Retrieval Of Soil Organic Matter Along Rivers Based On Machine Learning

Soil is an important environmental resource,and organic matter is an important standard to measure soil fertility.Therefore,it is of great significance to predict the content of soil organic matter.Compared with the traditional grid sampling method,remote sensing technology has the characteristics of low cost,high efficiency and high prediction accuracy.In this paper,the soil organic matter near the Le’an River and its runoff in Jiangxi Province is taken as the research object,and the spectral reflectance and vegetation index of landsat5tm single image are taken as the input characteristics to establish SVR and BP models optimized by genetic algorithm and SVR and BP models optimized by improved fish swarm algorithm;The terrain factor is extracted as an additional input variable through GIS technology to explore its impact on the accuracy of the model;On the basis of the above,multi temporal Landsat images are used to increase the response ability of spectral information to soil organic matter content,and the feature bands are selected by feature extraction method to establish the best inversion model;Finally,a new water body index is proposed to extract the water body in the study area,and the organic matter content of 150m buffer zone along the river is mapped by combining with the best inversion model:(1)Through correlation analysis,B1,B2,B3 and NDVI of Landsat5TM single image are selected as input variables,and six models of SVR,GA-SVR,ADAFSA-SVR,BP,GA-BP and ADAFSA-BP are established.Through comparative analysis,the best model of ADAFSA-BP is R2=0.361,RMSE=0.626,MAE=0.593.(2)Through correlation analysis,terrain factors,such as elevation,slope,plane curvature and section curvature,are added to the input variables,so that the training effect and prediction effect of the six inversion models are significantly improved.From the relative error of the predicted value of the test set,the terrain factor can significantly improve the prediction accuracy of low soil organic matter content.The optimal models of two machine learning methods are ADAFSA-SVR,R2=0.424,RMSE=0.594,MAE=0.52;ADAFSA-BP,R2=0.519,RMSE=0.543,MAE=0.504.Their corresponding RPD were 1.35 and 1.44,respectively.ADAFSA-BP model has certain reliability.(3)Compared with GA,ADAFSA needs fewer iterations in SVR parameter finding;while the former has more iterations in BP,it can jump out of the local optimal solution and find a more suitable solution.The learning speed of SVR is much faster than that of BP,and intelligent algorithm can improve the speed of grid search method commonly used in SVR.In the aspect of small sample learning ability,the best small sample learning model is ADAFSA-BP,but SVR model has the least impact on the prediction accuracy caused by the reduction of training set.(4)By superposition of landsat5tm images in the same region at different times,the multi-phase images similar to hyperspectral images are constructed.The features extraction methods of PCA,ICA and MNF are used to establish adafsa SVR,adafsa BP and elm models.The elm model with the first 10 most variance characteristic bands and terrain factors extracted by MNF method is the best.The training set r2=0.772,rmse=0.375,mae=0.297,the test set r2=0.602,rmse=0.521,mae=0.463,the relative analysis error RPD is 1.68.By introducing the multi-temporal image band information and terrain factors,the elm model is established,and a more reliable elm model is obtained The inversion model of organic matter mapping in the study area.The new water index icwi is used to solve the influence of NDWI on the extraction of water body by the shadow of small water and mountain body in the research area.The mapping of 150m along the river coast of the research area is carried out,but the mapping effect is not good.Figure[24]Table[20]Reference[81]...