Model-based neural network algorithm for coffee ripeness prediction using helios UAV aerial images

Over the past few years, NASA has had a great interest in exploring the feasibility of using Unmanned Aerial Vehicles (UAVs), equipped with multi-spectral imaging systems, as long-duration platform for crop monitoring. To address the problem of predicting the ripeness level of the Kauai coffee plantation field using UAV aerial images, we proposed a neural network algorithm based on a nested Leaf-Canopy radiative transport Model (LCM2). A model-based, multi-layer neural network using backpropagation has been designed and trained to learn the functional relationship between the airborne reflectance and the percentage of ripe, over-ripe and under-ripe cherries present in the field. LCM2 was used to generate samples of the desired map. Post-processing analysis and tests on synthetic coffee field data showed that the network has accurately leam the map. A new Domain Projection Technique (DPT) was developed to deal with situations where the measured reflectance fell outside the training set. DPT projected the reflectance into the domain forcing the network to provide a physical solution. Tests were conducted to estimate the error bound. The synergistic combination of neural network algorithms and DPT lays at the core of a more complex algorithm designed to process UAV images. The application of the algorithm to real airborne images shows predictions consistent with post-harvesting data and highlights the potential of the overall methodology.