Wide-Area Debris Field and Seabed Characterization of a Deep Ocean Dump Site Surveyed by Autonomous Underwater Vehicles

Disposal of industrial and hazardous waste in the ocean was a pervasive global practice in the 20th century. Uncertainty in the quantity, location, and contents of dumped materials underscores ongoing risks to marine ecosystems and human health. This study presents an analysis of a wide-area side-scan sonar survey conducted with autonomous underwater vehicles (AUVs) at a dump site in the San Pedro Basin, California. Previous camera surveys located 60 barrels and other debris. Sediment analysis in the region showed varying concentrations of the insecticidal chemical dichlorodiphenyltrichloroethane (DDT), of which an estimated 350–700 t were discarded in the San Pedro Basin between 1947 and 1961. A lack of primary historical documents specifying DDT acid waste disposal methods has contributed to the ambiguity surrounding whether dumping occurred via bulk discharge or containerized units. Barrels and debris observed during previous surveys were used for ground truth classification algorithms based on size and acoustic intensity characteristics. Image and signal processing techniques identified over 74,000 debris targets within the survey region. Statistical, spectral, and machine learning methods characterize seabed variability and classify bottom-type. These analytical techniques combined with AUV capabilities provide a framework for efficient mapping and characterization of uncharted deep-water disposal sites.


Supplementary Information
1 Sidescan Sonar Pre-processing AUVs with integrated sidescan sonars have port and starboard transducer arrays which transmit narrow acoustic beams that illuminate the seafloor.Range-gated processing of the backscattered signal, combined with subsequent pings from the moving platform, are used to create images of the seabed.To account for transmission loss due to spreading and absorption, time-varying gain (TVG) is generally applied to individual ping data.Additional gain adjustments can be applied to improve image contrast; the challenge in wide area surveys is to develop target detection metrics that account for the very large dynamic range that occurs from local variations in bottom type (bright or weak signal return) but also allows for debris target detection across the survey domain.
During the April 2021 survey, over 160GB of raw sidescan data were collected onboard the AUVs.Several pre-processing steps are performed on raw files prior to developing automatic target recognition (ATR) algorithms.Vehicle heading rate is used to identify and remove pings that occur during vehicle turns which lead to distorted imagery.Portions of the survey where the vehicle exceeded a turn rate of 2.5 • /sec are removed.Next, slant range, pitch, and roll corrections are performed to eliminate image artifacts that arise from vehicle motion.
The Bluefin and Remus surveyed at 15 m and 20 m altitude off the seafloor, respectively.
To equalize pixel intensity as a function of range, pixel intensities are normalized by the average of 60 bands in the range direction and 200 bands in the along-track direction with 50% overlap in both directions.This serves to remove any pixel intensity range dependence while also adapting to local features within the image (Figure 2).The sequences of pings form images that are georeferenced to a longitude and latitude grid using the along track position information, vehicle speed, and cross-track range values.
Finally, the survey was run with 200% overlap such that the same area was imaged twice, at different cross track ranges from the transducer array.Cross-track variability in backscatter intensity is due to range-and angular-dependent factors including seafloor bottom-type and the sonar beam pattern. 1 In this study, objects of interest are small (a few pixels or < 1.5 m) and the majority of bottom-types are sediment-covered and flat such that techniques to register and fuse 2 overlapping segments can lead to double counting of targets.For this reason, we choose to use sections of each scan at ranges 25-75% of the total swath width.For the Bluefin with 150 m swath width, this retains portions of the image at 37.5-112.5 m range and for the REMUS with 200 m swath width, ranges of 50-150 m from the transducers.This choice excludes side lobe interference at close ranges and eliminates geometric distortion at the furthest ranges of the image.The resulting images are stitched together and have no overlap.

Statistics
Statistical metrics provide information on large-scale patterns in the wide-area sidescan survey.The following information describes the definitions of variance, skewness, and entropy and how each relates to sidescan data.The statistics are computed using the pixel intensities, x, contained within each of the defined boxes.
Variance (V ) is defined as: where ν is the mean of A, and N represents the number of values in the distribution.High variance may be associated with the presence of targets or geological features.Skewness (S) is defined as: where ν is the mean of x, σ is the standard deviation of x and E(t) represents the expected value of the quantity t.Skewness is the statistical measurement of the asymmetry of a probability distribution.Areas of high skewness are associated with targets or geology which produce shadows shifting the intensity distribution towards lower values.Entropy (H) is defined as: where p contains the normalized histogram counts of the data, x.H provides a statistical metric for texture or randomness within a box.
While the range-dependent mean of each file has been equalized during pre-processing, the range-dependence of other statistical metrics may still be present.To eliminate rangedependence, the median of the data in each 12.5 m x 12.5 m box is corrected relative to the value nearest to the nadir.This process has a minor impact on the statistical values, but reduces range-dependence in the full survey map visualization.

S5
Prior to anomaly detection, a convolution filter is applied to the range-dependent intensities to reduce noise in the image.The ratio of the along-track to range pixel dimensions is approximately 27:1 for the Bluefin and 30:1 for the REMUS, so 27-point and 30-point Blackman windows are chosen and applied only in the range-direction, respectively.The Blackman window, a generalized cosine window, is defined as: where N is the window length and n is defined between [0, (N +1) 2 Each sidescan file is binarized using an intensity threshold of 4 and targets are counted as independent, connected components in the binary image (Figure 2).The threshold value is chosen to represent 3-5 standard deviations from the mean for most of the survey.Size and intensity characteristics of each target are stored including the number of cross-track pixels, number of along-track pixels, total number of pixels, maximum intensity from the pre-processed file (pre-filtering), and longitude and latitude position of the targets.Along and across track pixels are converted to size using the pixel resolutions of the sidescan sonar listed in Table 1.

Target Distributions and Maps
Detection of small targets in a wide area survey is more difficult in the presence of large objects, such as wrecks, and geological features.The sidewalls of the San Pedro Basin, the southwest and northeast portions of the wide area survey, have the most complex seafloor with small-scale variability in acoustic return.Areas with large rocks or a shipwreck often have dense, discontinuous, acoustic returns which present as small bright targets, similar to the targets of interest in this study.To account for seabed complexity, where it will require higher resolution acoustic and optical imagery to resolutely distinguish between targets and the background, we design a mask to remove targets in regions of known uncertainty.We use an entropy and class-based metric, computed in 12.5 m x 12.5 m seabed boxes, corresponding to the maps in the previous section.We select HRT and textured mud seabed classes and further subset using an entropy criteria to isolate the tails, values less than 5.15 and greater than 5.6 are removed from the target detections.

64 Figure 1 :
Figure 1: Statistical metrics of variance (blue), skewness (red), and entropy (green) are grouped into normalized probability distributions by spectral class.Each class distribution is shown relative to the full survey distribution (grey) for reference.

Figure 2 :
Figure 2: (A) Raw intensity as a function of range for one sidescan ping to port and starboard.(B) Pre-processed intensity includes normalization as a function of range, choice of the main lobe section (white), convolution filtering (green), and an intensity threshold (red) to identify targets.