Abstract
This paper proposes a multi-stage machine learning architecture designed to assist human "Shark Spotters" in South Africa by automating the detection of sharks in coastal waters. By combining translation, rotation, and scaling (TRS) invariant feature extractors with an ensemble of fused classifiers, the research demonstrates the mathematical feasibility of distinguishing sharks from other marine life regardless of their orientation or movement.
Problem Statement: The Limitations of Human Observation
In Cape Town, the "Shark Spotters" program relies on human observers positioned on mountainsides to scan the water with binoculars. This manual method is susceptible to two critical errors: False Positives, which lead to unnecessary beach evacuations and erode public trust, and False Negatives, which can result in fatalities.
Environmental factors such as rain, glare, and murky water significantly hinder human observation. The current system communicates safety levels through a flag-based protocol, where different colors indicate varying levels of shark activity and visibility.
The proposed "semi-automated" support system aims to provide a reliable second opinion. The primary technical challenge lies in the dynamic nature of sharks; they change shape, rotate, and move across the frame. Consequently, any detection system must be robust to geometric transformations while maintaining high accuracy in noisy maritime conditions.
Proposed Methodology
The architecture follows a "Bottom-Up" approach, progressing from raw pixels to a final decision through four distinct stages:
1. Image Preprocessing (IP)
To mitigate noise from wind and weather, the system employs Median Filtering. Otsu’s Algorithm is used for binarization, while the Random Walker Technique performs segmentation. This graph-based approach effectively isolates the shark's shape from the complex aquatic background.
2. Feature Extraction (FE)
Ensuring the system recognizes a shark regardless of its orientation (swimming left, right, or circling) requires Translation, Rotation, and Scaling (TRS) invariance. Three methods are utilized:
- Complex Moment Invariants (CMI): Provides rotation and scaling invariance.
- Generic Fourier Descriptors (GFD): Captures frequency-based shape information.
- Exact Legendre Moments (ELM): Extracts orthogonal space-based features.
3. Classification Ensemble
Rather than relying on a single model, the architecture suggests an ensemble of three classifiers with different underlying "philosophies":
- Artificial Neural Networks (ANN): Focused on minimizing training error.
- Genetic k-Nearest Neighbours (GKNN): Uses genetic algorithms to optimize the search for similar shapes.
- Support Vector Machines (SVM): Identifies the optimal global boundary between "shark" and "not shark."
4. Decision Fusion
The outputs of these classifiers are integrated using Dempster-Shafer Fusion (DSF). This mathematical framework is specifically chosen for its ability to handle uncertainty and reconcile conflicting evidence from multiple sources, making it ideal for high-stakes safety applications.
Validation and Results
Due to a lack of comprehensive, labeled aerial shark datasets, the research focused on validating the Feature Extraction stage. Using a sample set of shark images, transformed shark images, and other marine life (e.g., fish), the following was observed:
- Numerical Consistency: CMI values for a shark and its scaled/rotated counterparts were nearly identical (e.g., 9.0868 vs 9.0868), whereas values for non-shark marine life were significantly different (2.0727).
- Visual Confirmation: GFD and ELM plots showed that feature curves for transformed sharks overlapped almost perfectly, while the "fish" curve remained a clear outlier.
- Implementation: The framework was successfully implemented in Python using the
skimagelibrary, proving the theoretical feasibility of the pipeline.
Discussion and Future Work
While the theoretical framework is robust, data scarcity remains the primary hurdle for full implementation. Future efforts should focus on:
- Dataset Development: Creating a large-scale, labeled training dataset of aerial maritime imagery.
- Dynamic Tracking: Moving from static image analysis to video-based tracking to classify sharks based on movement patterns.
- Infrastructure: Deploying strategic cameras (e.g., on masts or balloons) to reduce reliance on human sightlines and improve data quality.
- Inter-Beach Coordination: Developing "hand-off" protocols to share tracking data as sharks move along the coastline.
The shift toward mathematical rigor over "black box" deep learning ensures that the model adheres to essential properties like surjectivity and uniqueness, providing a transparent and reliable tool for public safety.
Summary updated