Soil Moisture Retrieval Using AMSR-E Data By BP Neural Network For Sichuan Middle Hilly Area

Soil moisture is an important indicator of the parameters in the hydrology, meteorology, ecology, and agricultural scientific research fields. As one of the most important indicator parameters, soil moisture is the key factor of environmental processes in agricultural drought quantitative description, and accurately describing the soil moisture dynamic change is crucial for agricultural drought monitoring. Therefore, it is significant to monitor soil wetness status and spatial distribution pattern. Traditional soil moisture prediction models, such as the models based on soil moisture dynamics, empirical formula, water balance, have their advantages, but their shortcoming is also obvious for needing to measure many factors and process large amounts of inadequate data. In contrast, artificial neural network, such as BPNN (Back Propagation Neural Network), soil moisture prediction model has the ability to handle multi-factor and non-linear problems. And microwave remote sensing data, such as AMSR-E remote sensing data, has the advantage of penetrating atmosphere, clouds, fog, etc. Thus, in this paper, based on AMSR-E and BPNN we have built some models to retrieval soil moisture.The data used for this research are the observed soil moisture and AMSR-E soil moisture data from2006to2010in Sichuan Middle Hilly Area. Using the data we have built a number of different back BP neural network models by the way of adoption two preferred factors, three preferred factors and four preferred factors as inputs respectively. And then combining with precipitation data, evaporation data and the index proposed by Ma Zhuguo, we have selected two four-factor models (BPNN1model and BPNN2model) to validate, evaluate and analyze the simulation results. And the main conclusions are as follows:1. The accuracy of more-factor neural network models is generally higher than less-factor ones, but sometimes interference factors may reduce the accuracy of the models, The results show that Tb36.5GHz frequency band is interference factor as the input for soil moisture retrieval by BP neural network models, while6.9GHz and10.7GHz frequency band used in the models can help to reduce the simulation value dispersion and hence to improve the correlation between the simulated values and the observed values.2. The MSR-E soil moisture product is more sensitive to the precipitation dynamic change than the observed soil moisture and the simulated soil moisture data by BPNN1 model and BPNN2model for their certain lag to the time of less cumulative rainfall.3. The spatial distribution overall trend in magnitude of AMSR-E soil moisture is not very consistent with that of observed soil moisture. And the simulated soil moisture by BPNN1model and BPNN2model are more agreement with the observed than AMSR-E soil moisture product in magnitude and spatial distribution. From the comparison between BPNN1model and BPNN2model, it can be concluded that the trend of time series curve is consistent with the simulated values by BPNN1and BPNN2model, while the BPNN2simulation values have a greater fluctuation than BPNN1in the temporal dynamics. In terms of spatial distribution, the BPNN2simulated values always ignore the smaller and larger values, but the spatial distribution of magnitude of the model simulated value is more agreement with that of the observed soil moisture for the major drought events.4. When adding18.7GHz and36.5GHz frequency band as input factors, the BP neural network model can get some smaller simulated values and a smoother trend of soil moisture spatial distribution. The result indicates that the spatial distribution of arid areas reflected by BPNN2model’s simulated soil moisture spatial distribution is more close to that reflected by index proposed by Ma Zhuguo than that by BPNN1model.5. The trend of soil moisture spatial distribution by BP neural network model, which is lower in west and gradually increases in south and southeast, is just opposite to varied topography of the study area. While AMSR-E soil moisture distribution trend actions accord with the topography. That is to say, topographic factor has a certain influence on BP neural network model for soil moisture retrieval.