Uranium Production Pathway Classification Using Quantified Morphology Without Human Interaction

Facilities that handle special nuclear materials must account for the entire mass of materials in their control. Should that material be diverted, it must be traced back to its origins. This work explores whether uranium oxide precipitation routes can be classified through quantified morphology without human interaction. The system to achieve this started with a program to generate quantified morphology for algorithmically segmented images. The quantified morphological attributes for each segment were averaged before being used to train and test various machine learning (ML) classifiers on a five-class problem. Seven processing factors were evaluated in a design of experiments (DOE) study to determine their impact on classification accuracy using analysis-of-variance techniques. These processing factors included process steps like denoising, normalization, and augmenting the ML data with the final oxide type (UO₃, U₃O₈, or UO₂) and the image pixel width. The factors that offered the best accuracy included dataset augmentation with the final oxide and pixel width. The most accurate standard ML classifiers were the support vector machine (SVM) and random forest. The most beneficial DOE factors were applied to the most accurate classifiers, which were subsequently retrained resulting in an accuracy as high as 89% with an SVM. These experiments were repeated after dropping two attributes called “grayscale mean” and “gradient mean” from the training data to check for bias in the data, which reduced the best accuracy to 85%. This paper also includes an experiment to automatically select the watershed segmentation threshold using clustering analysis techniques. Those techniques used a dimension reduction algorithm and a clustering algorithm with silhouette scoring to select the best segmentation threshold for each image in the dataset. However, this reduced classifier accuracy by an average of 17%. Regardless, this work demonstrates that uranium oxide precipitation routes can be classified using morphology without human interaction with up to 89% accuracy.