The Application Of Deep Learning Methods In Land-Use Classification And Sea Surface Wind Retrieval From SAR Imagery

In recent years,deep learning has made amazing progress in computer vision,speech recognition,Natural Language Processing and other fields.These successes are generally based on a large number of labeled data.In practical applications,a large number of labeled data are hard to get.The situation of insufficient label data under specific problems is generally achieved through transfer learning or semi-supervised learning,and achieved good results.In this paper,the application of transfer learning and semi-supervised learning algorithm is carried out in the application of land-use classification and SAR sea surface wind field retrieval.The main work is as follows:First,implementation of transfer learning algorithm and its application in the land-use classification.Transfer learning is an algorithm that can transfer the information learned in the source domain to the related fields to improve the performance of the learner.At present,natural image tagged data sets,such as ImageNet,are easy to get.Some excellent deep learning models,such as ResNet,have achieved surpassing human performance on these datasets.Taking into account that natural images and SAR images do not obey the same distribution,but they have similar spatial correlation.We transfer the ResNet which is trained on ImageNet to the land-use classification through the methods of training by layer-by-layer,fine-tuning and so on,and achieve 96.19% accuracy on the UC-Merced dataset.Second,implementation of supervised learning algorithm and its application in the land-use classification.In the actual application,labeled data is not easy to obtain,but the acquisition of unlabeled data are often very easy.How to apply a small amount of data contained in the label and a large number of unlabeled data,namely semi supervised learning,has gradually become a hot research topic in the field of deep learning.The process of training convolution is transformed into the process of training DAE by using unlabeled data by using the similarity between DAE network internal operation and convolution operation and the performance of the proposed algorithm is verified on the UC-Merced dataset.Third,Application of sea surface wind retrieval using data from SAR.The above algorithms are applied to the retrieval of SAR sea surface wind field.The wind speed error of the semi supervised learning algorithm and the transfer learning algorithm is less than 4m/s,and the wind direction error is within 20 degrees.The best performance of the two algorithms can be achieved when the wind speed error is 1.895m/s and the wind direction error is 15.754 degrees.