Predicting Protein Content in Grain Using Hyperspectral Deep Learning
We assess the possibility of performing regression analysis on hyperspectral images utilizing the entire spatio-spectral data cube in convolutional neural networks using protein regression analysis of bulk wheat grain kernels as a test case. By introducing novel modifications of the well-known convolutional neural network, ResNet-18, we are able to significantly increase its performance on hyperspectral images. Our modifications consist of firstly applying a 3D convolution layer enabling learning of spectral derivatives that 2D spatial convolution is unable to learn, and secondly, the application of a (1 x 1) 2D convolution layer that downsamples the spectral dimension. Analysis of the responses learned by the convolution kernels in our modifications reveals meaningful representations of the input data cube that reduce noise and enable the subsequent ResNet-18 to perform more accurate regression analysis.