机器学习在黄土高原沟壑区插补二氧化碳浓度的潜力评估

Evaluation of the Potential of Using Machine Learning to Interpolate the CO₂ Concentration in the Gully Region of the Loess Plateau

摘要: 以陕西的淳化生态水文实验基地的沟底和塬上的实测二氧化碳浓度数据和空气温度数据，用沟底观测数据评价了多层感知机(MLP)、双向长短期记忆神经网络(Bi-LSTM)、随机森林(RF)、多元线性回归(MLR)等机器学习方法模拟空气二氧化碳浓度的潜力，遴选性能较好的机器学习方法对塬上二氧化碳浓度进行了模拟、插补，并评估插补值的质量。结果表明：在沟底植被覆盖度不同的区域，仅以实测二氧化碳浓度为输入，机器学习方法即可达到可接受的模拟精度；增加空气温度作为输入后，MLP和RF模型能进一步提升模拟精度。利用沟底二氧化碳浓度数据插补塬上浓度时，MLP和RF仍表现出较高的可行性，能有效还原变化趋势，虽然在部分时段对高值存在轻微低估，但其插补结果与实测值的均值、极大值、极小值的最大偏差仅为5%。MLP和RF可以用于进行区域二氧化碳浓度模拟与数据插补。

Abstract: Based on the measured data of carbon dioxide concentration and air temperature at the bottom of the gully and on the plateau of the Chunhua Ecological Hydrological Experimental Base in Shaanxi Province, the potential of machine learning methods such as Multilayer Perceptrons (MLP), Bidirectional Long Short-term Memory (Bi-LSTM), Random Forest (RF) and Multiple Linear Regression (MLR) in simulating air carbon dioxide concentration was evaluated using the observation data from the bottom of the gully. The machine learning method with better performance was selected to simulate and interpolate the carbon dioxide concentration on the tableland, and the quality of the interpolation value was evaluated. The results show that in areas with different vegetation coverage at the bottom of the ditch, when only the measured carbon dioxide concentration is used as the input, the machine learning method can achieve an acceptable simulation accuracy; when air temperature is added as an input, the MLP and RF models can further improve the simulation accuracy. When using the ditch bottom carbon dioxide concentration data to interpolate the concentration on the plateau, MLP and RF still show high feasibility and can effectively restore the changing trend, although there is a slight underestimation of high values at some times, the maximum deviation between the interpolation results and the measured values of mean, maximum and minimum values is only 5%. MLP and RF can be used for regional carbon dioxide concentration simulation and data interpolation.