Estimation Of Soil Carbon Emission From Summer Maize Field Based On Ground Measurement And UAV Remote Sensing

At the farmland scale,the chamber method and eddy correlation measurement,as the traditional monitoring method of soil respiration（R_S）,is time-consuming and laborious,and high cost respectively.Remote sensing technology,as a new monitoring method,has been widely used in the estimation of R_S from different vegetation types because of its low cost and high efficiency.However,due to the lower resolution and long revisit period,satellite remote sensing is not suitable for R_S estimation at the farmland scale.In addition,when it comes to predict the seasonal R_S for different vegetation types and treatment methods,it is relatively difficult to select the optimal model to estimate R_S due to the lack of unified model description.In this study,an UAV multi-sources remote sensing system integrating visible light,thermal infrared and multispectral cameras was developed.Based on the ground data and UAV remote sensing image data obtained from the precision irrigation experimental station in Zhaojun town,Dalate Banner,Inner Mongolia,2019,the biotic factor（vegetation index,VI）affecting crop photosynthesis and abiotic factor(soil surface temperature,T_SF)affecting the mechanism of R_S were considered comprehensively.Nine different VIs were retrieved from the multispectral remote sensing and T_SF was obtained from the thermal infrared remote sensing,respectively.The seasonal R_S estimation model（VHI model）was proposed based on the Vant’Hoff index equation.This paper compared and analyzed the performance of the VHI model,the BP artificial neural network model（BPNN）and the multiple linear regression model（MNR）.The optimal model was selected to estimate the instantaneous emission of R_S,and the relationship model between the instantaneous and measured R_S was established.The spatial distribution maps of daily total emission flux of R_S at different growth periods were generated.The feasibility and applicability of R_S estimation by UAV remote sensing and ground monitoring were analyzed.The main contents and conclusions of this study are as follows:（1）An UAV multi-sources remote sensing system with visible light,thermal infrared and multispectral cameras was developed,and the relevant parameters of the system were tested.The results showed that the system was stable during the flight,the load and endurance time could meet the operation requirements,the acquired remote sensing images performed a relatively high temporal and spatial resolution after mosaic preprocessing,and the system can be used to efficiently obtain the visible light,thermal infrared and multispectral remote sensing images at the farmland scale.（2）Based on the measured data,the sensitivity of soil temperature（T_S）and soil water filling porosity（WFPS）at different depths to R_S was analyzed at the daily scale.The results showed that T_S and WFPS at 5 cm exhibited the highest correlation coefficient with R_S.At the seasonal scale,the relationships between R_S and the biotic factors（LAI,chlorophyll content,i.e.,CHC,）and abiotic factors(T_SF,WFPS,T_S)were analyzed.The results indicated that T_SFand T_S at 5 cm depth performed the similar correlation coefficients with R_S,followed by CHC.Three machine learning algorithms,MNR,BPNN and support vector machine regression（SVR）,were used to construct R_S estimation models at the daily and seasonal scales under different water treatments.The results showed that the MNR,BPNN and SVR models could estimate R_S well,and R_S had a strong dependence on the driving factors of the models.BPNN model exhibited the best prediction of R_S both daily and seasonal scales.In addition,the asynchronous change between R_S and T_SF was deeply discussed.The results indicated that water stress affected the lag relationship between R_S and T_SF.The correlation between R_S and T_SF decreased with the increase of water stress,which affected the accuracy of R_S estimation.（3）T_SF estimated by thermal infrared remote sensing performed a good agreement with the measured（R²=0.85）.The average error of the four water treatment areas was 0.77℃,1.21℃,1.43℃and 1.32℃,and the root mean square error was 0.98℃,1.48℃,1.68℃and 1.55℃,respectively.The correlation analysis between R_S and nine VIs was carried out.The results showed that the three vegetation indices(normalized vegetation index 2(NDVI_g-b),normalized pigment chlorophyll index（NPCI）,simple pigment ratio index（SRPI）sensitive to crop coverage and CHC were highly correlated with seasonal R_S.The R_S estimation model（VHI model）was constructed based on VI estimated by multispectral remote sensing and T_SFestimated by thermal infrared remote sensing.The results indicated that the R² of VHI model(NDVI_g-b,NPCI,SRPI,respectively)was 0.71,0.73 and 0.75,and the RMSE was 182.7 mg m^-2 h^-1,136.9 mg m^-2 h^-1 and 133.3 mg m^-2 h^-1,respectively.Compared with the BPNN and MNR models,the results showed that the accuracy of VHI model was similar to that of BPNN（R²=0.81）,but better than that of MNR model（R²=0.66）.Considering that the BPNN model can’t be expressed by explicit expression,and the complexity of the model was high,the VHI model with a relatively high accuracy was selected to estimate the instantaneous R_S and daily total emission flux.（4）At the daily scale,the variation of R_S under different water treatments was analyzed.The growth rate of soil carbon flux exhibited a similar magnitude under different water treatments from 6:00 to 12:00 during the daytime.The more serious the water stress was,the less CO₂ the soil emitted.（5）A linear regression model was established between the instantaneous R_S estimated by VHI model and the daily total emission flux measured on the ground.The results showed that the fitting accuracy of the model was relatively high,R²=0.73,RMSE=2893.54 mg m^-2.The spatial distribution maps of daily total emission flux of R_S in different growth stages of summer maize were generated.Our results showed that it was feasible to use UAV remote sensing and ground monitoring data to estimate R_S.