Research On Remote Sensing Method Of Saline Land Information Extraction Based On Deep Learning

Soil salinization has become a major soil environmental problem,which not only causes the reduction of food production but also causes serious damage to the ecological environment,so it is very necessary to treat and restore saline lands.However,the restoration process of saline land is complicated and it is very easy to repeat salinization of saline land after restoration.This is because saline land is mostly distributed in arid and semi-arid regions,where the ecological environment is more fragile and inappropriate treatment means can lead to negative ecological effects.Therefore,it is especially important to use scientific management methods when implementing saline land management,and scientific management cannot be achieved without accurate and efficient access to saline land information.Since the use of remote sensing means to study saline land has the advantages of non-contact,wide range and block imaging speed,so how to use remote sensing means for efficient and accurate saline land information acquisition has been a hot spot of research.The deep learning method,as a powerful information processing tool,has proved to have good performance in the fields of remote sensing image classification and waveband information processing,and has become one of the important methods for remote sensing data processing.However,among the previous studies,the use of remote sensing methods based on deep learning for saline information processing is still little involved.Therefore,this study will be carried out at both macroscopic and microscopic scales to enrich the related research,which is also the innovation point of this paper.In the macroscopic aspect,this study proposes a deep neural network based on salinity indices using OLI sensor data from Landsat-8 with a typical saline area,Zhenlai County,western Jilin,as the study area,with the aim of extracting saline information with high automation and low human involvement.In addition,different salinity indices(SI,SI1 and SI2)and combinations were incorporated in this study,and several common accuracy evaluation metrics(Precision,Io U,F1-score and Recall)were used to evaluate the effects of different salinity indices and combinations on the accuracy of saline information extraction.Finally,this study also applied the network with the highest accuracy to the extraction task of saline land in Zhenlai region from 2016 to 2020,and obtained the spatial distribution characteristics of saline land in the region as well as the characteristics of saline land distribution over time.In the microscopic aspect,a Bidirectional Polarized Distribution Function(BPDF)model based on deep neural network was proposed in this study,and it was compared with several existing semi-empirical BPDF models and machine learning-based BPDF models based on the measurement results of polarization spectra in the laboratory,in order to find the most suitable BPDF model for saline soil.The following conclusions were obtained from this study:(1)A deep neural network based on a two-level nested U-shaped network(U~2-Net)is effective for extracting saline sites from medium and high-resolution remote sensing images.This deep neural network adopts a two-level nested U-shaped structure,and each of its internal parts is populated by a residual U module(Re Sidual U,RSU)module.The network is well implemented to combine global features and multi-scale features,which in turn allows more contextual information to be obtained.In the saline extraction results,the overall accuracy of the model reached 93.58%,which is already greater than 80%and can be considered as achieving good results.(2)The neural network model based on salinity indices has also achieved good results in saline extraction.No matter which salinity index or its combination was added,the extraction accuracy obtained reached more than 0.8.The inclusion of the salinity index as a global feature in the input section affected the accuracy of image classification.In addition,the classification accuracy improved with the addition of specific salinity indices compared with the training results of the multispectral group without the addition of salinity indices,with most accuracy metrics improving by about 1%.(3)The superimposition of salinity indices does not necessarily lead to higher accuracy.Overlaying more salinity indices allows more information to be input for training and provides more global features to the input data,which would theoretically produce better results.However,the results of this study do not support this hypothesis.On the contrary,overlaying too many salinity indices reduces the accuracy of classification.The possible reason is that adding too many salinity indices creates information redundancy and reduces the extraction accuracy of the network.When three salinity indices are added,all accuracy indicators of the extraction results are lower than those of the multi-wavelength group without salinity indices.(4)In this study,a deep neural network based on salinity indices was used to extract saline land in Zhenlai region from 2016 to 2020 to verify the usability of the network.And to make a generalization of the spatial and temporal distribution characteristics of saline land in the region,in addition to an attempt to analyze the reasons for the formation of this spatial and temporal characteristic.Specifically,the saline land in Zhenlai is mainly concentrated in the central and western regions,and the distribution of saline land has decreased over time.This is mainly due to long-term saline land management activities,including various ecological management projects(especially water conservation projects),development and screening of saline-tolerant crops,etc.(5)Several semi-empirical BPDF models and machine learning-based BPDF models were applied to saline soils,and the applicability was analyzed by the polarized reflectance data of saline soils measured in the laboratory.The results showed that several BPDF models showed adaptability for saline soils,with the best results being the newly proposed deep neural network-based BPDF model,and the evaluation results of several evaluation metrics(RMSE,Cor and R~2)were at the top of several BPDF models.