This project seeks to quantify the optimal strategies as well as limitations for using optical data to locate and identify underwater military munitions (UWMM). The factors are known for affecting use of optical data to search for and classify UWMM: resolution, noise, visibility, biofouling, and burial. The goal here is to quantify how these factors affect data quality, and then, in turn, assess how data quality affects machine-assisted mapping and classification of UWMM targets. There are three objectives of the project, which together respond to the “optical sensors” and “machine learning” topics.

1. To acquire an extensive imagery dataset of UWMM across relevant substrates, gradients of environmental conditions, and with targets both proud and partially buried.
2. To use the dataset from (1) to quantify the effects of these parameters on optical data quality.
3. To use the dataset from (1) and the model quality assessment from (2) to address data fusion and limited training data problems in machine learning for UWMM classification.

This project will assemble a suite of test objects consisting of 40-50 inert munitions and a similar number of other man-made objects (“clutter”). Half of the objects will remain in seawater to induce corrosion and biofouling over the three-year duration of the project, half will remain in their initial condition. This project will also acquire images of these objects underwater at regular intervals over the course of three years, thereby capturing data under a wide range of water optical properties and physical conditions of the targets. Sand, seagrass, rocky, and mixed habitats will be used, increasing the dataset diversity and allowing for partial burial of the objects. Three-dimensional (3D) models generated from these images using photogrammetric techniques will then be fed into classification algorithms to assess modern methods for automated UWMM classification.

The first benefit will be a quantitative assessment of the geometric accuracy of underwater photogrammetric 3D models as a function of environmental and imaging parameters (light, visibility, camera noise etc..). The second benefit will be an understanding of how 3D model accuracy cascades into image classification accuracy. Together, these will answer the question how good does your optical data have to be to map UWMM? Answering this question has clear benefits for future Department of Defense efforts to incorporate optics in systems for UWMM detection and monitoring. A third benefit will be a large, annotated dataset of UWMM images (>150 3D models, >150,000 images) that will be published and publicly available. Algorithm development by the imaging community benefits greatly from such pre-labeled datasets. Making such a dataset available leverages the work done in this project by encouraging potentially a huge pool of algorithm developers to work on this problem.