Calibration and Machine Learning-Based Prediction of Soil Moisture in High-Density Apple Orchards

Resistive sensors can provide affordable measurements of soil-water dynamics, although site-specific calibration is necessary for reliable interpretation. A low-cost sensor system with wireless communication was evaluated using clay-loam soil from a high-density apple orchard at SKUAST-Kashmir, India. Nine laboratory calibration levels were prepared by adding 0–400 mL of water to 1 kg of oven-dry soil; the resulting nominal mass-based water-content scale was 0–40%. The analog-to-digital converter (ADC) counts decreased approximately linearly with water addition. A linear model fitted to the nine observations yielded R² = 0.99964 and an in-sample root mean square error of 5.04 ADC counts, equivalent to 0.24 percentage points on the nominal gravimetric scale. A fifth-degree polynomial produced R² = 0.99989 and 2.81 ADC counts (0.14 percentage points), but its added complexity increases the risk of overfitting. Sensors were installed at 15, 20 and 25 cm in the orchard, and aggregate field comparisons with gravimetric and time-domain reflectometry measurements were summarised. Random forest, neural network and long short-term memory models were also explored using field-derived time-series observations. The available aggregate summaries do not establish independent predictive generalisation or quantify changes in irrigation water use. These findings support the feasibility of a site-specific sensor system while identifying reference-moisture measurement, chronological validation and longer-term testing as essential requirements for operational irrigation decisions.