Retrieval of the Soil Salinity From Sentinel-1 Dual-Polarized SAR Data Based on Deep Neural Network Regression
| Autor: | Dehai Zhu, Chao Zhang, Ruizhi Sun, Chen Qiqi, Li Li, Qianqian Zhang |
|---|---|
| Rok vydání: | 2022 |
| Předmět: |
Soil salinity
Mean squared error 0211 other engineering and technologies Regression analysis 02 engineering and technology Geotechnical Engineering and Engineering Geology Regression Random forest Support vector machine Multilayer perceptron Statistics Feature (machine learning) Electrical and Electronic Engineering 021101 geological & geomatics engineering Mathematics |
| Zdroj: | IEEE Geoscience and Remote Sensing Letters. 19:1-5 |
| ISSN: | 1558-0571 1545-598X |
| Popis: | Soil salinity is a significant environmental indicator, directly reflecting the land quality and its productivity. In this work, dual-polarized synthetic aperture radar (PolSAR) data were exploited to retrieve the soil salinity in saline-affected fields based on the deep neural network (DNN) theory. Two polarimetric features gotten from Cloude polarization decomposition, together with six other features, were calculated and selected from ten features according to the feature importance. The DNN regression model for soil salinity (DNNR-S) based on one-dimensional convolutional neural network, quantifying the relationship between the eight feature parameters and soil salinity, was established to retrieve the soil salinity in northeastern China. Compared with other five traditional machine learning regression methods [logistic regression (LOR), random forest regression (RFR), support vector regression (SVR), decision tree regression (DTR), and multilayer perceptron regression (MPR)], DNNR-S is the best-performing model with the performance of root mean square error (RMSE) = 0.28, mean absolute error (MAE) = 0.21, and r = 0.74. Using the proposed model, the soil salinity was retrieved, and 52.37% was found to be affected by salinization which was in general agreement with the government report. Due to the training criteria in the proposed model, it was a little more computationally intensive than all the others, but the DNNR-S had the advantage of easy spread to other study areas. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|