基于SVC和过采样的类别非均衡农业高光谱数据分类

Classification of Unbalanced Agricultural Hyperspectral Data based on SVC and Oversampling

摘要:

系统研究了农业高光谱数据中少数类的分类质量问题。为了提升少数类的分类质量，提出采用过采样SMOTE技术增加少数类新样本，同时研究了SMOTE技术中新样本生成策略和少数类采样倍率对高光谱数据中少数类分类结果的影响，以及不平衡数据集上分类器与模型的匹配度。在新的采样数据集上采用多类分类SVC技术对少数类分类，提升了非均衡高光谱数据集中少数类的分类质量。在真实数据集上进行了试验验证，并对不同的分类方法和系统参数进行了试验对比和分析，结果表明，本文方法能够显著地提高非均衡高光谱数据中少数类分类效果，平均分类精度不小于0.82，平均召回率提升幅度为11.11%～26.15%，F1提升幅度为5.81%～40.85%。

Abstract:

Hyperspectral technology is widely used in agricultural natural resources such as agro-ecological environment and land resource protection. Spectral imaging technology can effectively classify and identify ground objects. Therefore, the classification of hyperspectral data is one of the important contents of hyperspectral research. Category non-equilibrium problem is a common problem in agricultural hyperspectral data, and the classification quality of minority classes has great significance for the effective classification of hyperspectral data. However, the classification of minority classes is affected by the dominant majority classes. The general classification algorithm tends to the dominant majority classes classification, so that minority classes are usually submerged in the majority classes, bringing great challenge to classification accuracy and recall rate of the minority classes. The classification quality of the minority objects was studied in agricultural hyperspectral data. In order to improve the classification quality of minority classes, an oversampling technique SMOTE was proposed to add new samples for the minority classes. At the same time, the effects of new sample generation strategy and minority instance sampling rate on the classification results of minority samples in the agricultural hyperspectral data and the matching degree between the classifier and the model on the unbalanced data set were systematically studied. A multi-class classification SVC technique was used to classify minority classes on a new sampling data set, and it improved the classification accuracy of the minority classes in unbalanced agricultural hyperspectral dataset. The experimental verification was carried out on the real data set, and different classification methods and system parameters were tested and compared. The experimental results showed that the proposed method can greatly improve the effect of minority classification in unbalanced agricultural hyperspectral data. The weight precision can reach above 0.82, the weight recall rate was obviously improved from 11.11% to 26.15%，and F1 was increased from 5.81% to 40.85%. The method can provide a reference for the unbalanced agricultural hyperspectral data to improve the classification effect systematically.