Inference of Agronomic Information in Smallholder Systems from Low-Altitude UAV RGB Imagery 

Multispectral and hyperspectral drones provide rich agronomic information but remain costly 
for smallholder farming systems, where consumer RGB cameras are far more accessible. Yet, 
the extent to which simple RGB imagery can substitute for, or complement, more advanced 
sensing remains poorly quantified. This study evaluates that potential of cereal monitoring in 
Morocco, across four analytical axes: RGB spectral indices, GLCM texture, cross-modal 
Support Vector Regression (SVR) for continuous NDVI prediction, and Structure-from-Motion 
(SfM) crop-height estimation, complemented by altitude-robustness and sensor-sensitivity 
analyses.  
Multi-temporal data were acquired over wheat and barley at an experimental station (nine dates, 
2024–2025), with a second site used for cross-site transfer testing. The Green Leaf Index 
tracked NDVI at r = 0.929 and detected senescence onset two to three weeks earlier. The SVR 
model predicted NDVI at R² = 0.933 under spatially independent (Spatial K-Fold) validation, 
with spectral and textural features contributing comparably. Object-based classification 
recovered 82% of the multispectral discrimination capacity. Spectral indices remained stable at 
altitudes from 10 to 90 m altitude, whereas texture varied with resolution.  
Critically, direct cross-site transfer failed (R² = −1.401); however, a two-parameter linear 
recalibration restored performance (R² = 0.576). Furthermore, simulated cross-sensor tests 
preserved spatial structure (R² > 0.97) while shifting absolute values by −6% to −17%. These 
convergent results reveal a general principle: RGB imagery preserves the relative geometry of 
agronomic information but not its absolute scale. Consumer RGB drones therefore offer a 
reliable, low-cost temporal monitoring tool that complements, rather than replaces, 
multispectral sensing, provided that per-site and persensor recalibration is applied.