Soil Mapping Research Based On Multi-source Remote Sensing Data And Random Forest Algorithm

Soil spatial distribution information plays an important role in agricultural production and resource utilization.The method based on traditional soil survey is mainly used by soil experts to obtain soil-landscape models based on empirical knowledge and field surveys,combined with other data such as terrain data,satellite imagery,and aerial interpretation techniques to determine the spatial distribution of different types of soil.And the soil map was drawn by hand.Its process is time-consuming and laborious,and the mapping accuracy is low.With the increasing demand for modern agricultural production and development,traditional soil survey methods are difficult to meet the current development requirements of agricultural refinement.In recent years,with the continuous development of 3S technology and data mining algorithms,digital soil mapping has become an emerging soil survey technology.At present,in digital soil mapping research,how to choose environmental co-variables that reflect soil changes is a key issue.Spectral information,vegetation index,texture features and other data derived from remote sensing image data are gradually used as new variables in soil mapping due to their fast imaging,easy to obtain,high resolution,and rich information of underlying data.This paper takes the Huangshui River Basin of Huajiahe Town,Hong'an County,Hubei Province as the research area,and uses traditional soil maps as data sources,which uses remote sensing image data(landsat-8 OLI,sentinel-2A/B,GF-2)to Update soil type map by implementing random forest algorithms.Specific research includes inference mapping based on multi-resolution remote sensing data and multi-time series remote sensing data:(1)Soil inferred map based on multi-resolution remote sensing data.In this part of the study,a single-temporal image of Landsat-8(30 meters),Sentinel 2A/B(10 meters),and High Score 2(2 meters)was selected as the remote sensing data source.The remote sensing variables such as the first principal component,spectral index,and texture features were extracted separately,which combined with Parent matter and topographic factors to make up an environmental factor data set.In order to ensure the principle of unique experimental variables,no environmental factor screening is performed in this process.The random forest algorithm was used to predict and map soil types.The accuracy of the mapping results was evaluated through field sampling points.And the main environmental variables that affected the spatial distribution of soil types under different resolution remote sensing data were explored.From the mapping results and verification accuracy,it can be concluded that the overall mapping accuracy of the Landsat-8 OLI(30 meters)experimental group is 81.4%,the Kappa coefficient is 0.76.And the overall mapping accuracy of the Sentinel-2A/B(10 meters)experimental group is 81.8%.The Kappa coefficient is 0.77.The overall mapping accuracy of the GF-2(2meters)experimental group is 80.3%,and the Kappa coefficient is 0.75.The spatial distribution of soil in the three groups of experiments has the same overall trend.Among them,the spatial distribution of soil types in the Landsat-8 OLI experimental group isclosest to the traditional soil map,and the Sentinel-2A/B experimental group predicts the highest overall classification accuracy.The spatial resolution of remote sensing images has affected the consistency of soil spatial distribution of predicted soil maps and traditional soil maps and the degree of patch fragmentation of predicted soil maps to a certain extent.The higher the spatial resolution,the more broken the soil type distribution map,and the lower the spatial consistency with the traditional soil map.And the higher the spatial resolution,the greater the difference in area of soil types developed under the same parent material,which will lead to a reduction in classification accuracy to a certain extent.(2)Soil inferred map of remote sensing data based on annual time series.This part of the study selects the image with the highest mapping accuracy(Sentinel 2A/B)as the remote sensing data source,and introduces remote sensing factors based on annual time series as new auxiliary variables,which combined with environmental factors and terrain factors to make up an environmental factor data set.The principal components analysis,box plot,and mean processing were used to screen the remote sensing factors,and the classification features were extracted.The random forest algorithm was used to predict and map soil types,and the accuracy of the mapping results was evaluated through field sampling points.The effects of soil mapping using single-temporal remote sensing factors with topographic factors,multi-temporal remote sensing factors with topographic factors were compared.From the mapping results and verification accuracy,it can be concluded that the overall mapping accuracy of the multi-temporal remote sensing image experiment group is 83.3%,and the Kappa coefficient is 0.79.The overall spatial distribution of soil in the two groups of experiments is consistent.The spatial distribution of soil types in the multi-temporal remote sensing experiment group is closer to the traditional soil map,and the overall classification accuracy of the multi-temporal remote sensing image experiment group is better than that of the single-temporal remote sensing image experiment group.The classification accuracy of most soil types in the multi-temporal remote sensing image experiment group were improved compared with the single-temporal remote sensing image experiment group.The silt fields,silt soil and shallow moist silt fields in the predicted soil maps based on single-temporal remote sensing images are more fragmented than the predicted soil maps in multi-temporal images.The distribution of the predicted soil maps in multi-temporal images is more concentrated,and the area of the delineation of soil is increased,and the classification accuracy is also improved.This study discussed the impact of image resolution on soil mapping,and further explored the feasibility of using multi-temporal remote sensing factors as soil-environment co-variables.With the addition of remote sensing co-factor,the mapping accuracy of soil map is improved,and the implied soil-environment knowledge can be expressed in more detail.It provides theoretical support for expanding the soil-landscape knowledge acquisition of current remote sensing data.