From Single to Multi-Sensor UAV Strategies: Growth-Stage-Specific AI Modeling Improves Soil Moisture Estimation in Maize Field

Accurate soil-moisture estimation is crucial for optimizing irrigation and improving water-use efficiency. Although most drone-based sensors (e.g., RGB–thermal, hyperspectral, and ground-penetrating radar (GPR)) offer strong capabilities in estimating soil moisture, their performance varies across maize growth stages due to physiological and structural changes. Traditional single-level modeling, which treats the season uniformly, often misses these stage-specific dynamics and reduces predictive accuracy. To address this, we propose a multi-level framework that segments the model by growth stage (early, mid, late) and tailors feature selection accordingly. Our results show that multi-level models consistently outperform single-level models in accuracy, generalizability, and bias reduction. Across the full season, single-source single-level models achieved R² values of 0.61 (RGB–thermal), 0.65 (hyperspectral), and 0.74 (GPR), whereas stage-aware models raised early-season R² to 0.90–0.92 and reduced RMSE by ∼50% (e.g., RGB–thermal 1.48%, hyperspectral 1.60%, GPR 1.36%). Data fusion further improved early-season performance to R² = 0.93 with RMSE = 1.30%. External validations confirmed robustness: for RGB–thermal, multi-level vs. single-level RMSE was 1.40% (early), 3.62% (mid), and 2.57% (late); for hyperspectral, 3.14% (mid) and 4.38% (late). Multi-level modeling thus halves the error in early-stage estimates. RGB–thermal within the multi-level framework delivered high accuracy at the lowest cost, making it practical for early-season deployment. GPR was most reliable in mid to late stages because it penetrates dense canopies and senses subsurface moisture. Hyperspectral data performed moderately well for low-moisture detection, reflecting sensitivity to pigment and structural change, but degraded under high-moisture conditions due to spectral saturation and physiological decoupling. While data fusion produced the strongest accuracy in both single- and multi-level settings, it also incurred higher costs. A cost-efficiency assessment showed that multi-level modeling allows lower-cost sensors to achieve high accuracy by leveraging their optimal performance windows across the season—an important consideration for growers balancing performance and affordability. We recommend RGB–thermal with a multi-level model for early-season monitoring and GPR for mid-season soil-moisture assessment. Although fusion models offer the highest performance, their cost may limit widespread adoption. Overall, growth-stage–specific modeling provides a scalable, accurate, and economically feasible pathway for soil-moisture monitoring in maize fields.