基于多目标遗传随机森林特征选择的面向对象湿地分类

以多时相Landsat8影像和SRTM DEM为数据源,对南瓮河流域进行了面向对象湿地分类。为削弱高维特征集对分类精度的影响,提出一种多目标遗传随机森林组合式特征选择算法(MOGARF)进行特征集优化。利用Relief F算法对完整特征集进行特征初选,再以基于随机森林的封装式多目标遗传算法进一步提取优化特征集。将所得特征集结合随机森林分类法提取湿地信息。并将结果分别与基于完整特征集和仅采用Relief F算法及Boruta算法提取的优化特征集的3种随机森林分类结果对比。试验结果表明,采用MOGARF算法特征选择后,特征维度降低至原来的10%,且分类精度最高,总体精度为92.61%,比其他分类方案提高0.35%~1.94%,Kappa系数为0.907 5,袋外误差为7.77%,比其他分类方案降低0.91%~1.48%。利用MOGARF特征选择的随机森林分类法是湿地分类的有效方法。

Object-oriented Wetland Classification Based on Hybrid Feature Selection Method Combining with Relief F, Multi-objective Genetic Algorithm and Random Forest

Recently, researchers adopted object-oriented method to extract wetland distributions. Multi-temporal and multi-sources of data can facilitate the extraction process but meanwhile it enlarges the amount of features. It needs a large quantity of experiment based on the expert knowledge to determine the optimal feature sets and the threshold values. In order to improve the classification accuracy and relief the researchers from large amount of work, a filter-wrapper hybrid feature selection method combining relief F, multi-objective genetic algorithm and random forest was proposed, which was a two-step method. In the first step, relief F algorithm was adopted to select features with class separability. In the second step, multi-objective genetic algorithm based on random forest (MOGARF) was built. Four measures such as out-of-bag (OOB) error of random forest algorithm, dimension of the feature space, correlations among features and the variable weight of relief F algorithm were acted as four objectives of MOGA. The probability whether the feature was expressed was determined by the variable importance measures from random forest algorithm. The crowded distance of each feature collection was calculated and the feature collection with the least crowded distance was the optimal feature set. Nanweng river basin was taken as the study site. Object-oriented classification using random forest classifier was conducted based on the optimal feature set. Then the result was compared with three other random forest classification schemes by using the entire feature set or the feature set selected by relief F algorithm or the Boruta algorithm. The classification scheme with MOGARF had the best performance and the feature dimension was reduced to 10% of the entire one. The overall accuracy reached 92.61% which was 0.35%~1.94% higher than those of the other three schemes with Kappa coefficient of 0.9306. The OOB error of MOGARF was 7.77% which was 0.91%~1.48% lower than those of the other schemes. All these indicated that the MOGARF feature selection method was an effective feature selection method when it was combined with random forest classifier.