Remote sensing inversion models and validation of aboveground biomass in soybean with introduction of terrain factors in black soil area

Abstract: Crop biomass plays an important role in food security and global carbon cycle. Achieving the timely and accurate monitoring of biomass is vital for precise and reasonable agricultural management. Undoubtedly, remote sensing technique has been proven to be an effective tool for biomass estimation. Along with traditional means, it reduces the actual operation and investigation of ground surveys. In ordered to accurately estimate the crop aboveground biomass at the field scale and improve the precision and stability of soybean aboveground biomass inversion model, this paper obtained SPOT-6 6-meter multi-spectral data on July and August 2016 of the study area, as well as the soybean aboveground biomass of different terrain slopes. At the same time, the terrain data of the study area were measured and the topographic factors such as elevation, slope and aspect were extracted. We intended to use above measured data to build three models, which were the traditional linear regression model, the multiple regression model and the neural network model. Firstly, the correlation of the relationships between enhances vegetation index (EVI), normalized difference vegetation index (NDVI) and observed date of soybean aboveground biomass were analyzed by linear regression model. Then we added the terrain factors which were related to the aboveground biomass for establishing multilayer perception stepwise multiple regression and neural network inversion model. Through the model accuracy comparison and estimation accuracy analysis, the results were following: 1) In the linear regression model established by the two vegetation indexes, NDVI Model fitting degree was higher then EVI, and the coefficient of determination (R2) reached 0.712, plus root mean square error (RMSE) was 0.116 kg/m2. The results could be explained that the use of traditional single vegetation index model to predict soybean aboveground biomass was feasible. 2) The neural network multilayer sensor model had the highest precision and reliability among all above (R2= 0.904, RMSE = 0.047 kg/m2). The results of model validation showed that the average absolute and relative error of using neural network model were the smallest, and the values were 0.113 kg/m2 and 0.212, respectively. In the three types of inversion models, the inversion results of the neural network model were closest to the actual data of crop aboveground biomass distribution. The inversion results of this study were in good agreement with the terrain, topography, temperature and precipitation characteristics of the plot and accurately reflected the space distribution features of crop condition and growth. Our research provided a reliable and scientific basis for dynamic monitoring and precise management of soybean aboveground biomass. The method was meaningful in precision agriculture, especially in yield and production prediction.