| 摘要: |
| 籽粒蛋白含量（Grain protein content,GPC）是衡量小麦品质优劣的重要标准，快速准确的预测小麦GPC有利于其品质评价和分级管理。本研究分别以卫星光谱参数、农学氮素参数以及气象因子为影响因素，并运用多元线性回归模型(Multi-Linear regression, MLR)、极限学习机算法（Extreme Learning Machine, ELM）、地理加权回归模型（Geographically Weighted Regression Model, GWR）三种方法实现对冬小麦GPC 的预测，最终构建及评价基于不同自变量和不同方法的GPC 预测模型。结果表明：1）小麦开花期氮素参数(Leaf nitrogen content, LNC)，小麦冠层光谱参数（Greenness, GI）与小麦籽粒蛋白品质的关系显著相关，影响小麦籽粒蛋白品质的关键性气象因子包括5.26-5.30日降雨和（Rainfall from 26 May to 30 May,rain526）、5月中旬-6月上旬日照和（Sunshine time from mid-May to early June,sun5x6s）、3月上旬-6月上旬积温和（Accumulated temperature from early March to early June,jiwen3s6s）；2）以卫星光谱参数、农学氮素参数和气象因子为自变量，分别采用多元线性回归、极限学习机和地理加权回归三种方法构建小麦GPC的预测模型；其中，基于MLR构建的GPC模型决定系数（Coefficient of determination, R2）为0.598，验证集标准均方根误差(Normalized root mean squared error, nRMSE)和平均绝对误差（Mean absolute error, MAE）分别为10.36%、1.091，验证结果较稳定； 基于ELM构建的GPC模型R2为0.483， 验证nRMSE和MAE分别为10.895、1.111 ；基于GWR的GPC模型建模精度及验证精度相对最优，其建模R2为0.616，验证nRMSE及MAE分别为8.58%、0.956，为最优选择；综合分析模型的精度评价指标可知，考虑空间数据不稳定性构建的地理加权回归模型的预测精度最好，能更加准确的预测冬小麦籽粒蛋白含量，为精确反演冬小麦GPC区域间和年际间的预测提供可靠的依据，具有广泛的应用前景。 |
| 关键词: 遥感 模型 籽粒蛋白含量 极限学习机 地理加权回归 冬小麦 |
| DOI： |
| 分类号: |
| 基金项目:国家重点研发计划（2016YFD0300603，2016YFD0700303）, 国家自然科学基金项目（41371349；41471285） |
|
| Remote Sensing Prediction of Grain Protein Content in Regional Winter Wheat |
|
wangqi, songxiaoyu
|
|
National Engineering Research Center for Information Technology in Agriculture
|
| Abstract: |
| Grain protein content(GPC) is an important factor to evaluate the quality of wheat. Predicting the GPC quickly and accurately is beneficial to grade the wheat quality . In this study, several factors that correlated with GPC evaluation, including meteorological factors in study area, wheat plant nitrogen parameters in flowering stage, as well as satellite spectral parameters for wheat samples, were analyzed of using multi-linear regression (MLR), extreme learning machine algorithm(ELM) and geographical weighted regression (GWR) methods. ThentheGPC prediction models based on different independent variables and methods were evaluated. The results showed that: 1) The nitrogen parameters at wheat flowering stage and the wheat canopy spectral parameters have good correlation with wheatGPC. While rainfall from 26 May to 30 May, sunshine time from mid-May to early June and accumulated temperature from early March to early June are the key meteorological factors affecting the GPC of wheat. 2) The coefficient of determination(R2) of MLR model is 0.598, while the validation precision of Normalized root mean squared error (nRMSE) and Mean absolute error (MAE) are 10.36% and 1.091, respectively. The R2 of the GPC model based on ELM model is 0.483, and the standard nRMSE and MAE of the Validation Set are 10.895% and 1.111, respectively. The R2 of the GPC GWR model is 0.616, and the standard nRMSE and MAE of the validation set are 8.58% and 0.956, respectively. According to the precision evaluation index of the comprehensive analysis model, the multivariate parameter model is superior to the univariate parameter model. The multivariate parameter GWR model, which takes into account the instability of spatial data, has the best prediction accuracy and can predict the GPC more accurately. This study provides a reliable basis for accurately predicting the GPC in different regions and years, and has a broad application prospect in the future. |
| Key words: remote sensing models grain protein content extreme learning machine geographically weighted regression model winter wheat |