Crop Hyperspectral Dataset Unmixing Using Modified U-Net Model

The hyperspectral unmixing procedure allows the extraction of different material information (endmembers) and their abundances from pixel spectra of the hyperspectral image for further analysis. An extensive study of papers on hyperspectral unmixing methods that use variational autoencoders, deep convolutional networks, or visual transformers shows a lack of diversity in the hyperspectral imaging (HSI) datasets. In this paper, we propose using a modified U-Net architecture of a deep convolutional neural network to analyze a new UAV-gathered HSI dataset for crop health analysis. Secondly, a data analysis methodology was created to extract a set of endmembers with a variation in the samples from the original HSI data for ground truth creation. Our U-Net model was modified to extract endmember class data at the bottleneck layer, and the SoftMax activation function was used during the decoding phase to extract the abundance map. Multiplication of abundance and endmember matrices reconstructs the HSI upon which the reconstruction losses are calculated. We used an image-patching approach to improve the model’s ability to learn endmember data from the HSI dataset. The model performance was evaluated on our newly created data and openly available hyperspectral datasets such as Washington DC Mall, Samson, and Urban.