Biological fouling of ship hulls is an important issue to the U.S. Navy because it impacts ship performance, fuel efficiency and may provide a pathway for the transport of non-indigenous species (NIS). To date, most monitoring and control programs have focused on ballast water as the dominant mechanism for transport of invasive species. However, recent studies suggest hull fouling may represent a significant pathway for transport of invasive species. The Department of Defense (DoD) needs to quantify environmental risks associated with hull fouling, specifically the risks associated with DoD vessels which may differ from civilian vessels in terms of operational scenarios, hull coatings, and maintenance procedures. Most information on hull fouling is based on diver inspections, which are costly, sometimes dangerous, and often subjective. Improved methods are needed to effectively conduct underwater hull inspections and objectively quantify the abundance and diversity of fouling organisms on ship hulls.

The objective of this project was to develop automated image processing and image understanding algorithms coupled with an artificial neural network (ANN)-based classification scheme for estimating area coverage and diversity of fouling organisms on ship hulls.

ANN used to automatically classify organisms in digital images based on characteristic features.

The automated quantification and classification system for hull fouling organisms utilizes the following four processing modules: (1) image acquisition, (2) image pre-processing, (3) feature extraction, and (4) classification. Estimates of fouling organism abundance are derived from digital images through image pre-processing and feature extraction techniques. The extracted features are then used to identify and classify fouling organisms via digital image processing and ANN classifiers that can self-learn the rules (i.e., features) hidden in the images to quantify and classify the fouling condition. Because of the self-learning capabilities, the unsupervised neural network does not require human intervention except to specify the number of classes (or degrees) of fouling desired for classification.

These methods can be adapted to suit a remotely operated vehicle, developed at the Naval Surface Warfare Center (Carderock Division), for underwater hull evaluation. This technology may facilitate more effective assessment of the degree and type of hull fouling, while providing an improved capability for management and control strategies.

Sample hull fouling data was generated by suspending metal plates in San Diego Bay and imaging the panels with a digital camera. A total of 112 images, covering different fouling conditions, were collected and sub-sampled randomly 2-4 times, producing a total of 360 sub-image samples. The 360 sub-image samples were divided into two datasets, one with 300 samples for training and the other with 60 samples for testing a feature-based ANN and a Self-Organizing Map (SOM). The training and testing datasets were subjectively classified into the following four fouling conditions: minimal, mild, moderate and heavy. For each image, a total of 32 feature variables were extracted using the Spatial Gray Level Difference Method (SGLDM) and entered into the ANN for supervised training of the image classification into the four pre-defined fouling conditions. Results show that the ANN is able to classify images from the testing dataset with approximately 70 percent accuracy. When the SOM was used for initial pre-classification of raw feature data and the ANN was used for final classification, the success rate for classification of the fouling conditions was improved to approximately 99 percent. This project was completed in FY 2002.