Unraveling low abundance intimate mixtures with deep learning

The high-confidence detection and identification of very low abundance, subpixel quantities of solid materials in nonlinear/intimate mixtures are still significant challenges for hyperspectral imagery (HSI) data analysis. We compare the ability of a traditional, shallow neural network (NN), deep learning with a convolutional neural network (DL/CNN), and a support vector machine (SVM) to analyze spectral signatures of nonlinear mixtures. Traditional mainstay algorithms (e.g., spectral unmixing, the matched filter) are also applied. Using a benchtop shortwave infrared (SWIR) hyperspectral imager, we acquired several microscenes of intimate mixtures of sand and neodymium oxide (Nd₂O₃). A microscene is a hyperspectral image measured in a laboratory. Several hundred thousand labeled spectra are easily and rapidly generated in one HSI cube of a microscene. Individual Petri dishes of 0, 0.5, 1, 2, 3, 4, and 5 weight-percent (wt. %) Nd₂O₃ with a silicate sand comprise a suite of microscenes furnishing labeled spectra for analysis. The NN and the DL/CNN both have average validation accuracies of ≥ 98 % (for the low wt. % classes); the SVM yields similar performance. As wt. % Nd₂O₃ increases, accuracies decrease slightly—perhaps due to the dominance of the Nd₂O₃ signature in the mixtures, which causes an increasing difficulty in separation. For example, this could affect the 4 and 5 wt. % classes in which the Nd₂O₃ would be easily detected and identified with traditional, mainstay HSI algorithms. The fact that neural network methods can separate such low quantity classes (e.g., 0, 0.5, and 1 wt. %), though not unexpected, is encouraging and demonstrates the potential of NNs and DL/CNNs for such detailed HSI analysis.