Remote Sensing Image Terrain Classification Based On Active Learning

Active Learning(AL)is an iterative learning paradigm in machine learning field.In each iteration,the former trained model is used to predict unlabeled candidate instances at first,then some most informative instances are selected and labeled as new training set to improve the generalization ability.In this thesis,we mainly combine active learning with three kinds of classification models to deal with terrain classification problem of remotely sensed image.Experimental datasets are Polarimetric Synthetic Aperture Radar(Pol SAR)image and hyperspectral image.Pol SAR is an advanced radar system,sending coherent multichannel microwave actively and receiving the scattering microwaves rebounded from object,which can observe land-cover and land-use in all-weather and all-time.Hyperspectral imaging system can record continuous spectrum of terrain target with high spectral resolution.It has been widely used in ground observation.In this thesis,active learning is combined with three kinds of classification models,including Extreme Learning Machine(ELM),Convolutional Neural Networks(CNN)and Gradient Boost Decision Tree(GBDT).Our works are mainly as follows:For ELM algorithm,an Online Active ELM(OA-ELM)is proposed in order to improve the generalization ability and training efficiency.OA-ELM is online active learning algorithm,which merges AL-ELM and Online Sequential ELM(OS-ELM).In addition,based on the output property of ELM,a new sampling method called discrepancy sampling is proposed,which can make full use of the ELM's real-valued outputs directly.Discrepancy sampling measures the uncertainty level by calculating the difference between largest and second largest outputs.The experimental results on the handwritten dataset and the two Pol SAR datasets verify that the proposed discrepancy sampling is effective,and OA-ELM is accurate and fast.For CNN algorithm,an Active Learning for Weighed CNN(AL-WCNN)algorithm is proposed to improve the generalization ability,which combines sample weighted method and active sampling method.In sample weighted method,at first the former learned model is used to predict the importance of training sample,and then the sample weight of the loss function is modified according to its importance.Margin sampling is adopted in AL-WCNN to select more informative samples as training set in order to improve accuracy.The experimental results show that AL-WCNN achieves good classification performance on Pol SAR images and hyperspectral images.For GBDT algorithm,margin sampling is adopted in AL-GBDT algorithm.In addition,a post-processing method is proposed to revise classification result via Gaussian Filtering,which is called Gaussian Filter for Label Revising(GFLR).The two parts form an semi-supervised active learning system,which improves classification performance significantly.Experimental results on two Pol SAR images and a hyperspectral image indicate that AL-GBDT improves accuracy by selecting training samples actively,and GFLR further improves accuracy by incorporating spatial neighborhood information.In this thesis,three kinds of active learning algorithms are proposed,including OA-ELM,AL-WCNN and AL-GBDT.The terrain classification results demonstrate the effectiveness of the proposed algorithms.