This project addressed the problem of detecting and classifying shallow water underwater munitions using datasets collected from synthetic aperture sonar (SAS) systems. This project discusses an adaptive scheme for unexploded ordnance (UXO) detection and how to discriminate between UXO from non-UXO objects using manifold learning principles when applied to datasets collected from SAS systems.

Technical Approach

The researchers’ detection hypothesis was that the presence of munitions in the sonar backscatter collected from a hydrophone array will inherently lead to a low-rank component in the pulse compressed data from multiple pings. A statistical hypothesis test was developed to determine when this low-rank component is present using the Generalized Likelihood Ratio Test (GLRT).

The researchers’ classification hypothesis was that the sequence of measurements collected from an object in a linear SAS survey results in data that lies in some low-dimensional subspace which is locally linear but globally non-linear, i.e. the data is assumed to lie on a low-dimensional manifold embedded in a high-dimensional space. The coordinates on that low-dimensional manifold and their behavior could then be used to discriminate among various UXO and non-UXO objects that may be encountered in practice. Techniques were developed to not only learn the low-dimensional manifold, but to also provide an out-of-sample embedding for newly acquired training data. The manifold features from the training set were then used to construct local linear subspaces for representing each newly embedded testing feature. A statistically motivated technique was then used to select the most likely class label by finding the class which best represents the data.

Test results for both the detector and classifier were then presented using an experimental data set which was designed to collect sonar data from underwater objects in a relatively controlled and clutter-free environment. Various experiments were also conducted to observe the proposed system's robustness to various forms of mismatch that may enter the data collection process. Results are presented using standard performance metrics such as probability of detection (Pd), probability of correct classification (Pcc), probability of false alarm (Pfa), as well as Receiver Operating Characteristic (ROC) curve and confusion matrix characteristics.


The results of this project showed that the detection algorithim, applied within these fairly ideal data sets was capable of discriminating sonar returns of objects lying on the seafloor from the background with a probability of detection of Pd = 98% and an average of 2.4 false alarms per image (with each image covering approximately 20 m2). Images of the likelihood ratio produced by the detector also demonstrated its ability to localize each object on the seafloor. Classification results generated using the same experimental data set demonstrated the ability of the proposed classification technique to accurately discriminate the sonar returns of UXO objects from those of non-UXO objects with a probability of correct classification of Pcc = 93%. Moreover, the proposed method was able to correctly classify nearly 70% of the testing data for the ‘real’ UXO which was not included during the training process.


This problem is technically challenging due to the variability in environmental conditions, as well as obscuration of the munitions. Thus, new methods were needed to rapidly and reliably assess large areas that are potentially contaminated with munitions and detect, localize, and identify each individual threat with a high degree of accuracy. To this end, this research addressed an important shortcoming of the existing Automatic Target Recognition (ATR) algorithms that use sonar data by developing new environmentally adaptive algorithms for the detection and classification of military munitions in shallow underwater environments using data collected from low frequency broadband SAS systems. This research directly benefits the Department of Defense by reducing the number of false detections of UXO and by allowing more efficient detection of munitions in complex underwater environments.