TL;DR
This paper presents a robust and efficient automated system for detecting ships in high-resolution optical satellite imagery. By combining mathematical morphology, wavelet transforms, and Radon transforms, the system effectively handles challenging conditions like cloud cover and sea clutter, offering a valuable tool for maritime surveillance.
What Problem Does This Paper Address?
Maritime security is a critical concern, encompassing traffic surveillance, illegal fishing prevention, and pollution monitoring. While Vessel Monitoring Systems (VMS) exist, they don't cover all vessels, especially smaller ones. Satellite imagery offers a way to monitor these vessels, but detecting small ships in complex ocean environments (like those with clouds or wave patterns that mimic ships) is challenging. Existing methods often struggle with false alarms or are computationally expensive. This paper aims to develop an automated, efficient, and robust system to overcome these limitations.
Methods
The proposed system employs a three-stage approach:
- Preprocessing:
- Image Tiling: Large SPOT 5 images are divided into smaller tiles for faster processing and better local information handling.
- Cloud Masking: A statistical approach using histogram analysis identifies and masks cloud pixels, assuming cloud pixels are the brightest.
- Contrast Enhancement: Linear contrast stretching is applied to improve the visibility of features, using the modal value as the minimum intensity.
- Prescreening: This is the core detection phase, focusing on identifying potential ship pixels.
- Connected Operators (Mathematical Morphology): The system utilizes connected operators, which are powerful for image filtering and simplification. They work by building a "component tree" representing connected regions in the image at different grey levels.
- Filtering Strategy: The component tree is pruned based on criteria like "height" (brightness difference) and "area" (size of the connected component). This step aims to isolate bright, small features that are likely ship targets, while removing noise and false positives like wave crests. An adaptive threshold is used for height filtering to account for varying image conditions.
- Postprocessing: This stage refines the detections from the prescreening phase.
- Wavelet Transform (DWT): Used to analyze the image at different scales, enhancing the signature of ship targets and distinguishing them from surrounding noise. The Haar wavelet is chosen for its suitability in spot detection.
- Radon Transform (RT): Applied to further enhance linear features, particularly the wakes of ships, which can be strong indicators. The RT helps to differentiate ship wakes from other linear features like wind-wave crests.
- Logistic Regression Model: A statistical model is used to assign a "membership probability" to each detected target, classifying it as either a "ship" or "other." This probability is based on features extracted from the DWT and RT.
Datasets
The algorithm was calibrated and validated using SPOT 5 panchromatic satellite imagery with 5-meter resolution.
- Calibration Dataset: 15 SPOT 5 images were used for model calibration.
- Validation Dataset: 37 SPOT 5 images, acquired over the Exclusive Economic Zone (EEZ) of French Guiana, were used for performance evaluation. These images were not used in the model building process.
- Operational Validation: The system was also validated in a pre-operational context over Rijeka Harbour, Croatia, using one-eighth of a SPOT 5 scene.
Key Insights and Ideas
- Component Trees for Robust Detection: The use of connected operators and component trees is a novel and effective way to simplify images and extract relevant features for ship detection, especially in cluttered environments.
- Multi-stage Approach for Accuracy: Combining preprocessing, a morphology-based prescreening, and signal processing techniques (wavelet and Radon transforms) in postprocessing creates a powerful pipeline that significantly reduces false alarms.
- Adaptive Thresholding is Crucial: The development of an adaptive threshold for height filtering in the prescreening phase is key to handling variations in image brightness and contrast across different scenes.
- Wavelets and Radon Transforms Complement Each Other: DWT excels at multi-scale analysis and noise reduction, while RT is effective at highlighting linear features like ship wakes, making their combination highly beneficial for discrimination.
- Human-in-the-Loop for Final Decision: The system provides membership probabilities, allowing an experienced operator to make the final decision, balancing sensitivity and false alarm rates.
Key Results
- The algorithm is efficient, taking only 5-10 minutes for processing a large SPOT 5 image on a standard 2 GHz Pentium IV machine.
- In the validation over French Guiana, with 79 ship targets identified by human operators, the algorithm achieved:
- A detection rate (DR) of 89.8% with a probability threshold of 30%.
- A detection rate (DR) of 63.3% with a probability threshold of 70%, but with a significantly reduced false alarm rate (FAR) of 34.1%.
- The system demonstrated good performance in an operational validation over Rijeka Harbour, with only five false alarms for targets with membership probabilities greater than 70%.
- The algorithm showed a significantly lower FAR (one false positive for 100% DR) compared to a previous approach using neural networks and genetic programming on similar challenging images.
Limitations and Future Work
- False Alarms from Wind Turbulences: A significant source of false alarms is local wind turbulences and wave patterns that can mimic ship signatures.
- Multiple Detections of Large Ships: Large ships can sometimes be detected as multiple smaller targets, inflating the false positive count.
- Ground Truth Data: The unavailability of comprehensive ground truth data made precise validation challenging in some cases.
Future work suggested includes:
- Automatic Wake Detection: Developing a module to automatically detect ship wakes would further confirm detections and potentially estimate ship heading and speed.
- Bounds on Target Area: Implementing bounds on target area (size) to reduce multiple detections of large ships.
- Integration with Other Systems: Incorporating information from other maritime monitoring systems (like VMS) and cross-cuing with other sensors (like SAR) for improved validation and detection.
- Adaptation to Different Resolutions: Adapting the algorithm to various ship sizes and types, and testing it on very high spatial resolution optical imagery (sub-metric).
- Automatic Threshold Determination: Further research into automatically determining thresholds, particularly for cloud screening and height criteria.
- Enhanced Discrimination: Investigating additional characteristics and directional morphological filters to better discriminate true targets from wind-wave crests.
Summary updated