A geospatial framework to model within-field phosphorus efficiency via proximal sensing and machine learning

Phosphorus (P) is essential for crop production, yet inefficient management contributes to nutrient losses, water pollution, and eutrophication. Phosphorus use efficiency (PUE) is a key metric for balancing agronomic productivity with environmental sustainability. However, within-field spatial variability of PUE remains poorly understood. This study presents a spatially explicit framework integrating proximal sensing, field measurements, and machine learning to assess and map PUE in corn (Zea mays L.) systems from Central Texas, USA. Grain yield was measured with an Ag Leader yield monitor, while grain protein, oil, and starch were assessed using a CropScan grain quality sensor mounted to the combine. Apparent soil electrical conductivity (EC_a) was mapped using a Veris platform to characterize soil spatial variability. Grain and soil P contents were determined from strategically selected locations using conditioned Latin hypercube sampling and scaled across fields through regression with CropScan measurements. PUE was calculated as the ratio of grain P removal to residual soil P plus applied fertilizer P. A Random Forest (RF) model was trained using EC_a and terrain attributes to predict spatial patterns of PUE. The proximal sensing approach effectively captured P dynamics, with CropScan-based grain P predictions achieving R² up to 0.97. The RF model predicted PUE with high accuracy (R² = 0.78; RMSE = 0.01). EC_a, elevation, and wetness index were the dominant drivers of PUE variability, with predicted values ranging from 0.02 to 0.25. Fields with higher residual soil P exhibited lower PUE, while P-limited fields showed greater efficiency. This framework enables high-resolution assessment of within-field PUE and supports precision P management to enhance productivity while reducing environmental impacts.