Research On Soil Moisture Inversion Based On Neural Network And Machine Learning

Soil moisture,also known as soil moisture,is a key indicator of moisture exchange between the atmosphere and the Earth's surface.It directly controls the climate change and water cycle of the Earth's ecological environment.According to the soil moisture index,climate and environmental changes can be efficiently monitored.This is of great significance for monitoring agro-forestry droughts and floods,regional climate change,and surface plant evapotranspiration.In order to meet the needs of these practical ap-plications,soil moisture products should have the following characteristics:(1)High time and spatial resolution.(2)It can cover large areas.However,traditional observa-tion methods cannot meet the needs of large-scale detection,and remote sensing tech-nology cannot obtain soil moisture data with high spatial resolution.In view of this situation,using the latest remote sensing soil moisture data and Earth observation data to invert and downscale the soil moisture data,improving the spatial resolution of soil moisture data is a difficult point that needs to be studied at present.In this study,the most recent visible image of the near-infrared spectrum of the Tiangong-2 Wide-band Imaging Spectrometer(WIS)and the SMAP(Soil Moisture Ac-tive and Passive)soil moisture data were selected as the down-scale inversion data source.The relationship between spectral information and soil moisture data was estab-lished by Bayesian neural network algorithm improved by GA(Genetic Algorithm)ge-netic algorithm and random forest algorithm improved by GA genetic algorithm respec-tively.The SMAP soil moisture data is down-scaled and the spatial resolution is in-creased from 3km to 100m.The Bayesian neural network model with different number of hidden layer nodes and the random forest model with different decision tree numbers are compared and analyzed.The model precision and fitting effect are deeply studied,and the algorithm complexity is given to explore the applicability of neural network and machine learning algorithms in the inversion of remote sensing data.Finally,the correlation between spectral reflectance of each channel and soil moisture is analyzed.The main research contents and conclusions are as follows:(1)Data acquisition and preprocessing.Atmospheric correction of remote sensing images,and then using the spectral index based cloud and shadow detection algorithm to mask the thick clouds in the Tiangong-2 image.The thick cloud in the image is iden-tified and extracted and rejected so that it does not participate in the calculation of the sample training.Three parameters of longitude,latitude and soil moisture were extracted from the SMAP/Sentinel-1 L2 soil moisture data and matched with the se-lected Tiangong-2 image.(2)The GA algorithm is used to improve the Bayesian neural network and random forest algorithms,and the improved algorithm is used to down-scale the soil moisture.Based on the MATLAB 2018A neural network toolbox and GUI page,the reflectivity data of 14 channels collected by the Tiangong-2 is used as input data,and the SMAP/Sentinel-1 L2 soil moisture data is used as output data for learning and training.(3)The results show that the spatial resolution of SMAP soil moisture data is im-proved from 3km to 100m with the inversion of neural network and improved machine learning algorithm.When using the GA-modified Bayesian neural network to invert,the training effect is best when the number of hidden layer nodes is 24,R~2 is 0.755,and the root mean square error RMSE is 0.161.When using the GA-modified random forest machine learning algorithm to invert,the best effect is when the number of decision trees is 60,R~2 is 0.809,and the root mean square error RMSE is 0.120.When dealing with big data samples,the improved random forest algorithm based on GA has lower time complexity than the improved Bayesian neural network algorithm based on GA.This study found that the spectral reflectance of 8,9 and 10 channels of the Tiangong-2 WIS has a stronger correlation with soil moisture.When the SMAP soil moisture data is down-scaled,the random forest model has higher precision and better fitting effect than the Bayesian neural network model.A more accurate large-scale soil moisture re-duction scale inversion can be achieved.