Prediction of morpho-physiological traits in sugarcane using aerial imagery and machine learning

Abstract

Changes in morpho-physiological traits are important for biotic or abiotic stresses in sugarcane (Saccharum spp. interspecific hybrids). Ground measurements of such traits are labor-intensive and time-consuming. Therefore, predicting them using aerial imagery can be important for detecting stress and timely management. In this study, ground data and aerial imagery were collected from plant cane and first ratoon (two site-years) field trials of the last stage (Stage IV) of the Canal Point sugarcane cultivar development program in Florida. There were multiple genotypes (12 plant cane: 9 first ratoon) with six replications in each trial. Data were collected on soil plant analysis development (SPAD), leaf area index (LAI), plant height, normalized difference vegetation index (NDVI), and number of millable stalks per hectare. Aerial imageries were collected using a hyperspectral sensor, and ground data using handheld sensors and manual readings on multiple dates (April, July, and September) to determine the best timing in morpho-physiological trait prediction. The gradient boosting regression tree was selected as the best prediction model. The mean absolute percentage error (MAPE) was utilized to determine the model's prediction accuracy. Results showed that SPAD was predicted with higher accuracy (89%) compared to other traits. July was observed as the best time for predicting most of the morpho-physiological traits in plant cane and first ratoon. Furthermore, the NDVI values collected by GreenSeeker and UAV imaging were compared using the Bland-Altman degree of agreement, and it was found that the mean difference of NDVI values between the two sensing systems was low (0.09).