Tree-level irrigation inference using UAV thermal imagery and convolutional neural networks

Abstract

Irrigation management has been one of the keys to achieve sustainability and marketability in agriculture, which is estimated to account for over 70% of global water use. The optimal use of water through irrigation is important for the evolution of agriculture. Many progressive growers make irrigation decisions using crop evapotranspiration (ETc). With the advent of Unmanned Aerial Vehicles (UAVs), lightweight sensors, such as thermal camera, can be mounted on the UAVs to take high-resolution images. Compared with satellite imagery, the spatial resolution of the UAV images can be at the centimeter level. Thus, in this article, the authors proposed a reliable individual tree-level irrigation inference system using a small UAV platform and Convolutional Neural Networks (CNNs). A field study was conducted at the USDA-ARS Research Center in Parlier, California to train and test CNN models using images of the pomegranate trees. The pomegranate field was randomly designed into 16 equal blocks to test two irrigation levels, the low irrigation volume (35% and 50% of ETc) and high irrigation volume (75% and 100% of ETc), measured by a weighing lysimeter in the field. Results showed that the trained CNN model could successfully classify the individual tree using the thermal UAV imagery into the targeted irrigation levels. The overall prediction accuracy was around 87%, which showed a state-of-art performance and indicated that UAV thermal imagery could infer the irrigation levels at individual tree level.