# Dual-Channel Convolutional Neural Network for Bare Surface Soil Moisture Inversion Based on Polarimetric Scattering Models

^{1}

^{2}

^{3}

^{*}

## Abstract

**:**

## 1. Introduction

- Based on two scattering models, namely the X-Bragg model and the IEM, we make full use of the amplitude and phase information of PolSAR data, which can expand the applicable range of roughness and improve the inversion accuracy.
- The dual-channel convolutional neural network, which makes full use of spatial information and adds a dropout layer to reduce the overfitting, is used to perform the feature fusion of the parameters extracted from the scattering model.
- We use the dual-channel convolutional neural network to design a framework, which can perform coarse-grained qualitative classification and fine-grained quantitative regression to suit different tasks.

## 2. Methods

#### 2.1. Extraction of H, A, and α by the X-Bragg Model

#### 2.2. Extraction of ${\sigma}_{hh}^{0}$,${\sigma}_{vv}^{0}$, and ${\sigma}_{hh}^{0}/{\sigma}_{vv}^{0}$ by the IEM

#### 2.3. Design of the Dual-Channel Convolutional Neural Network

#### 2.4. Data Preprocess and Training

## 3. Datasets

#### 3.1. Simulated Dataset

#### 3.2. Field Dataset

## 4. Experiments and Discussion

#### 4.1. Simulated Data

- Inversion accuracy (IA) and average IA$$\mathrm{IA}(\%)=\frac{{\displaystyle \sum _{i=1}^{K}}I\{{y}_{Classification}^{i}={y}_{real}^{i}\}}{K}\ast 100\%$$$$\mathrm{Average}\mathrm{IA}(\%)=\frac{{\displaystyle \sum _{i=1}^{N}}I\{{y}_{Classification}^{i}={y}_{real}^{i}\}}{N}\ast 100\%$$
- Root mean square error (RMSE)$$\mathrm{RMSE}=\sqrt{\frac{1}{N}\sum _{i=1}^{N}{({y}_{Regressiom}^{i}-{y}_{real}^{i})}^{2}}$$
- Coefficient of determination (${\mathrm{r}}^{2}$)$$\mathit{SST}=\sum _{i=1}^{N}{({y}_{real}^{i}-{\stackrel{\u203e}{y}}_{real}^{i})}^{2}$$$$\mathit{SSE}=\sum _{i=1}^{N}{({y}_{Regressiom}^{i}-{y}_{real}^{i})}^{2}$$

#### 4.2. Field Data

^{2}are given in Table 11. We can see from Table 11 that the dual-channel CNN has a lower RMSE and higher coefficient of determination than other methods, which means that the dual-channel CNN has higher inversion performance. The dual-channel CNN can make full use of spatial information and reduce mutual interference between features by using dual channels. Compared with the traditional neural network and convolutional neural network, the dual-channel CNN can reduce the RMSE by about 0.39–1.34% and increase r

^{2}by about 0.11–0.48. We verify the effectiveness of the proposed method by using L-band airborne data. It has weak penetration for C-band data. However, since this paper focuses on the bare soil region and there is no interference from vegetation, we consider that this method can also be applied to the C-band satellite data for soil moisture inversion of bare soil.

#### 4.3. Different Size Patches and Visualization of Feature Maps

## 5. Conclusions

^{2}and RMSE are 0.88 and 0.98% respectively. The above statistical results are superior to one-channel convolutional neural network and MLP, which indicates that the proposed method can be accurately applied to soil moisture inversion task in bare soil regions. The dual-channel CNN makes full use of spatial information and merges the information between different polarization parameters. A small number of samples can be used to get high inversion accuracy. Due to the physical models used in this paper are only applicable to bare soil and sparse vegetation cover areas, the effect of vegetation is not taken into account. In the future, we plan to combine the dual-channel convolutional neural network with the inversion model suitable for vegetation covered area for soil moisture inversion task in vegetation-covered areas.

## Author Contributions

## Funding

## Acknowledgments

## Conflicts of Interest

## References

- King, C.; Baghdadi, N.; Lecomte, V.; Cerdan, O. The application of remote-sensing data to monitoring and modelling of soil erosion. Catena
**2005**, 62, 79–93. [Google Scholar] [CrossRef] - Bronstert, A.; Creutzfeldt, B.; Gräff, T.; Hajnsek, I.; Heistermann, M.; Itzerott, S.; Jagdhuber, T.; Kneis, D.; Lück, E.; Reusser, D.E.; et al. Potentials and constraints of different types of soil moisture observations for flood simulations in headwater catchments. Nat. Hazards
**2012**, 60, 879–914. [Google Scholar] [CrossRef] - Berthet, L.; Andréassian, V.; Perrin, C.; Javelle, P. How crucial is it to account for the antecedent moisture conditions in flood forecasting? Comparison of event-based and continuous approaches on 178 catchments. Hydrol. Earth Syst. Sci.
**2009**, 13, 819–831. [Google Scholar] [CrossRef] [Green Version] - Brisco, B.; Pultz, T.J.; Brown, R.J.; Topp, G.C.; Hares, M.A.; Zebchuk, W.D. Soil moisture measurement using portable dielectric probes and time domain reflectometry. Water Resour. Res.
**1992**, 28, 1339–1346. [Google Scholar] - El Hajj, M.; Baghdadi, N.; Zribi, M.; Bazzi, H. Synergic use of Sentinel-1 and Sentinel-2 images for operational soil moisture mapping at high spatial resolution over agricultural areas. Remote Sens.
**2017**, 9, 1292. [Google Scholar] [CrossRef] [Green Version] - He, L.; Hong, Y.; Wu, X.; Ye, N.; Walker, J.P.; Chen, X. Investigation of SMAP active–passive downscaling algorithms using combined sentinel-1 SAR and SMAP radiometer data. IEEE Trans. Geosci. Remote Sens.
**2018**, 56, 4906–4918. [Google Scholar] [CrossRef] - He, L.; Panciera, R.; Tanase, M.A.; Walker, J.P.; Qin, Q. Soil moisture retrieval in agricultural fields using adaptive model-based polarimetric decomposition of SAR data. IEEE Trans. Geosci. Remote Sens.
**2016**, 54, 4445–4460. [Google Scholar] [CrossRef] - Panciera, R.; Tanase, M.A.; Lowell, K.; Walker, J.P. Evaluation of IEM, Dubois, and Oh radar backscatter models using airborne L-band SAR. IEEE Trans. Geosci. Remote Sens.
**2013**, 52, 4966–4979. [Google Scholar] [CrossRef] - Hoshikawa, K.; Nagano, T.; Kotera, A.; Watanabe, K.; Fujihara, Y.; Kozan, O. Classification of crop fields in northeast Thailand based on hydrological characteristics detected by L-band SAR backscatter data. Remote Sens. Lett.
**2014**, 5, 323–331. [Google Scholar] [CrossRef] - Bertuzzi, P.; Chaànzy, A.; Vidal-Madjar, D.; Autret, M. The use of a microwave backscatter model for retrieving soil moisture over bare soil. Int. J. Remote Sens.
**1992**, 13, 2653–2668. [Google Scholar] [CrossRef] - Rice, S.O. Reflection of electromagnetic waves from slightly rough surfaces. Commun. Pure Appl. Math.
**1951**, 4, 351–378. [Google Scholar] [CrossRef] - Tsang, L.; Kong, A.J.; Ding, K.H. Scattering of Electromagnetic Waves: Theories and Applications; John Wiley & Sons: Hoboken, NJ, USA, 2004. [Google Scholar]
- Fung, A.K.; Li, Z.; Chen, K.S. Backscattering from a randomly rough dielectric surface. IEEE Trans. Geosci. Remote Sens.
**1992**, 30, 356–369. [Google Scholar] [CrossRef] - Fung, A.K. Microwave Scattering and Emission Models and Their Applications; Artech House: Norwood, MA, USA, 1994. [Google Scholar]
- Dobson, M.C.; Ulaby, F.T.; Hallikainen, M.T.; El-Rayes, M.A. Microwave dielectric behavior of wet soil-Part II: Dielectric mixing models. IEEE Trans. Geosci. Remote. Sens.
**1985**, 1, 35–46. [Google Scholar] [CrossRef] - Jung, S.G.; Hong, Y.; Oh, Y. Verification of Surface Scattering Models and Inversion Algorithms with the Polarimetric Backscatter Measurements of a Bare Soil Surface. In Proceedings of the Asia-Pacific Microwave Conference, Bangkok, Thailand, 11–14 December 2007; pp. 1–3. [Google Scholar]
- Qiu, C.; Chen, Y.; Tong, L.; Jia, M.; Pang, S. The method for soil moisture inversion based on ground-based scattering measurement. In Proceedings of the 2011 IEEE International Geoscience and Remote Sensing Symposium, Vancouver, BC, Canada, 24–29 July 2011; pp. 3086–3088. [Google Scholar]
- Oh, Y.; Sarabandi, K.; Ulaby, F.T. An empirical model and an inversion technique for radar scattering from bare soil surfaces. IEEE Trans. Geosci. Remote Sens.
**1992**, 30, 370–381. [Google Scholar] [CrossRef] - Dubois, P.C.; Van Zyl, J.; Engman, T. Measuring soil moisture with imaging radars. IEEE Trans. Geosci. Remote Sens.
**1995**, 33, 915–926. [Google Scholar] [CrossRef] [Green Version] - Merzouki, A.; McNairn, H.; Pacheco, A. Evaluation of the Dubois, Oh, and IEM radar backscatter models over agricultural fields using C-band RADARSAT-2 SAR image data. Can. J. Remote Sens.
**2010**, 36 (Suppl. 2), S274–S286. [Google Scholar] [CrossRef] - Sahebi, M.R.; Angles, J. An inversion method based on multi-angular approaches for estimating bare soil surface parameters from RADARSAT-1. Hydrol. Earth Syst. Sci.
**2010**, 14, 2355–2366. [Google Scholar] [CrossRef] [Green Version] - Verhoest, N.E.C.; Lievens, H.; Wagner, W.; Álvarez-Mozos, J.; Moran, M.S.; Mattia, F. On the soil roughness parameterization problem in soil moisture retrieval of bare surfaces from synthetic aperture radar. Sensors
**2008**, 8, 4213–4248. [Google Scholar] [CrossRef] [Green Version] - Hajnsek, I.; Pottier, E.; Cloude, S.R. Inversion of surface parameters from polarimetric SAR. IEEE Trans. Geosci. Remote Sens.
**2003**, 41, 727–744. [Google Scholar] [CrossRef] - Mirsoleimani, H.R.; Sahebi, M.R.; Baghdadi, N.; El Hajj, M. Bare soil surface moisture retrieval from sentinel-1 SAR data based on the calibrated IEM and dubois models using neural networks. Sensors
**2019**, 19, 3209. [Google Scholar] [CrossRef] [Green Version] - Baghdadi, N.; Cresson, R.; Hajj, M.E.; Ludwig, R.; La Jeunesse, I.; Baghdadi, N.; Cresson, R.; El Hajj, M.; Ludwig, R.; La Jeunesse, I. Estimation of soil parameters over bare agriculture areas from C-band polarimetric SAR data using neural networks. Hydrol. Earth Syst. Sci.
**2012**, 16, 1607–1621. [Google Scholar] [CrossRef] [Green Version] - Said, S.; Kothyari, U.C.; Arora, M.K. ANN-based soil moisture retrieval over bare and vegetated areas using ERS-2 SAR data. J. Hydrol. Eng.
**2008**, 13, 461–475. [Google Scholar] [CrossRef] - Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. Adv. Neural Inf. Process. Syst.
**2012**, 25, 1097–1105. [Google Scholar] [CrossRef] - Botalb, A.; Moinuddin, M.; Al-Saggaf, U.M.; Ali, S.S. Contrasting convolutional neural network (CNN) with multi-layer perceptron (MLP) for big data analysis. In Proceedings of the 2018 International Conference on Intelligent and Advanced System (ICIAS), Kuala Lumpur, Malaysia, 13–14 August 2018; pp. 1–5. [Google Scholar]
- Shang, R.; He, J.; Wang, J.; Xu, K.; Jiao, L.; Stolkin, R. Dense connection and depthwise separable convolution based CNN for polarimetric SAR image classification. Knowl.-Based Syst.
**2020**, 194, 105542. [Google Scholar] [CrossRef] - Song, T.; Kuang, L.; Han, L.; Wang, Y.; Liu, Q.H. Inversion of rough surface parameters from SAR images using simulation-trained convolutional neural networks. IEEE Geosci. Remote Sens. Lett.
**2018**, 15, 1130–1134. [Google Scholar] [CrossRef] - Allain, S.; Ferro-Famil, L.; Pottier, E. Two novel surface model based inversion algorithms using multi-frequency polSAR data. In Proceedings of the IGARSS 2004 IEEE International Geoscience and Remote Sensing Symposium, Anchorage, AK, USA, 20–24 September 2004; Volume 2, pp. 823–826. [Google Scholar]
- Lee, S.; Schuler, D.L.; Ainsworth, T.L. Polarimetric SAR data compensation for terrain azimuth slope variation. IEEE Trans. Geosci. Remote Sens.
**2000**, 38, 2153–2163. [Google Scholar] - Cloude, S.R.; Pottier, E. A review of target decom position theorems in radar polarimetry. IEEE Trans. Geosci. Remote Sens.
**1996**, 34, 498–518. [Google Scholar] [CrossRef] - Ioffe, S.; Szegedy, C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the International Conference on Machine Learning, Lille, France, 6–11 July 2015; pp. 448–456. [Google Scholar]
- Srivastava, N.; Hinton, G.; Krizhevsky, A.J.; Sutskever, I.; Salakhutdinov, R. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. J. Mach. Learn. Res.
**2014**, 15, 1929–1958. [Google Scholar] - Kingma, D.P.; Ba, J. Adam: A method for stochastic optimization. arXiv
**2014**, arXiv:1412.6980. [Google Scholar] - Hajnsek, I.; Bianchi, R.; Davidson, M.; D’Urso, G.; Gomez-Sanchez, J.; Hausold, A.; Horn, R.; Howse, J.; Low, A.; Lopez-Sances, J.M.; et al. AgriSAR 2006—Airborne SAR and optics campaigns for an improved monitoring of agricultural processes and practices. Geophys. Res. Abstr.
**2007**, 9, 04085. [Google Scholar] - Lee, J.S.; Jurkevich, L.; Dewaele, P.; Wambacq, P.; Oosterlinck, A. Speckle filtering of synthetic aperture radar images: A review. Remote Sens. Rev.
**1994**, 8, 313–340. [Google Scholar] [CrossRef] - Liang, J.; Zhou, J.; Qian, Y.; Wen, L.; Bai, X.; Gao, Y. On the sampling strategy for evaluation of spectral-spatial methods in hyperspectral image classification. IEEE Trans. Geosci. Remote Sens.
**2016**, 55, 862–880. [Google Scholar] [CrossRef] [Green Version]

**Figure 4.**Pauli images of the field dataset. (

**a**) E1, E2, E3, and E4 are dated 19 April 2006, and (

**b**) E5 is dated 20 April 2006.

**Figure 5.**The scattering entropy for different soil moisture values. (

**a**) 3% soil moisture; (

**b**) 18% soil moisture.

**Figure 6.**Simulated data inversion results of RMSE and r

^{2}by the dual-channel CNN in different levels of noise.

**Figure 7.**Classification results on the field dataset. (

**a**) Pauli RGB. (

**b**) Ground truth map of field data. (

**c**) Ground truth map of random sampling. (

**d**) Ground truth map of spatially disjoint sampling. (

**e**) Classification result map under random sampling. (

**f**) Classification result map under spatially disjoint sampling.

**Figure 9.**Regression results on the field dataset. (

**a**) Ground truth map of field data. Regression results of the (

**b**) dual-channel CNN, (

**c**) MLP, (

**d**) SVM, (

**e**) X-Bragg-CNN, (

**f**) IEM-CNN, and (

**g**) 6CH-CNN.

**Figure 10.**Visualization process of dual-channel CNN. (

**a**) Visual images of the X-Bragg-CNN input. (

**b**) Visual images of the IEM-CNN input. (

**c**) The visualized feature maps after the first convolution and ReLU operation in the X-Bragg-CNN. (

**d**) The visualized feature maps after the first convolution and ReLU operation in the IEM-CNN.

Traditional Model (Physical, Empirical, Semi-Empirical Models) | Neural Network (MLP) | Convolutional Neural Network (CNN) | |
---|---|---|---|

The modeling method | Analytical method or numerical method | Train data to fit model automatically | Train data to fit model automatically |

Data for modeling | Single pixel | One-dimensional vector of a target pixel | Local patch centered on the target pixel |

The scope of model application | Affected by the model itself and the calibration factors calculated in a specific area | Affected by training data | Affected by training data |

The number of model parameters | Small | Depend on the number of layers in the network | Depend on the number of layers in the network |

The robustness of the model | Normal | Normal | Strong |

Major source of errors | The fitting function is not exact. Soil surface modeling is not ideal. There is noise in the fitting data. | Overfitting due to limited training data | Overfitting due to limited training data |

The inversion results | Due to the model errors and the validity range of the model, a large number of pixels cannot be retrieved. For predicted pixels, the results are reliable. | The predicted soil moisture value can be obtained for all the input pixels. When the soil moisture of the input prediction data does not appear in the training, wrong results may occur. | The predicted soil moisture value can be obtained for all the input pixels. When the soil moisture of the input prediction data does not appear in the training, the wrong results may occur. |

Type | Dual-Channel CNN | |
---|---|---|

X-Bragg-CNN | IEM-CNN | |

Convolution layer | 8@3 × 3/1/0 | 8@3 × 3/1/0 |

Convolution layer | 16@3 × 3/1/0 | 16@3 × 3/1/0 |

Convolution layer | 24@3 × 3/1/0 | 24@3 × 3/1/0 |

Convolution layer | 32@3 × 3/1/0 | 32@3 × 3/1/0 |

Fully connected layer | 120 | 120 |

Merge layer | 240 | |

Fully connected layer | 84 | |

Fully connected layer | C (Class number) In this paper, C of the simulated dataset is 8, and the field dataset is 5 | |

Dropout (ratio:0.5) |

Type | Dual-Channel CNN | |
---|---|---|

X-Bragg-CNN | IEM-CNN | |

Convolution layer | 8@3 × 3/1/0 | 8@3 × 3/1/0 |

Convolution layer | 16@3 × 3/1/0 | 16@3 × 3/1/0 |

Convolution layer | 24@3 × 3/1/0 | 24@3 × 3/1/0 |

Convolution layer | 32@3 × 3/1/0 | 32@3 × 3/1/0 |

Fully connected layer | 120 | 120 |

Merge layer | 240 | |

Fully connected layer | 84 | |

Fully connected layer | 32 | |

Fully connected layer | 1 | |

Dropout (ratio:0.5) |

Filed ID | Date | Average Soil Moisture (%) |
---|---|---|

E1 | 19-Apr-06 | 20.1 |

E2 | 19-Apr-06 | 23 |

E3 | 19-Apr-06 | 24.3 |

E4 | 19-Apr-06 | 26.8 |

E5 | 20-Apr-06 | 13.8 |

ENL | Average IA | 3% | 8% | 13% | 18% | 23% | 28% | 33% | 38% |
---|---|---|---|---|---|---|---|---|---|

4.5-look | 97.96% | 100% | 99.80% | 98.87% | 98.74% | 99.54% | 97.63% | 95.22% | 93.89% |

4-look | 96.05% | 99.95% | 99.02% | 99.15% | 98.42% | 98.65% | 94.46% | 91.77% | 86.96% |

3-look | 92.56% | 99.81% | 98.68% | 98.27% | 97.02% | 93.68% | 87.20% | 88.39% | 77.42% |

2-look | 87.59% | 98.97% | 98.46% | 97.91% | 96.03% | 89.20% | 75.55% | 75.51% | 69.09% |

ENL | RMSE (%) | r^{2} |
---|---|---|

4.5-look | 0.65 | 0.99 |

4-look | 1.60 | 0.98 |

3-look | 1.89 | 0.97 |

2-look | 2.91 | 0.94 |

Field ID | The Number of Training Samples | The Number of Test Samples |
---|---|---|

E1 | 500 | 616,480 |

E2 | 500 | 120,282 |

E3 | 500 | 716,987 |

E4 | 500 | 344,954 |

E5 | 500 | 64,263 |

Field ID | Random | Disjoint |
---|---|---|

E1 IA | 97.61% | 97.28% |

E2 IA | 99.68% | 99.77% |

E3 IA | 95.81% | 95.50% |

E4 IA | 95.08% | 90.00% |

E5 IA | 98.18% | 97.52% |

Average IA | 96.59% | 95.39% |

Category | Dual-Channel CNN | MLP | SVM |
---|---|---|---|

E1 IA | 97.28% | 94.36% | 94.62% |

E2 IA | 99.77% | 97.96% | 97.92% |

E3 IA | 95.50% | 90.15% | 89.63% |

E4 IA | 90.00% | 79.74% | 82.05% |

E5 IA | 97.52% | 77.62% | 85.88% |

Average IA | 95.39% | 89.82% | 90.33% |

Testing time (s) | 50 | 20 | 30 |

Category | Dual-Channel CNN | X-Bragg-CNN | IEM-CNN | 6CH-CNN |
---|---|---|---|---|

E1 IA | 97.28% | 88.99% | 77.39% | 96.52% |

E2 IA | 99.77% | 97.10% | 98.93% | 99.09% |

E3 IA | 95.50% | 83.68% | 90.66% | 92.50% |

E4 AI | 90.00% | 89.86% | 73.06% | 86.06% |

E5 IA | 97.52% | 94.89% | 88.91% | 97.14% |

Average IA | 95.39% | 87.76% | 83.42% | 93.18% |

Testing time (s) | 50 | 31 | 31 | 35 |

**Table 11.**Field data RMSE and r

^{2}of the dual-channel CNN, MLP, SVM, X-Bragg-CNN, IEM-CNN, and 6CH-CNN.

Net | RMSE (%) | r^{2} |
---|---|---|

Dual-channel CNN | 0.98 | 0.88 |

MLP | 2.14 | 0.44 |

SVM | 2.02 | 0.50 |

X-Bragg-CNN | 1.82 | 0.59 |

IEM-CNN | 2.32 | 0.40 |

6CH-CNN | 1.37 | 0.77 |

Patch Size | Average IA | Testing Time (s) |
---|---|---|

7 $\times $ 7 | 93.16% | 45 |

11 $\times $ 11 | 95.39% | 50 |

15 $\times $ 15 | 95.65% | 65 |

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. |

© 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

## Share and Cite

**MDPI and ACS Style**

Yin, Q.; Li, J.; Ma, F.; Xiang, D.; Zhang, F.
Dual-Channel Convolutional Neural Network for Bare Surface Soil Moisture Inversion Based on Polarimetric Scattering Models. *Remote Sens.* **2021**, *13*, 4503.
https://doi.org/10.3390/rs13224503

**AMA Style**

Yin Q, Li J, Ma F, Xiang D, Zhang F.
Dual-Channel Convolutional Neural Network for Bare Surface Soil Moisture Inversion Based on Polarimetric Scattering Models. *Remote Sensing*. 2021; 13(22):4503.
https://doi.org/10.3390/rs13224503

**Chicago/Turabian Style**

Yin, Qiang, Junlang Li, Fei Ma, Deliang Xiang, and Fan Zhang.
2021. "Dual-Channel Convolutional Neural Network for Bare Surface Soil Moisture Inversion Based on Polarimetric Scattering Models" *Remote Sensing* 13, no. 22: 4503.
https://doi.org/10.3390/rs13224503