Abstract: Scientific analysis of grain production forecasting and its influencing factors is of great significance for making food production decisions and ensuring food security. In this paper, the grain data of Chongqing from 1997 to 2021 were collected, sorted out and analyzed, and it was found that the characteristics and the grain yield of Chongqing were not simple linear, so the nonlinear model was used to fit the functional relationship between the grain yield and its influencing factors, and the Gaussian process regression (GPR) model of three nuclear functions was trained and combined prediction, and the experimental results showed that the combined model had good generalization ability. Taking the 2020—2021 data as the test set, the absolute percentage error of the combined prediction model for the 2020 and 2021 data predictions was 0.074 4% and 0.632 4%, respectively. However, the functional relationship between grain yield and its influencing factors was not easy to obtain by GPR, which made it difficult to analyze the importance of influencing factors using GPR model, and then the influencing factors of Chongqing were analyzed by means of taylors formula for multivariate functions and partial least squares regression (PLSR). Through the PLSR model, it was found that the major factors affecting the grain production in Chongqing were the grain sown area, the total mechanical power, the labor input and the disaster area. The increase of the total mechanical power has reduced the negative impact of the reduction of the grain sown area. Finally, it is recommended to ensure the food security of Chongqing by protecting the area of cultivated land, developing the innovation of agricultural science and technology,  encouraging returntohome entrepreneurship and employment, and strengthening the climate monitoring.

Key words: grain yield prediction, Gaussian process regression, partial leastsquares regression, machine learning

摘要： 科学进行粮食产量预测及其影响因素分析对作出粮食生产决策及保障粮食安全有重要意义。对重庆市1997—2021年粮食数据进行收集、整理、分析，发现各特征与重庆市粮食产量间不是简单线性关系，因此使用非线性模型拟合粮食产量与其影响因素之间的函数关系，训练三种核函数的高斯过程回归(GPR)模型并进行组合预测，试验结果显示所得的组合模型具有很好的泛化能力。以2020—2021年数据为测试集，组合预测模型对2020年和2021年数据预测的绝对百分比误差分别为0.074 4%和0.632 4%。但GPR不易获得粮食产量与其影响因素之间的函数关系，导致使用GPR模型进行影响因素重要性分析很困难，进而借助多元函数泰勒公式及偏最小二乘回归(PLSR)对重庆市粮食产量进行影响因素分析。通过PLSR模型发现对重庆市粮食产量影响较大的因素是粮食播种面积、农用机械总动力、劳动力投入和成灾面积；农用机械总动力的增加降低了粮食播种面积减少等带来的负面影响。最后提出保护耕地面积、发展农业科技创新、鼓励返乡创业就业以及加强气候监测等建议来保障重庆市粮食安全。

关键词: 粮食产量预测, 高斯过程回归, 偏最小二乘回归, 机器学习