Multiparametric Regressions and Machine Learning: UAV Imagery for Temperature Prediction of Irrigated Forage Cactus in the Brazilian Semiarid

Abstract

Vegetation surface temperature is a key indicator of crop water status, particularly in semi-arid regions where water availability constrains agricultural production. This study evaluated whether spectral bands and vegetation indices derived from UAV multispectral imagery can predict the surface temperature of irrigated forage cactus (Opuntia spp.) grown with reused water in the Brazilian semiarid region. A randomized block experiment was conducted in Pesqueira, Pernambuco, Brazil. Canopy temperature was measured with a handheld thermal camera (FLIR E6), and multispectral imagery was acquired with a DJI Phantom 4 UAV across three flights. Linear, polynomial, and multiparametric regression models were compared against four machine learning algorithms (Random Forest (RF), epsilon-support vector regression (SVR), Extreme Gradient Boosting (XGBoost), and K-Nearest Neighbors (KNN)), each evaluated over 1500 randomized train/test splits. Vegetation indices were negatively correlated with canopy temperature, consistent with higher vegetative vigor being associated with cooler canopies. The seven-variable multiparametric regression model achieved the best regression performance (validation R2 = 64.70%, Nash–Sutcliffe efficiency = 0.84), while KNN was the best-performing machine learning algorithm (mean test R2 = 78.11%). The findings suggest that UAV-derived spectral indices represent a promising indirect proxy for estimating forage cactus canopy temperature under the evaluated semiarid conditions. However, independent validation across diverse environments and management conditions is necessary prior to operational implementation as an alternative to direct thermal sensing.