Deep learning for identification of water deficits in sugarcane based on thermal images

Thermal images of plants have been used as a way for monitoring water status since it does provide a non-destructive method that allows its remote evaluation. Even with the facilities provided, it does require a high level of expertise for a human to evaluate these images and this is a time-consuming task. In this context, machine learning techniques are crucial tools that can automatically evaluate the images and predict the water status of plants with more accuracy and as fast as possible to deal with the dynamic in time and space of agriculture fields. In this work, a method for the evaluation of thermal images of sugarcane crop predicting its water status is proposed. The method can be used to evaluate thermal images taken with any type of portable thermal camera available on the market today. For this an artificial neural network model is used, Inception-Resnet-v2 network, that is a kind of a deep learning model combined with the transfer learning technique that allow archive high accuracy with low time and cost, compared with traditional methods. The proposed model performance was compared to a human evaluating the same set of thermal images. Overall results showed that the developed system achieved superior performance compared to the assessments made by the human and aided in classifying the water stress of plant’s thermal images in a non-destructive manner. In addition to superior performance accuracy by 23 %, 17 %, and 14 % for the AWC classes of 25 %, 50 %, and 100 %, respectively, the deep learning model demonstrated a greater ability to distinguish between the classes of thermal stress. Thus, the system developed in this study is a less time-consuming, affordable, non-destructive and effective tool for estimating the water status of sugarcane plants in the field, based on thermal images.