Research On Maize Crop Residue Coverage Estimation Model Based On Sentinel-2 Data

Crop residue coverage（CRC）as a key indicator of protective farming technology,how to accurately estimate CRC using remote sensing technology has been a major concern for agricultural managers.Therefore,the maize CRC estimation model studied in this paper is of great reference value to the government agricultural administration.This study area is located to the southern part of Songnen Plain,the hinterland of the Northeast Plain,where we first measured maize CRC after harvesting using the Line Transect methods and Photographic methods,then constructed input variables for the corn CRC estimation model using Sentinel-2 data,and finally established the study area maize CRC estimation model.First,in view of the problems of multicollinearity as well as fewer feature variables in part of the spectral indices constructed from sentinel-2 data,we propose to construct textural features using gray level co-occurrence matrix（GLCM）algorithm,which can be used as input variables for the estimation model of maize CRC.The results showed that different sliding window sizes,moving directions and step sizes had different effects on texture features,and the vast majority of texture features and maize CRC were poorly correlated,Only Band4_Mean-5*5,Band5_Mean-5*5,Band6_Mean-5*5,Band3_Mean-3*3,Band8AMean-3*3,Band4Mean-3*3,Band7Mean-5*5,Band8Mean-5*5 and Band8AMean-5*5were strongly correlated with maize CRC,and most of the optimal sliding windows for these textural features were 5*5,the direction of movement was 45°,and the step size was 1.Second,for the problems of low prediction accuracy and weak robustness of the established univariate linear regression model,Partial Least Squares Regression（PLSR）algorithm was chosen to establish the maize CRC estimation model in this paper.The results of this study showed that the R² predicted by the model using the Line Transect methods and Photographic methods established with spectral indices were 0.803 and0.781,respectively,and when the textural features were added,the R² predicted by the model were 0.822 and 0.814,respectively,and it could be found that the addition of textural features could improve the prediction accuracy of the maize CRC estimation model.Finally,to further explore the question of whether machine learning algorithms are feasible in maize CRC.In this paper,Random Forest（RF）algorithm and Back Propagation Neural Network（BPNN）algorithm were selected to establish the remote sensing estimation model of maize CRC.The results showed that the R² predicted by the Line Transect methods and Photographic methods established with the RF algorithm was 0.828 and 0.816,respectively,the BPNN algorithm was 0.884 and 0.887,respectively,with the best prediction.All three methods of estimating maize CRC provided in this study can accurately obtain CRC information,provide a new idea for macro understanding the characteristics of straw distribution at the regional scale,provide a reference for crop residue studies in this area as well as the Northeast Plain,and provide a scientific basis and data support for surface monitoring such as crop residue return,government decision-making,and integrated crop residue use.Through comparative analysis,the accuracy of the three estimation methods was ranked from highest to lowest by BPNN,RF,PLSR.however,BPNN and RF algorithm have large operation volume of data in practical application and relatively low efficiency,while PLSR algorithm is more efficient and slightly low accuracy.