BP神经网络算法在河西绿洲玉米生产碳排放评估中的应用及算法有效性研究

针对作物生产碳排放预测较为困难的实际问题,提出基于BP神经网络算法的玉米生产碳排放预测模型。选择地处河西走廊石羊河下游的民勤绿洲246家农户,面对面调查玉米种植户农场内生产投入数据,将玉米生产投入数据作为神经网络输入层;查阅和梳理国内外相似区域玉米生产环节碳排放系数,运用碳足迹生命周期法计算得到的碳排放值作为神经网络输出层;基于BP人工神经网络算法,运用试凑法确定网络隐含层节点个数,建立河西绿洲玉米生产碳排放预测模型,选择多元线性回归模型、多元非线性回归模型,对该模型有效性进行评估。研究结果表明,3层且各层节点数9、10、1的神经网络结构能够准确预测河西绿洲玉米生产碳排放,其碳排放预测值为0.763 kg(CO_2-eq)·kg~(-1)(DM);9-10-1结构的神经网络预测模型的相关系数(R~2=0.984 7)高于多元线性和非线性回归模型,该神经网络结构模型的均方根误差(RMSE=0.069 1)、平均绝对误差(MAE=0.051 3)均低于其他模型,BP神经网络算法预测性能明显优于其他预测模型。该研究为准确预测农业生产碳排放提供了新思路和可操作方法。

Application and validity of BP neural networks on prediction of carbon emissions from corn production in Hexi Oasis

Back-propagation (BP) neural network has been widely used in global climate change researches in recent years. There is also increasing research interests in the application of BP neural network on predicting carbon emission from agricultural lands. Hexi Oasis in the northern side of Qilian Mountain accounts for over 30% of total grain and over 70% of commercial grain production in Gansu Province, of which corn is the primary food crop. However, there has been little research in carbon emissions from corn fields in Hexi Oasis. Therefore, the objectives of this study were to predict carbon emissions from corn production in Hexi Oasis using BP neural network algorithm and to validate the performance of BP neural network algorithm against multiple linear regression and non-linear regression models. This study was done in Minqin Oasis (103°05''E, 38°38''N) located at the downstream of Shiyanghe River in Hexi Corridor. Data were collected on 246 local farms in a face-to-face questionnaire-driven survey. The data of production inputs were used as the inputs for the model in farm and the value of carbon emissions calculated using life-cycle assessment based on carbon emission factors published in the literatures about the similar regions and default figures reported by Inter-governmental Panel on Climate Change (IPCC). In order to predict carbon emissions based on BP neural network, the numbers of node in the hidden layer were calculated by trial and error. The results indicated that neural network structure with three layers predicted carbon emissions in corn productions in Hexi Oasis and the number of nodes for the input layer, hidden layer and output layer were 9, 10 and 1, respectively. The evaluated carbon emission was 0.763 kg(CO₂-eq)·kg^-1(DM) in the study area. To verify the validity of the BP neural network model, multiple linear regression and non-linear regression models were developed using the same dataset. The results indicated that the correlation coefficient (R²=0.984 7) of BP neural network model with the 9-10-1 structure was higher than that for the corresponding multiple linear regression and non-linear regression models. Also the root mean square error (RMSE=0.069 1) and mean absolute error (MAE=0.051 3) of BP model were lower than those of the corresponding multiple linear regression and non-linear regression models. Therefore, the performance of BP neural network model was better than that of the regression models. The BP neural network model developed in this study using data collected from the local farms in Hexi Oaiss combined the local practices and regional carbon emission factors, consequently providing a practical tool applicable in the prediction of carbon emissions in corn fields. Moreover, the validity of BP neural network model was also verified through comparison with multiple linear regression and non-linear regression models, which improved the reliability of its practical application. Therefore, the results of this study contributed new ideas and development methods to accurately predict carbon emissions in agricultural fields for the government and scientific community.