Application Research Of Transfer Learning In Remote Sensing Monitoring Of Water Source Risk Sources

At present,the environmental safety situation of drinking water in China is severe.Including agricultural land,mining land,industrial land,residential areas,etc,have become an important source of water pollution to varying degrees.Rapid,timely and effective dynamic monitoring of water source risk sources to further protect the water environment is a research focus in the field of remote sensing.The sample plays an important role in the remote sensing classification monitoring,but the sample selection work is time-consuming,labor-intensive and costly.Based on the idea of transfer learning,the application of sample migration can solve the problem of insufficient samples of remote sensing images.Taking the construction area which has an important influence on the water source environment as an example,the two methods of support vector machine and full convolution network are applied to study the feasibility of remote sensing image sample migration based on transfer learning.The main contents include:1)For the Landsat 8 data with a resolution of 30 meters,according to the actual characteristics of the features,the idea of layered extraction is adopted.Firstly,the features in the image with large changes in phenology are removed,and then the support vector machine classifier is used for other features extraction in the image.Combined with this idea.and the method of transductive transfer learning,sample migration applications are applied to data sources in different periods to verify the possibility of sample migration based on sample attributes.2)For the Sentinel 2 data with 10 m resolution and combined with the full convolution network,the migration of samples under different phenology and different imaging conditions was tested by the inductive and transductive transfer learning.Combined with the sample migration results of the full convolutional network,this method is used to extract the high-rise and low-rise building areas along the Central Line Project of South-to-North Water Diversion under the 10-meter resolution data.The all results show that:(1)In the face of sample attribute migration of 30-meter resolution data,a good sample migration effect can be obtained when the sample attribute and the data to be processed have strong consistency in the feature space,and sample migration has a strong application space.(2)In the case of sample migration with 10 m resolution data,different phenological conditions have a greater impact on the results.Negative samples with the phenological changes cause the accuracy of the extraction results to decrease;while atmospheric conditions such as haze can be increased to some extent;The atmospheric conditions such as smog can increase the robustness of the sample to a certain extent,which makes the sample migration have strong operability.At the same time,combined with phenological and imaging weather conditions,a highly applicable set of labeled samples can be obtained.In addition,the practical application of the method of convolution network around the Central Line Project of South-to-North Water Diversion indicates that the method has strong generalization.