基于多模态多目标优化的端元束提取方法研究

Endmember Bundle Extraction Method Based on Multi-modal and Multi-objective Optimization

为解决高光谱影像受传感器及分辨率的影响所产生的光谱变化给解混造成的困扰,提出基于多模态多目标优化的端元束提取方法(MOPSOSCD)。对高光谱图像进行标号编码,采用基于索引的环形拓扑结构进行邻域的个体交互,通过邻域最优改进粒子群速度更新方式并整数化粒子位置更新。同时,根据高光谱图像空间特征,通过改进决策空间拥挤距离提高决策空间的多样性,再结合目标空间的拥挤距离进行综合排序,实现多模态多目标优化的粒子筛选。当粒子定向移动概率pm为0.2、粒子数P为30及迭代次数M为400时,算法在MUUFL数据集上均方根误差(RMSE)及平均光谱角距离(mSAD)分别为0.0088、0.1112。通过对比试验,本文方法相较于VCA、DPSO等方法具有更高的提取精度和效率,为高光谱解混提供了更加准确的端元束提取方法。

Hyperspectral image has continuous spectral information of ground objects,which is an essential means of remote sensing monitoring.On this basis,the endmembers of the features can be extracted by decomposing the mixed pixel spectrum and exploring the degree of each endmember participates in the mixing.However,specific spectral changes cause trouble for spectral unmixing due to the sensor and the image’s resolution.To solve this problem,an endmember bundle extraction method based on multi-modal and multi-objective particle swarm optimization by special crowding distance(MOPSOSCD)was proposed.Firstly,for a three-dimensional hyperspectral image,the label coding was carried out pixel by pixel,and the index-based ring topology was used for individual interaction in different neighborhoods.Secondly,for particle velocity and position update,the position update method of PSO was adopted and the particle swarm velocity update method and the integer particle position update were improved through neighborhood optimization.The objective function selection was measured by two RMSEs,that was,the unconstrained least squares method was used to solve the RMSE of the abundance map anti-mixing and the original map,and the fully constrained least squares method was used to solve the RMSE of the abundance map anti-mixing and the original map.At the same time,according to the spatial characteristics of hyperspectral images,decision space diversity was improved by improving the crowded distance of decision space.Finally,the crowding distances of the decision space and the target space were combined and sorted,and the particles were updated according to the sorting results.When the particle directional movement probability was 0.2,the number of particles was 30,and the number of iterations was 400,the results of RMSE and mSAD on the MUUFL dataset were 0.0088 and 0.1112,respectively.Through the comparative test,the method had higher extraction accuracy and efficiency than VCA and DPSO,providing a more accurate end beam extraction method for hyperspectral unmixing.