Hyperspectral And Simulated Multispectral Quantitative Inversion Modeling Of Soil Carbon Fraction In Cotton Fields Of Alar Reclamation Area

As an important component of the soil carbon pool,changes in the content of soil organic carbon play a key role in the global carbon cycle and climate change.At the meantime,SOC,as a key indicator of soil quality,also has a controlling effect on farm productivity and soil ecological service function through its dynamic changes.Compared with SOC,human understanding of soil inorganic carbon（SIC）and its impact on climate change is still at a preliminary stage,and it is believed that SIC is very stable and plays a limited role in the global carbon cycle.In recent years,the SIC has become increasingly unstable as anthropogenic disturbances（land use change,chemical fertilizer application,agricultural irrigation,etc.）and global climate change（increased rainfall and temperature,extreme rainfall events,etc.）have intensified.Traditional chemical measurement methods are cost efficient and difficult to meet realistic needs.The development of remote sensing technology has realized the transformation of soil property monitoring from static to dynamic.Efficient and accurate monitoring of SOC and SIC contents is of practical significance for soil carbon pool research and achieving sustainable agricultural development.In this study,198 soil samples were collected from 0～100 cm depth in the cotton field of Alar Reclamation Area,Xinjiang,and the soil samples were divided into five 20 cm levels.The feasibility of predicting SOC and SIC contents of different depth profiles and different soil layers based on surface spectra was explored;at the same time,Landsat 8 satellite multispectral data were simulated to construct a model for SOC and SIC prediction based on simulated multispectral.The study found that:（1）The trends of soil reflectance spectra in the study area were basically similar,showing a trend of decreasing spectral reflectance with increasing SOC content and increasing spectral reflectance with increasing SIC content over the entire wavelength range.The six modeling methods performed differently on SOC and SIC data,and the overall prediction effects were LSTM>1D-CNN>RF>PLSR>2D-CNN>DBN in order.The LSTM model has the highest overall accuracy,and it is a relatively new model in soil spectral analysis that provides good predictions of both SOC and SIC contents(for the SOC validation set R²=0.83,RMSE=0.85 g kg^-1,RPD=2.35,RPIQ=3.18;for the SIC validation set R²=0.85,RMSE=1.82 g kg^-1,RPD=2.57,RPIQ=3.47).（2）The six algorithms SA,ACO,IRIV,PSO,IRF and CARS feature band selection algorithms can all effectively select the sensitive bands of SOC and SIC.The number of bands screened by all algorithm accounts for less than 5%of the total number of bands,which can significantly reduce the computational volume of the model and improve the computing efficiency.Where the feature bands of SOC are near 900 nm,1400 nm,1900 nm,2200 nm,and 2350 nm.The six algorithms select the same SIC bands near 570 nm、645nm、805 nm、1420 nm、2348 nm and 2400 nm,which may be the characteristic bands of SIC.The models built using the feature-selected bands were all able to achieve quantitative predictions of SOC and SIC（model validation set RPD>2.0）.For the prediction of SOC and SIC,the models built with the IRIV selected bands had the highest accuracy,with model validation sets R² of 0.83 and 0.87,RMSE of 0.78 g kg^-1 and 1.77 g kg^-1,RPD of 2.19 and 2.65,and RPIQ of 3.08 and 3.58,respectively,and the models were able to predict SOC and SIC contents well.Collectively,the IRIV algorithm was found to be the optimal method for SOC and SIC feature band selection.（3）In the prediction of SOC and SIC based on surface spectra for 0～40 cm,0～60 cm,0～80 cm and 0～100 cm depth profiles,the accuracy of both PLSR and LSTM models in predicting SOC and SIC showed a gradual decrease with the increase of soil profile depth.The prediction of SOC and SIC contents in different depth profiles can be achieved by using the surface soil sample spectra,but the prediction accuracy is not high.the LSTM model can better predict SOC in profiles above 0～60 cm and SIC contents in profiles above 0～80 cm.In the prediction of SOC and SIC of 20～40 cm,40～60 cm,60～80 cm and 80～100 cm soil layers based on the surface spectrum,the correlation between SOC and SIC of surface soil samples and SOC and SIC of the remaining soil layers,respectively,reached a highly significant level.The LSTM model can quantify the SOC of 20～40 cm soil layer and SIC of 40～60 cm soil layer The LSTM model can predict the SOC and SIC of 20～40 cm soil layer and 40～60 cm soil layer with high accuracy.（4）SOC and SIC have characteristic responses in the Coastal（430～450 nm）,Blue（450～510nm）,Green（530～590 nm）,Red（640～670 nm）,and NIR（850～880 nm）bands of simulated Landsat8 OLI,and the correlations all reach 0.01 significant level.The prediction effects of the models established by the five methods were LSTM>1D-CNN>RF>PLSR>MLR in order.the SOC and SIC prediction models based on the LSTM algorithm had the highest prediction performance with the validation set R² of 0.77 and 0.82,RMSE of 0.97 g kg^-1 and 2.16 g kg^-1,RPD of 2.05 and2.17,and RPIQ of 2.76 and 2.93,respectively.This study provides new ideas and methods for the monitoring of SOC and SIC.