Mapping Tree Water Deficit with UAV Thermal Imaging and Meteorological Data

The mapping of forest stands and individual trees affected by drought stress is a crucial step in targeted forest management, aimed at fostering resilient and diverse forests. Unoccupied aerial vehicle (UAV)-based thermal sensing is a promising method for obtaining high-resolution thermal data. However, the reliability of typical low-cost sensors adapted for UAVs is compromised due to various factors, such as internal sensor dynamics and environmental variables, including solar radiation intensity, relative humidity, object emissivity and wind. Additionally, accurately assessing drought stress in trees is a complex task that usually requires laborious and cost-intensive methods, particularly in field settings. In this study, we investigated the feasibility of using the thermal band of the Micasense Altum multispectral sensor, while also assessing the potential for modelling tree water deficit (TWD) through point dendrometers and UAV-derived canopy temperature. Our indoor tests indicated that using a limited number of pixels (< 3) could result in temperature errors exceeding 1 K. However, enlarging the spot-size substantially reduced the mean difference to 0.02 K, validated against leaf temperature sensors. Interestingly, drought-treated (unwatered) leaves exhibited a higher root mean squared error (RMSE) (RMSE = 0.66 K and 0.73 K) than watered leaves (RMSE = 0.55 K and 0.53 K), likely due to lower emissivity of the dry leaves. Comparing field acquisition methods, the mean standard deviation (SD) for tree crown temperature obtained from typical gridded flights was 0.25 K with a maximum SD of 0.59 K ( n = 12). In contrast, a close-range hovering method produced a mean SD of 0.09 K and a maximum SD of 0.1 K ( n = 8). Modelling the TWD from meteorological and point dendrometer data for the 2021 growth season ( n = 2928) yielded an R 2 = 0.667 using a generalised additive model (GAM) with vapor pressure deficit (VPD), wind speed, and solar radiation as input features. A point dendrometer lag of one hour was also implemented. When predicting individual tree TWD with UAV-derived tree canopy temperature, relative humidity, and air temperature, an RMSE of 4.92 (μm) and R 2 of 0.87 were achieved using a GAM. Implementing leaf-to-air pressure deficit (LVPD) as an input feature resulted in an RMSE of 6.87 (μm) and an R 2 of 0.71. This novel single-shot approach demonstrates a promising method to acquire thermal data for the purpose of mapping TWD of beech trees on an individual basis. Further testing and development are imperative, and additional data from drought periods, point dendrometers, and high-resolution meteorological sources are required.