二元数据融合下的高光谱影像地物识别

The Combination of Sentinel-2 and GF-5 Spectral Image Was Used for Ground Object Recognition

摘要: 针对目前高分五号AHSI (the Advanced Hyperspectral Imager)相机存在的330个波段但是空间分辨率较低的问题，提出了高光谱影像结合Sentinel-2号卫星的3波段(560 nm)、4波段(664.5 nm)波段分别进行融合的方法，来解决高分五号卫星面临的地物分类精度不足的问题，进而提高精度。为了避免处理的过程中出现维数灾难的问题，首先对高分五号AHSI影像和Sentinel-2号影像分别预处理，再采用G-S (Gram-Schmidt)融合，最后利用随机森林算法进行分类识别，利用用户精度(User Accuracy, UA)、生产者精度(Producer Accuracy, PA)、整体分类精度(Overall Accuracy, OA)以及Kappa系数来评价精度。结果表明，利用随机森林算法对经过改进的数据在总体分类精度上有显著提高，相对于原始高光谱数据分别能够提升9.33%和10.17%，其中整体分类精度分别为97.57%和98.35%，KAPPA系数分别为0.9625、0.9731。

Abstract: Based on the current GF-5 AHSI (the Advanced Hyperspectral Imager) 330 bands exist the camera but low spatial resolution of the problem, put forward the Hyperspectral image combining satellite Sentinel-2 satellite’s 3 band (560 nm), 4 band (664.5 nm) band fusion method, to solve the high score 5 satellite faces the problem of insufficient feature classification accuracy, precision can be improved. In order to avoid dealing with the problem of dimension disaster, appear in the process of the first to score 5 AHSI image and Sentinel-2 image preprocessing, and then use the G-S (Gramm-Schmidt), the final random forests for classification recognition, using User Accuracy (User Accuracy, UA), the precision of producers (Producer Accuracy, PA), the Overall classification Accuracy (Overall Accuracy, OA) and Kappa coefficient to evaluate precision. The results show that compared with the original hyperspectral data, the overall classification accuracy of the fused high- resolution No. 5 image can be improved by 9.33% and 10.17%, respectively, and the overall classification accuracy of the fused high-resolution No. 5 image is 97.57% and 98.35%. The KAPPA coefficients were 0.9625 and 0.9731.