Estimation of citrus yield from airborne hyperspectral images using a neural network model

This research was conducted as a preliminary step for developing a methodology for estimating tree crop yield from airborne hyperspectral images. Using an Airborne Imaging Spectrometer for Applications (AISA) Eagle system, hyperspectral images in the 72 visible and near-infrared (NIR) wavelengths (from 407 to 898 nm) were acquired over a citrus orchard in Japan during the months of April, May and June of 2003. Average spectral reflectances for the canopies of 31 selected tree samples were extracted using ERDAS IMAGINE 8.6 software. Fruit yield data on individual citrus trees were collected during the local harvest season in 2003. A backpropagation neural network algorithm was applied to relate the average canopy reflectance to citrus yield for individual trees. Ten thousand experiments of neural network training were carried out for each of the three hyperspectral data. The best fit model was then identified from these 10,000 models for each hyperspectral data. The best fit model as well as the 10,000 ensemble models analyses indicated that the models corresponding to the hyperspectral data collected in May predicted citrus yield more accurately than those collected in April and June. These results demonstrate that the neural network model could work well for the hyperspectral data observed in a specific season, and suggest a potential of using airborne hyperspectral remote sensing to predict citrus yield.