Prediction of citrus yield from airborne hyperspectral imagery

Recent advances in spectral imaging technology have enabled the development of models that estimate various crop parameters from spectral imagery data. We developed partial least square (PLS) models to predict fruit yield of Satsuma mandarin using airborne hyperspectral imagery obtained several months before harvesting. Hyperspectral images in the 72 visible and near-infrared (NIR) wavelengths (from 407 to 898 nm) were acquired over a citrus orchard during the early growing seasons of 2003, 2004 and 2005. The canopy features of individual trees were identified using pixel-based average spectral reflectance values for all 72 wavelengths from the acquired images. The acquired canopy features were then used as prediction variables to develop yield prediction models. These were developed using three techniques: (1) normalized difference vegetation index (NDVI), simple ratio (SR) and photochemical reflectance index (PRI), (2) conventional multiple linear regression (NILR) models, and (3) PLS regression models. As we intended to predict yield several months before the harvesting season (generally late December), the conventional techniques (vegetation indices and NILR) did not predict well. In contrast, PLS models gave successful predictions for the three years. These results confirmed the hypothesized correlation between canopy features and citrus yield. The successful forecasting of yields several months or even one year ahead of the harvest season is expected to contribute to planning harvest schedules, generating prescription maps for dealing with fluctuations of yield in specific trees, control measures, and management practices.