Selecting Hyperspectral Bands for Leaf Mass Per Area Prediction by Means of Neural Networks

Accurate prediction of Leaf Mass per Area (LMA) from leaf-level hyperspectral reflectance measurements is fundamental for plant health watching. Limiting the number of required bands to a very low number while maintaining a good accuracy of LMA prediction is of very high interest for cost-effective implementations. For such purpose, the best current model consists in a combination of vegetation indices and have been obtained by performing computationally intensive brute force approach. Further accuracy improvements are, at present, limited by current computational capacity limit. We propose to use a neural network-based methodology for selecting a very low number of hyperspectral bands for further improving the LMA prediction accuracy. Testing on the publicly available Angers dataset, we diminished the Root Mean Square Error of prediction by 26% and 19% as compared to the best literature vegetation index and combination of vegetation indices, respectively.