Vessel Classification from Overhead Imagery Using the Random Forest Algorithm
Abstract
Classification of satellite ship imagery is an active topic of research, and multiple types of classifiers have been considered over the years. This study explores the viability of the random forest algorithm in vessel type classification and compares performance to that obtained in earlier work by Rainey et al., published in a 2012 SPIE Proceedings article, and by Parameswaran and Rainey, published in a 2015 SPIE Proceedings article. Random forest is advantageous due to its relative ease of use, resistance to overfitting, and built-in model validation. Results indicate that random forest performance is comparable to or better than time-tested machine learning methods, such as support vector machines, when applied to preprocessed vessel images. Feature extractors that capture spatial information yielded highest accuracies. Previous workhas indicated that the visual bag of words (VBOW) representation is flexible and effective in feature coding the vessel images. Therefore, in this work various weighting schemes augmented the VBOW, which was evaluated on both original and preprocessed vessel image datasets as input to the random forest, with limited success.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 2020
- Accession Number
- AD1127094
Entities
People
- Thomas J. Shaheen
Organizations
- Naval Postgraduate School