
• The system supports data-driven, plant-specific approaches in smart farming.
• Future work should involve diverse crop/soil validation and spectral data integration.