Research On Hyperspectral Image Classification Based On Weakly Supervised Ensemble Learning

Hyperspectral remote sensing images containing continuous spectral bands and abundant spatial information have been widely used in many fields.Classification task is often a basic work of various applications,so how to improve the classification accuracy is a focus of current research.In recent years,some traditional machine learning methods have been widely used in hyperspectral image classification and achieved good classification results.More recently,deep learning-based methods have achieved remarkable results in hyperspectral image classification.In general,these methods require a large number of labeled samples to train reliable classification models.However,when the number of labeled samples is very small,the classification results often fail to meet the requirements.In addition,labeling remote sensing images requires professional knowledge,and manually labeling remote sensing images is laborious and time-consuming.To address aforementioned issues,this paper starts from weakly supervised learning and ensemble learning,and analyzes the problems of weakly supervised learning and ensemble learning in hyperspectral image classification.Then,the classification models or algorithms are improved according to the characteristics of hyperspectral data.Finally,this paper unifies the weakly supervised learning and the ensemble learning into one framework for hyperspectral image classification.The main work of this paper is summarized as follows:(1)A cascaded random forest is proposed for hyperspectral image classification.Two different enhancements are embedded into the random forest to improve the classification accuracy.In particular,the out-of-bag error is used to update the resampling weights of samples,which can avoids the problem that Boosting easily falls into overfitting.In order to select more effective features to build the classification model,a hierarchical random subspace method based on neighborhood rough sets is proposed to select features.(2)An active semi-supervised random forest is proposed for hyperspectral image classification.Both active learning and semi-supervised learning are embedded into the random forest to improve the performance of the model.In order to avoid the bias caused by active learning model,this paper uses clustering technique to mine the internal structure information of hyperspectral data for data division.In order to make full use of spatial information,this paper constructs a new active learning query function with spectral-spatial constraint,which improves the discriminability of selected samples.(3)A multi-view-based random rotation ensemble pruning method is proposed for hyperspectral image classification.The proposed method constructs an ensemble classifier based on multi-view learning to improve the generalization performance of the model.In order to improve the diversity of base classifiers in ensemble learning,the spectral features of correlation analysis are used to divide multiple views and random rotation is introduced to increase the diversity of base classifiers.In order to remove redundant base classifiers,this paper uses ensemble pruning method to pruning classifiers.(4)A multi-view-based label propagation ensemble method is proposed for hyperspectral image classification.In order to make full use of spatial information,morphological features are used to construct the graph representation in this paper.In order to alleviate the instability of single graph model,the label propagation results of different graphs are fused to obtain the labels of unknown samples.In order to verify the effectiveness of the proposed methods,experimental results on multiple real hyperspectral data sets show that the proposed methods have better classification accuracy than similar algorithms.Especially when the number of labeled samples is limited,these methods proposed in this paper can significantly improve the classification accuracy.