The Research On Remote Sensing Scene Classification Algorithm Based On Ensemble Learning

In recent years,remote sensing scene classification based on deep learning has become one of the hot spots of research in remote sensing,and the classification models represented by convolutional neural networks have achieved high accuracy and efficiency in remote sensing scene classification.However,the sizeable inter-class similarity of remote sensing scene images may lead to inconsistent classification results.In this paper,with remote sensing scene classification as the research background,a new loss function is proposed to alleviate the class ambiguity problem and improve the classification accuracy of remote sensing scenes.For remote sensing scenes with different spatial resolutions,a remote sensing scene classification method based on multi-scale image pyramids is proposed.And a multi-output integrated learning algorithm is designed for the problem that a single model cannot break through the performance bottleneck.The main research contents of this paper are as follows.(1)Remote sensing scene classification method based on multi-scale image pyramid.Since image samples have different spatial resolutions of remote sensing scenes,the lower the image resolution is,the fewer feature details there are.Low-resolution images can analyze the overall content of the image,and high-resolution can explore the details of the image.It will be more advantageous to study them with different resolutions.In this paper,we construct image pyramids with multi-scale features for the same original image and input the three resolution images into three different convolutional neural networks to form an integrated network with complementary scale feature information.(2)Classification algorithm based on class ambiguity constraint of remote sensing scene images.The similarity between classes in remote sensing scene images is high,and different objects may appear in the same category with different scales and directions.At the same time,the same things may also appear in other scenes,thus causing class ambiguity.In this paper,to alleviate class ambiguity,a new loss function is proposed,in which the cross-entropy function is used as a classification constraint.The hyperparameter a is introduced to the cross-entropy loss function to reduce the loss of easily classified samples.The Top-k loss function is used as a class ambiguity constraint,which can improve the Top-k classification accuracy at the same time while ensuring the Top-1 accuracy.It focuses more on confusing samples,alleviates the class ambiguity problem of remote sensing scene images,and introduces migration learning into the model training to save training time.The effectiveness of the loss function proposed in this paper is verified by multiple comparison experiments on three remote sensing scene datasets.(3)Remote sensing scene classification algorithm based on multi-output integrated learning.The classification effect of a single convolutional neural network model on remote sensing scene classification dataset is not satisfactory.To improve the classification performance,this paper proposes a remote sensing scene classification method based on multioutput integrated learning,which achieves mutual compensation for unbalanced classification data by combining the advantages of multiple single network models.Then,the Top-1 predictions and Top-2 predictions of each sub-network output are fused,and the Top-2 predictions are used to correct the final results of the model.Comparative experiments on the integration algorithm and several existing algorithms were conducted on three remote sensing scene datasets.The experiments show that the multi-classifier integration algorithm has a better classification effect than a single classifier,and the accuracy comparison with SOTA method verifies the effectiveness of the algorithm in this paper.