Research On Improved Self-training Algorithm And Applied To Remote Sensing Image Classification

Remote sensing image classification plays a key role in many applications.These include geographic image retrieval,vegetation mapping,land cover determination,natural disaster detection,environmental detection,urban planning,geospatial object detection and so on.To improve the classification accuracy of remote sensing image is the primary task of the above applications.In the hyperspectral remote sensing image,when the number of training samples is limited,the classification accuracy increases first with the increase of the number of image bands,and then decreases with the increase of the number of image bands after reaching a certain extreme value.It is difficult for the neural network to learn the representation of images under the condition of having unbalanced or few data.When using high-dimensional eigenvectors,the number of samples of each category is required to be higher than the feature dimension,however,it is difficult to obtain samples in hyperspectral classification.Based on the above issues,the self-training algorithm was improved in this thesis,which can still train a neural network model with stronger generalization ability and robustness even when the number of training samples is limited.The main contents of this thesis are as follows: 1.A strong constraint mechanism in the self-training algorithm to filter the low-quality pseudo label data,so that the model generalization ability will become stronger as the number of algorithm iteration increasing.The strong constraint mechanism is to strictly control the pseudo label quality in the early stage of the algorithm and relax the filtering conditions for pseudo label in the late stage of the algorithm,to reduce the number of model iterations.2.Early stopping strategy and transfer learning are used to make the model to converge fast.3.A list is used to record the number of sample growth,and then data augmentation technique is used to balance the data and reduce the bias of the model.Extensive experiments are done in this thesis to verify the feasibility of the algorithm.On a 50% ratio of AID training data and the NNWPU-RESISC45 as an unlabeled dataset,the proposed algorithm achieves 96.01% on the rest of the AID data.On a 20% ratio of NNWPU-RESISC45 data and the AID as an unlabeled dataset,this algorithm achieves 93.03% on the rest of NNWPU-RESISC45 data.And on a single NNWPU-RESISC45 data set,the data set is split into different ratios,and compared with other self-training algorithms,the accuracy of this algorithm is the best.