FASTENER Feature Selection for Inference from Earth Observation Data.

Area of Science:

• Machine Learning
• Data Science
• Remote Sensing

Background:

• High-dimensional data presents challenges for machine learning model accuracy and efficiency.
• Feature selection is crucial for optimizing model performance and reducing computational costs.
• Existing methods may not effectively balance feature reduction with classification accuracy.

Purpose of the Study:

• Introduce FASTENER, a novel multi-objective feature selection algorithm.
• To maximize machine learning model accuracy using a minimal set of features.
• To improve upon existing feature selection techniques in terms of speed and performance.

Main Methods:

• FASTENER employs a multi-objective approach within an iterative genetic algorithm framework.
• It utilizes entropy-based measures, including mutual information, during the crossover phase.
• The algorithm was evaluated on Sentinel-2 land-cover classification data and open datasets.

Main Results:

• FASTENER demonstrates faster convergence to optimal feature subsets compared to POSS, DT-forward, and FS-SDS.
• It achieves superior classification accuracy over current similarity and information theory-based methods in earth observation.
• The algorithm yields state-of-the-art results for land-cover classification and comparable/superior results on other datasets.

Conclusions:

• FASTENER offers an efficient and effective solution for feature selection in high-dimensional data.
• The algorithm provides state-of-the-art performance, particularly in remote sensing applications.
• FASTENER is broadly applicable to any supervised machine learning task.