In-field calibration of uav thermal imagery for precision agriculture

Abstract

My study investigates in-field calibration of drone based thermal imagery as a precision tool for assessing water stress. High-resolution canopy temperature mapping serves as a critical indicator of plant water use dynamics. Collecting actionable measurements requires accurate sensor calibration to overcome bias in unmanned aerial vehicle (UAV) mounted uncooled long-wave infrared (LWIR) cameras. I obtained reference temperature measurements using two high-emissivity foam panels monitored by NIST-traceable Apogee SI-121 infrared radiometers. Based on these target temperatures I evaluated seven calibration method variants on 106 synchronized radiometer-sensor pairs from a single flight over a barley (Hordeum vulgare) field located near Bozeman, Montana. I compared the factory baseline against calibration models using empirical line method (ELM) with temporal drift correction and generalized versions, both incorporating the camera's focal plane array (FPA) temperature as a covariate. RANSAC (RANdom SAmple Consensus) robust regressions minimized the influence of outliers during model fitting, with Monte Carlo cross-validation (200 iterations) and bootstrap resampling (1,000 iterations) used to quantify uncertainty. The factory calibration produced an RMSE of 4.11°C, with opposite-sign gain errors of approximately +4.3°C on the cold panel and -2.5°C on the hot panel. Post-deployment calibration reduced RMSE by up to 44.8%, with the best-performing method, generalized LWIR+FPA, achieving RMSE=2.27°C. Applying this calibration to the full flight mosaic shifted mean canopy temperature from 39.97°C to 32.42°C and produced agronomically plausible Crop Water Stress Index values consistent with late-season dry-farmed barley under high vapor pressure deficit; whereas the uncalibrated imagery placed the entire canopy distribution implausibly above the non-transpiring upper limit. These results demonstrate that incorporating brief reference target imaging at takeoff and landing substantially improves the accuracy of UAV thermal data products for crop water stress assessment.