A Study on Feature-Based Target Detection for Satellite SAR Imagery

Abstract

Feature-based target detection of Synthetic aperture radar (SAR) images is essential for monitoring areas in situations where large amounts of data are difficult to obtain, such as tactical regions. Many features have been studied for target detection in SAR images, but their performance depends greatly on the characteristics of the image, and the indiscriminate use of features reduces both efficiency and performance. Therefore, this study proposes a detection framework that ensures efficient and excellent detection performance in general SAR images using previously studied features. The proposed method consists of the steps of pretreatment, co-discrimination stage (CDS), and fine discrimination stage (FDS). First, in the pretreatment stage, speckle and noise removal are performed using Lee filter, and target pixels are detected in images using Constant false alarm ratio (CFAR) detectors with fixed thresholds, and then pixels detected through the Density-based spatial clustering with noise (DBSCAN) algorithm are clustered. There are still many false detection pixels by artificial and natural clutter in this pre-processed image, and next, CDS removes false detection pixels using simple features to remove them. Finally, in FDS, the classification performance of each feature's target and clutter is evaluated and features suitable for classification are selected. In addition, by constructing a feature space through Karhunen–Loève (KL) transformation, redundancy of selected features can be reduced and classification performance can be maximized. Additionally, it is ideal for a single target to form a single cluster through clustering, but if multiple dense targets exist in the image, the target detection performance decreases when multiple targets are included in one cluster due to limitations in resolution. To improve this, this paper proposes two features for distinguishing dense target clusters after clustering. Performance evaluation is performed by applying the proposed methods to real TerraSAR-X (TSX) images, and it has been confirmed that the proposed methods are effective in distinguishing targets of interest, excluding most of the misdetection clusters detected in the images.