The Department of Defense (DoD) needs cost-effective methods for locating and identifying navigation and safety hazards related to underwater military munitions (UWMM). This is a broad and diverse problem because historical information relating to the locations of underwater munitions is often limited and not always accurate. Furthermore, some munitions were dispersed over large areas, being dumped by vessels, dropped via aircraft, or shot as projectiles on or adjacent to live target ranges.
Technologies are needed that can efficiently and objectively detect, identify, and map UWMM. Furthermore, knowledge of benthic environments adjacent to UWMM is critical for remediation decisions. Managers need to know, for example, if detected munitions and explosives of concern are encrusted in a reef or are partially buried; are they intact or cracked open; are they mobile or hard-stuck in silty mud? Currently, both wide-area searches and detailed mapping of UWMM rely on multiple acoustic (e.g., side-scan or multi-beam sonar) and/or metal detection methods.
Underwater optical images of the seabed could benefit surveys for UWMM as well as efforts to understand the environments around UWMM because optical images have very high spatial resolution. Objects on the seabed as small as a few centimeters in diameter can be easily resolved with optical imagery. In many cases, the type or condition of munitions, such as whether they are intact or not, can be discerned as well. Currently, however, analysis is a bottleneck for quantitative assessment of underwater images; therefore, imagery tends to be used for UWMM response in only a qualitative way involving visual inspection and interpretation by an analyst.
The project team recently developed a new seabed classification algorithm that has been shown to accurately classify benthic images from coral reefs. Three technical objectives were identified to evaluate the potential of underwater seabed images to improve both wide-area and detailed surveys for UWMM. First, how well can the algorithm classify munitions targets? Second, how well can the algorithm classify seabed types, thereby characterizing the environments around munitions targets? Third, would further development on the algorithm yield improvements in the ability to discriminate munitions and aspects of the environment?
Prior to this project, the supervised classification algorithm the team had developed had been tested with images of coral reefs from three different locations and one dataset of seabed images from a location where UWMM were found. Based on positive results from these preliminary tests, the approach for this project was to use two large datasets to test the algorithm in a wider variety of settings. The first dataset was collected specifically for this project, using inert surrogate munitions placed on the seabed as known targets. The second set of tests exploited an existing dataset of more than 30,000 images collected south of Pearl Harbor, HI, covering an area where actual chemical and conventional munitions were discarded over a 50-year period. Together, these datasets helped provide an assessment of the capabilities of the current classification algorithm across a wide spectrum of depths, substrates, benthic biota, and munitions types.
In addition to tests of the existing supervised classification algorithm, the project team investigated the utility of adding additional features to the classifier that were based on the local relief, or height, of the seabed. Height data was generated from the input images themselves using structure-from-motion computer vision techniques. A third dataset, also from Hawaii, was used to develop the modified algorithm, which was then tested on the newly collected dataset from Miami.
The image classifier by itself was shown to distinguish munitions from non-munitions (background) with generally high (> 80%) accuracy. This was accomplished at multiple sites in shallow water over seagrass, reef, and sand, and at depths greater than 500 m in sand. Discrimination of environments was high for the major seabed types. For example, sand and mixed sand-seagrass were classified with 80-100% accuracy in both shallow and deep water.
The image classifier by itself did not achieve > 80% accuracy for every class in every situation, however, even with a binary munitions/non-munitions scheme. False positive matches for munitions were observed in the Miami dataset over reef and seagrass, and false negatives were observed in the Hawaii dataset due to confusion of munitions with other anthropogenic clutter. These limitations indicated that, indeed, there remains a need for improvements to the algorithm.
Extending the existing algorithm to also use height data derived from stereo reconstruction showed that incorporating such so-called “2.5-D” data greatly improved the classification results. Using the 2.5-D information reduced the number of false positives in the Miami dataset. Furthermore, improved accuracy was observed not only on the basic, binary munitions / non-munitions classes. Adding 2.5-D information improved the capability to discriminate different types of munitions from one another.
2.5-D reconstruction of a portion of the Ordnance Reef dataset displayed as a textured surface (left) and as a triangular mesh (right).
UWMM response would be enhanced by adding quantitative image analysis to the toolbox of survey methodologies. Furthermore, any other application requiring quantitative assessment of benthic communities would also benefit from the same tools. Thus, an accurate, automated algorithm that can classify seabed images from diverse environments would be a benefit both to DoD and the wider scientific community.