Abstract

Feature/band selection (FS/BS) for target detection (TD) attempts to select features/bands that increase the discrimination between the target and the image background. Moreover, TD usually suffers from background interference. Therefore, bands that help detectors to effectively suppress the background and magnify the target signal are considered to be more useful. In this regard, three supervised distance-based filter FS methods are proposed in this paper. The first method is based on the TD concept. It uses the image autocorrelation matrix and the target signature in the detection space (DS) for FS. Features that increase the first-norm distance between the target energy and the mean energy of the background in DS are selected as optimal. The other two methods use background modeling via image clustering. The cluster mean spectra, along with the target spectrum, are then transferred into DS. Orthogonal subspace projection distance (OSPD) and first-norm distance (FND) are used as two FS criteria to select optimal features. Two datasets, HyMap RIT and SIM.GA, are used for the experiments. Several measures, i.e., true positives (TPs), false alarms (FAs), target detection accuracy (TDA), total negative score (TNS), and the receiver operating characteristics (ROC) area under the curve (AUC) are employed to evaluate the proposed methods and to investigate the impact of FS on the TD performance. The experimental results show that our proposed FS methods, as compared with five existing FS methods, have improving impacts on common target detectors and help them to yield better results.

Highlights

  • Hyperspectral imagery (HSI) have properties that provide scientists with various applications, such as crop and mineral identification in agriculture and geology, improved classification map production [1], subpixel target and anomaly detection [2,3,4,5,6,7], spectral unmixing [8], and data fusion

  • False alarm (FA) is important in target detection (TD), since it can reduce the costs of wrong decisions

  • In the initial and middle feature subsets, the proposed feature selection (FS) methods led to much lower false alarms (FAs) as compared to other methods

Read more

Summary

Introduction

Hyperspectral imagery (HSI) have properties that provide scientists with various applications, such as crop and mineral identification in agriculture and geology, improved classification map production [1], subpixel target and anomaly detection [2,3,4,5,6,7], spectral unmixing [8], and data fusion. Huge data volumes are produced due to the high number of spectral bands. Studies have been conducted on reducing the data dimensionality by feature/band selection (FS/BS) and feature extraction (FE) methods. The goal of dimensionality reduction (DR) is to reduce data volume, speed up computing, and improve the accuracy of analyses [9,10]. There is a new type of FE methods that is based on the deep learning concept, such as the convolutional neural network (CNN) and deep belief network (DBN), which are used to extract new high-level features in order to improve the classification accuracy [16,17]

Objectives
Methods
Results
Conclusion
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call