Research on Band Selection Method Based on Subspace Division and Visual Recognition
JIN Chun-bai1, YANG Guang1*, LU Shan2*, CHEN Qiang1 , 3, ZHENG Nan1
1. Aviation University Air Force, Changchun 130022, China
2. School of Geographical Science Northeast Normal University, Changchun 130022, China
3. 93116 Troop of PLA, Shenyang, 110000 China
Abstract：In the face of the increasingly abundant airborne and spaceborne hyperspectral sensors and the accompanying increase in hyperspectral data, the problems of excessive data volume and band redundancy have always been the major difficulties in hyperspectral image processing and interpretation. At the same time, the use of hyperspectral remote sensing technology to reveal camouflaged targets has always been the key point of modern remote sensing application technology research. In the detection of massive spectral data of ground objects and redundant spectral information, the design of appropriate data dimensionality reduction technology plays a vital role. The band selection method among the main methods of dimensionality reduction processing can not only reduce the spectral information of the image data without distortion, but also accurately distinguish the camouflage target and its background based on it. Today, the use of hyperspectral technology is an important technical means for military applications, and it is also a research hotspot of many scholars at home and abroad. It is a commonly used research method to use various indicators to calculate the different performances between the bands, and to select the most representative bands according to their parameters for feature identification or classification to test the pros and cons of the method. However, there are still few experimental studies on specific band selection methods for special features, such as vegetation camouflage targets. In the study, green steel plates, green camouflage nets, and green fake turf were selected and placed in a background environment containing green healthy vegetation, wet bare ground, and dry bare ground. Band selection and classification experiments were carried out by simulating camouflage targets and background objects in the real environment. First analyze the spectral curve, select a significant feature band. Secondly, the band screening is performed based on the sub-space divided according to the phase relationship between the band. The visual model is then established according to the image brightness of the property target. Finally obtains a band selection collection with relatively independence and optimal recognition. And adopt two kinds of common algorithms, support vector machine classifier and Mahalanobis distance classifier and full-band combination to compare the performance of classification experiments. It is found that the band selection results of the proposed method have improved accuracy and speed compared with the common algorithm selection band results and full band combination to perform classification. Compared with the application of full-band overall classification accuracy, the overall classification accuracy is increased by 4.5592% and 2.3648%, the Kappa coefficient is increased by 0.0594 and 0.0312, and the classification time is reduced by 6.83s. Experiments show that this method can effectively classify vegetation camouflage targets and background objects, and has great practical application value.
