Extraction of rice from remote sensing images based on deep heterogeneous transfer learning
In order to achieve high-quality construction and reuse of rice extraction models from remote sensing images based on het-erogeneous with only unlabeled samples in the target domain,a deep heterogeneous feature transfer learning model based on temporal and spatial constraints was constructed.Firstly,unlabeled sample groups in the source domains and target domains were constructed based on spatial location,and their deep features were extracted;secondly,in order to reduce the negative transfer impact of features and realize precise transfer of heterogeneous features,a heterogeneous feature transferring model was constructed by using a composite loss function including corresponding sample feature conversion loss,corresponding sample feature regular loss,and sample recon-struction loss;finally,in order to improve the accuracy of classification,a semi-supervised classification model was established,and HingLoss was introduced to eliminate the impact of wrong pseudo labels.The results showed that the research method could realize sam-ple feature transfer between images at different resolutions.Compared with the case without feature transfer,the accuracy rate was im-proved by 27.68 percentage points,and the F1 score was improved by 17.3 percentage points.
unlabeled sampleextraction of ricehigh resolution remote sensing imagesdeep heterogeneous transfer learning
NETL
NSTL
万方数据
