Predicting Corn Response to Variable Synthetic Fertilizer Treatments Using UAV-Derived Imagery

Abstract

Efficient nutrient management is essential for optimizing corn (Zea mays L.) productivity while minimizing environmental and economic costs. Traditional methods for assessing crop responses to nutrients are often damaging and labor-intensive, limiting accurate assessment of spatial and temporal variations. Accurate in-season yield potential estimation plays a vital role in guiding nutrient management decisions and supporting grain marketing strategies. This study evaluated the potential of Unmanned Aerial Vehicle (UAV) derived imagery to estimate chlorophyll (Chl) status and predict corn grain yield potential in-season under variable nitrogen (N), phosphorus (P), and potassium (K) fertilizer treatments across different growth stages. Two field trials (NP and K) were conducted at two locations in Virginia, Kentland Farm in Blacksburg (Kentland), Valley and Ridge province, and the Northern Piedmont Center in Orange (Orange), Piedmont province. These sites vary in altitude, soil type, and climatic conditions, providing contrasting environments for evaluating crop responses to fertilizers. In both trials, factorial arrangement of treatments (varied N, P, and K fertilizer rates) with four replications was implemented with a randomized complete block design (RCBD). Chlorophyll readings (ChlR) were collected using the Soil Plant Analysis Development (SPAD)-502 and atLEAF Chl meters at three growth stages: early vegetative (EV), late vegetative (LV), and reproductive (Repr). These measurements were synchronized with UAV flights performed on the same day. UAV flights were conducted using DJI Mavic equipped with an RGB sensor for visible light and four monochrome sensors for multispectral imaging (red: 650 nm ± 16 nm, green: 560 nm ± 16 nm, near-infrared (NIR): 840 nm ± 26 nm, red-edge: 730 nm ± 16 nm). UAV-derived vegetation indices (VIs) responsive to Chl and indicative of crop yield potential were computed to model ChlR and yield through single and multi-index regression analyses. Multi-index model performance was evaluated through repeated k-fold cross-validation (CV) (k = 5; 30 repetitions). Indices included the Normalized Difference Vegetation Index (NDVI), Chlorophyll Index Red-Edge (CIRE), Normalized Difference Red-Edge Index (NDRE), Green Normalized Difference Vegetation Index (GNDVI), MERIS Terrestrial Chlorophyll Index (MTCI), Normalized Difference Chlorophyll Index (NDCI), Canopy Chlorophyll Content Index (CCCI), and Optimized Soil-Adjusted Vegetation Index (OSAVI). Weather variation, early-season drought, and late-season rainfall strongly influenced yield formation and grain moisture, overshadowing fertilizer treatment effects. No significant yield differences were detected among N, P, or K levels (p > 0.05). UAV-derived VIs demonstrated significant correlations with both ChlR and yield, with stronger relationships observed during the LV stage when canopy closure and Chl concentration were most stable. In the K trial at Kentland, the relationships between VIs and both ChlR and yield were generally moderate, while in the K trial at Orange, correlations were consistently strong and significant. For ChlR prediction, green and red-edge based indices (GNDVI, NDRE, CIRE, and MTCI) were the most reliable indices, explaining 40 to 55 percent of the variation across both sites and trials. For yield prediction, GNDVI, NDVI, and NDCI consistently exhibited strong relationships at Orange during the LV stage, with R² values ranging from 0.50 to 0.72 across both trials. In contrast, Kentland showed comparatively lower predictive performance, with only moderate relationships observed in the K trial during the LV stage. The use of a polynomial regression (quadratic) model further improved prediction accuracy compared to the linear model in all trials. Multi-index regression further improved predictive accuracy. The best-performing yield models were observed in the K trial at Orange during the LV stage, achieving CV R² values up to 0.71 (CV RMSE of 13.9), while the best ChlR models were found in the NP trial at Orange with CV R² of 0.46 (CV RMSE of 2.45). Model performance was lower for EV and Repr stages. Overall, these findings demonstrate that UAV-based multispectral imaging is an effective tool for monitoring corn canopy Chl status and assessing yield potential, with prediction accuracy varying across growth stages.