【摘要】 针对高光谱数据波段间存在强相关性与高冗余性导致机器学习识别难以获得理想精度的问题，该文提出一种结合分段主成分(SPCA)和粒子群优化向量机(PSO-SVM)的沿岸湿地植被识别方法。选取洱海五种典型沿岸湿地植被(狗牙根、杨柳、池杉、菰、槐叶萍)的高光谱数据作为目标样本，首先计算波段间相关系数得到相关系数矩阵，将全波段(400.55～848.58 nm,波段数为217)划分为4个分段，分别对各分段及全波段进行主成分变换，将各分段的前5个主成分(累计贡献率>0.98)作为光谱特征。最后，构建PSO-SVM模型进行识别，PSO-SVM模型通过对惩罚参数(c)和核函数参数(g)的自动迭代寻优，有效避免人工调参过程中存在的主观性问题。结果表明：模型在第三分段(552.46～678.83 nm,波段数为62)的精确率PR为0.929 2,召回率RR为0.925 3,F1分数为0.926 7及AUC值为0.914 4;识别准确率达到0.916 7,高于全波段的0.866 7;同时识别耗时1.84 s,少于全波段的2.65 s。故可用第三分段的62个波段代替全波段的217个波段对植被进行更为准确和快速的识别。更多还原

【Abstract】 Aiming at the challenges posed by strong inter-band correlation and high redundancy in hyperspectral data—which often reduce the classification accuracy of machine learning models—this study proposes a wetland vegetation classification method that integrates segmented principal component analysis(SPCA) with a particle swarm optimization-based support vector machine(PSO-SVM). Hyperspectral reflectance data from five representative wetland vegetation species along the eastern shore of Erhai Lake(cynodon dactylon, willow, taxodium ascendens, zizania latifolia, and salvinia natans) were selected as target samples. The correlation coefficient matrix was computed to analyze spectral redundancy, and the full spectral range(400.55～848.58 nm, 217 bands) was divided into four segments accordingly. Principal component analysis was then applied to each segment and the full band, with the first five principal components(cumulative contribution>0.98) from each segment extracted as spectral features. Subsequently, a PSO-SVM model was constructed for vegetation classification, in which the PSO algorithm was employed to automatically optimize the penalty parameter(c) and kernel parameter(g),effectively mitigating the subjectivity and inefficiency of manual parameter tuning. Experimental results show that in the third segment(552.46～678.83 nm, 62 bands),the model achieved a precision of 0.929 2,recall of 0.925 3,F1-score of 0.926 7,and AUC of 0.914 4;the classification accuracy reached 0.916 7,surpassing that of the full spectrum(0.866 7),with a shorter computation time of 1.84 s compared to 2.65 s. These findings suggest that the 62 bands in the third segment can effectively replace the full 217 bands for more accurate and efficient vegetation classification.更多还原