Ship Detection Feature Analysis in Optical Satellite Imagery through Machine Learning Applications

Ship detection remains an important challenge within the government and the commercial industry. Current research has focused on deep learning and has found high success with large labeled datasets. However, deep learning becomes insufficient for limited datasets as well as when explainability is required. There exist scenarios in which explainability and human-in-the-loop processing are needed, such as in naval applications. In these scenarios, handcrafted features and traditional classification algorithms can be useful. This research aims at analyzing multiple textures and statistical features on a small optical satellite imagery dataset. The feature analysis consists of Haar-like features, Haralick features, Hu moments, Histogram of Oriented Gradients, grayscale intensity histograms, and Local Binary Patterns. Feature performance is measured using 8 different classification algorithms, including K-Nearest Neighbors, Logistic Regression, Gradient Boosting, Extreme Gradient Boosting, Support Vector Machine, Random Decision Forest, Extremely Randomized Trees, and Bagging. The features are analyzed individually and in different combinations. Individual feature analysis results found Haralick features achieved a precision of 92.2% and were computationally efficient. The best combination of features was Haralick features paired with Histogram of Oriented Gradients and grayscale intensity histograms. This combination achieved a precision score of 96.18% and an F1 score of 94.23%