An Edge Transfer Learning Approach for Calibrating Soil Electrical Conductivity Sensors.

Area of Science:

• Agricultural Technology
• Machine Learning
• Sensor Networks

Background:

• Electrical conductivity (EC) sensors are crucial for smart agriculture but suffer from drift, necessitating recalibration.
• Existing EC sensor calibration methods often rely on standard sensors and cloud-based processing, which can be costly and inefficient.

Purpose of the Study:

• To develop an efficient, edge-based machine learning approach for recalibrating EC sensors in smart agriculture.
• To enable on-site AI training and transfer learning for EC sensor calibration, reducing reliance on cloud infrastructure.

Main Methods:

• Proposed SensorTalk3, an ensemble of XGBOOST and Random Forest models executable on edge devices like Raspberry Pi.
• Integrated soil temperature and moisture sensor data as key features for calibration.
• Developed a dual-sensor detection solution for determining recalibration needs.

Main Results:

• SensorTalk3 achieved a Mean Absolute Percentage Error (MAPE) as low as 1.738%, a significant improvement over the original sensor's 7.792% error.
• Accurate EC calibration was achieved even with uncalibrated moisture and temperature sensors (errors ≤ 8.3%).
• On-site AI training and transfer learning were successfully demonstrated at the edge node.

Conclusions:

• SensorTalk3 offers a cost-effective and accurate solution for EC sensor recalibration in smart agriculture.
• Edge-based AI training and transfer learning represent a significant advancement for on-site data processing.
• The proposed method enhances farming intelligence through improved sensor data reliability.