基于双树复小波分解的Boosting集成学习土地覆被分类研究

近年来，集成学习（Ensemble Learning，EL）分类方法成为土地覆被分类的研究热点，尤其是Boosting集成分类方法具有分类精度高、泛化能力强，在土地覆被分类中得到了显著的应用。但是，Boosting集成分类方法对噪声很敏感，如果训练样本含有噪声时，Boosting算法可能会失效，这是该方法的局限性。为了解决Boosting集成方法在土地覆被分类中存在的问题，有效克服噪声的影响，减少分类结果中的“椒盐”现象和提高分类精度，提出了基于双树复小波分解的Boosting集成学习分类方法。该方法对影像的光谱波段进行一层双树复小波分解，降低图像的噪声，将分解后的各波段作为Boosting集成学习的输入，得到最终的分类结果。实验先后比较了GBDT、XGBoost、LightGBM 3种Boosting集成学习算法在SPOT 6和Sentinel-2A影像上的分类效果。结果表明：①在SPOT 6影像上，3种Boosting集成算法总体分类精度均高于90%；DTCWT-LightGBM分类总体精度最高，达到94.73%，Kappa系数为0.93，比LightGBM总体精度提高了1.1%，Kappa系数提高了0.01；LightGBM分类总体精度比XGBoost分类总体精度提高了1.99%，Kappa系数提高了0.03，比GBDT分类总体精度提高了2.9%，Kappa系数提高了0.04；②在Sentinel-2A影像上，DTCWT-LightGBM分类总体精度最高，达到93.25%，Kappa系数为0.91，比LightGBM分类总体精度提高了1.53%，Kappa系数提高了0.01；LightGBM分类总体精度比XGBoost分类总体精度提高了1.14%，Kappa系数提高了0.02，比GBDT分类总体精度提高了2.53%，Kappa系数提高了0.03；③基于双树复小波分解的Boosting集成学习分类方法，降低了影像的噪音，减少了分类结果中存在的“椒盐”现象，区域一致性更强，提高了分类精度。

Study on Boosting Ensemble Learning Land Cover Classification based on Dual-Tree Complex Wavelet Transform

Ensemble Learning （EL） classification method has become a research hotspot of land cover classification in recent years. Boosting Ensemble Learning classification method has high classification accuracy and strong generalization ability particularly， which has been significantly applied in land cover classification. However，Boosting Ensemble classification method is sensitive to noise. If the training sample contains noise， Boosting algorithm may lose effectiveness， which is the limitation of the method. In order to solve the problems existing in Boosting Ensemble method in the classification of land cover，effectively overcome the influence of noise， reduce the salt and pepper phenomenon in the classification results and improve the classification accuracy， a Boosting Ensemble Learning classification method based on the dual-tree complex wavelet transform is proposed. In this method， the spectral band of the image is transformed by a layer of dual-tree complex wavelet to reduce the image noise. The extracted low-frequency features are taken as the input of Boosting Ensemble Learning to obtain the final classification result. Boosting Ensemble Learning GBDT， XGBoost and LightGBM algorithms are respectively compared classification accuracy and efficiency for SPOT6 and Sentinel-2A image. The results show as follow： （1）For SPOT6 image， the overall classification accuracy of the three Boosting Ensemble algorithms is higher than 90%.LightGBM algorithm after DTCWT has the highest classification accuracy.The overall classification accuracy and Kappa coefficient are 94.73% and 0.93 respectivesly.Two precision values are higher than without the transform of dual-tree complex wavelet by 1.1% and 0.01. LightGBM algorithm classification accuracy and Kappa coefficient are higher than the XGBoost algorithm by 1.99% and 0.03，and are higher than the GBDT algorithm by 2.9% and 0.04.（2） For sentinel-2A image， LightGBM algorithm after DTCWT has the highest classification accuracy.The overall classification accuracy and Kappa coefficient are 93.25% and 0.91 respectivesly.Two precision values are higher than without the transform of dual-tree complex wavelet by 1.53% and 0.01. LightGBM algorithm classification accuracy and Kappa coefficient are higher than the XGBoost algorithm by 1.14% and 0.02，and are higher than the GBDT algorithm by 2.53% and 0.03.（3） After the transform of dual-tree complex wavelet， the Boosting Ensemble Learning classification can reduce the noise of the image， reducing the salt and pepper phenomenon in the classification results， having stronger regional consistency， improving the classification accuracy.