Empirical Model Selection And Feature Extraction For Heavy Metal Concentration In Soil Using Multi-temporal Landsat-8 Images

Soil is a precious natural source which is closely related to environmental quality and human health,so monitoring soil quality by soil heavy metal concentration retrieval is of great importance.Traditional approach relies on field grid sampling.Applying remote sensing technique to soil quality retrieval benefit soil monitoring with low cost and high efficiency.This study attempts to enhance the empirical model generalization capability via data preparation,model selection and feature extraction,aiming at improving the precision of retrieving soil heavy metal concentration(HMC)using multi-spectral remote sensing images.(i)In terms of data selection,this study takes the advantage of easy data availability,short revisit period and wide range coverage of Landsat-8 images to retrieve soil concentration of three heavy metal,i.e.,copper(Cu),arsenic(As),mercury(Hg).To overcome the shortage of low spectral resolution of Landsat-8,this study adopts multi-temporal Landsat-8 images to enlarge the spectral information.In this manner,the weak signal,i.e.,HMC in soil can be better captured by enriched spectrum for higher retrieval precision of HMC.Comparative experiments on a certain study area demonstrates that the model precision achieved by time-series is much higher than that achieved by single date image,which indicating that time-series images could take full of informative spectrum to effectively reflect spectral response of the HMC variation.(ii)For model selection,a systematic model unbiased evaluation approach is applied to select the best model among partial least square regression,artificial neural network and support vector regression which are widely utilized methods.Model selection turns out to be a necessary process for HMC retrieval and PLSR performs the best on all three HMC precision.(iii)To choose the optimal feature combination,this study evaluates the importance of original bands of time-series images,principal components and minimum noise fraction features for HMC concentration.Experiments shows that the selected features combination performs better than original bands on all three HMC retrieval,which indicating the extracted features are able to more effectively extract the spectral response of HMC variation in soil.This study applies the combination of best data,model and features to HMC prediction of the entire study area and then HMC thematic maps are obtained.To evaluate the quality of the obtained map,we compared them to the related materials like classification map of that area,which demonstrate the reliability of the precision maps.