Neural network algorithm for coffee ripeness evaluation using airborne images

A NASA unmanned aerial vehicle (UAV) was deployed over a commercial coffee plantation during the 2002 harvest season. An on-board digital camera system collected a set of high-resolution surface reflectance images in three spectral bands (580, 660, and 790 nm). An intelligent and robust algorithm operated on the multispectral images to estimate absolute percentages of under-ripe (green), ripe (yellow), and over-ripe (brown) coffee cherries displayed on the canopy surface. The procedure was based on a coupled leaf/canopy radiative transfer model (LCM2), modified to include fruiting bodies as photon scattering and absorbing elements. A neural network (NN) set was trained on simulated data, and then used to invert LCM2 for retrieval of fruit and leaf percentages from empirical canopy reflectance data. A projection technique was implemented to systematically mitigate situations where the observed reflectance data fell outside the NN training set domain and the inversion thus initially rendered non-physical solutions (fruit percentages outside of range 0 to 100%). The algorithm was applied to three study fields representing a broad gradient of mature (ripe plus over-ripe) fruit ranging from 28% to 61%. Correlation between predictions and yield data across all ripeness levels was 0.78, with a mean absolute error of 11% (range 1% to 26%). By comparison, a standard ground-based harvest readiness assessment produced a correlation 0.64 with yield, mean absolute error of 13% (range 5% to 23%). The procedure was designed to operate on a reasonably modest set of a priori specifications and, by coupling with remote sensing, potentially represents an efficient method for monitoring ripeness progression or other agricultural phenomena that alter visible and near-infrared crop canopy reflectance.