Estimating stomatal conductance of citrus orchard based on UAV multi-modal information in Southwest China
Quanshan Liu, Zongjun Wu, Ningbo Cui, Shunsheng Zheng, Shouzheng Jiang, Zhihui Wang, Daozhi Gong, Yaosheng Wang, Lu Zhao, Renjuan Wei
Abstract
Stomatal conductance (Gs) reflects the extent of water stress experienced by crops, which plays a crucial role in precision irrigation and water resource management. High spatiotemporal resolution multimodal remote sensing data from unmanned aerial vehicles (UAV) offers great potential for accurately predicting crop stomatal conductance to monitor crop water stress. In this study, multispectral and thermal infrared remote sensing data of citrus canopies were acquired using UAV. Multimodal features, including RGB, spectral, and thermal information of the citrus canopy, were extracted. Simultaneously, Gs of citrus and soil moisture content (SMC) were collected. The Black-winged Kite Algorithm (BKA) was employed to optimize both the Extreme Learning Machine (ELM) and Kernel Extreme Learning Machine (KELM) models. Gs estimation models for citrus were constructed by incorporating RGB, multispectral (MS), and thermal infrared (TIR) data, as well as their combinations, using the BKA-KELM, BKA-ELM, KELM, and ELM algorithms. The results showed that Gs had the highest correlation with the average soil moisture content (SMCa) at a depth of 0–40 cm (R² = 0.674, P < 0.05). Additionally, Gs exhibited a strong correlation with 20 cm and 40 cm soil moisture content (SMC20 and SMC40), with R 2 of 0.638 and 0.606, respectively (P < 0.05). The fusion of RGB, MS, and TIR multimodal information significantly improved the accuracy of Gs estimation. The Gs models constructed using RGB, MS and TIR as inputs demonstrated the best estimation performance, with R² ranging from 0.859 to 0.989, and RMSE from 1.623 mmol to 5.369 mmol H₂O m⁻²·s⁻². The BKA optimization algorithm effectively enhanced the predictive performance of the KELM and ELM models. The BKA-KELM7 model, using RGB+MS+TIR feature information as inputs, was identified as the optimal model for estimating citrus Gs, with R² ranging from 0.906 to 0.989, and RMSE from 1.623 mmol to 3.997 mmol H₂O m⁻²·s⁻². This study showed that combining multimodal information from low-cost UAV with the optimized machine learning algorithm can provide relatively accurate and robust estimates of citrus Gs. It offers an effective method for estimating Gs using only UAV data, providing valuable support for precision irrigation and field management decisions. • Citrus stomatal conductance (Gs) exhibited a stronger correlation with deep soil moisture content. • The fusion of multimodal remote sensing data has significantly improved the Gs estimation accuracy. • The use of the BKA to optimize ELM and KELM has enhanced the prediction accuracy and stability of the Gs model. • Compared to ELM, KELM exhibits significant potential for the accurate remote sensing estimation of Gs.