基于机器学习的土壤温度预估研究综述

A Review of Soil Temperature Estimation Research Based on Machine Learning

土壤温度是地球科学研究中的重要物理量。在陆-气相互作用研究中,土壤温度不仅影响土壤内部的物理、生物、化学过程,而且对陆-气之间能量和物质交换起重要作用。随着可获取的相关数据越来越丰富,机器学习方法已经被越来越多的研究人员引入到土壤温度预估中,在很多任务中已经超过了统计模型、物理模型的性能。本文对比了统计模型、物理模型和机器学习方法这三种土壤温度常用计算方法的异同,简要介绍了应用于土壤温度研究的各类机器学习模型的原理和特点,综合国内外文献归纳了传统机器学习和深度学习在土壤温度空间分布、时间变化和时空变化三方面的研究进展。在土壤温度空间分布研究中,传统机器学习方法能够通过影响因子的空间异质性学习空间特征,并利用站点观测数据计算土壤深处的温度,但随土壤深度增加模型效果减弱,而深度学习模型有能够提取空间特征的结构,但对数据量要求高,当前研究中仅用于地表温度的遥感反演;在土壤温度时间序列研究中,加入了周期性信息的传统机器学习方法具有更好的模型效果,深度学习中的序列学习模型能自动捕捉土壤温度变化规律,结合了非平稳序列分析方法的混合模型能充分考虑土壤温度变化的连续性和周期性;由于陆面过程复杂性,土壤温度时空变化方面研究较少。基于模型特点和研究现状,本文总结了机器学习在土壤温度预估时的适用性,对数据选择、模型选择和模型评估方法进行了讨论。不同数据条件、研究目的决定着数据和模型的选择,决策树类方法可以可视化提供一定可解释性,支持向量机可适用于数据量较少的情况,极限学习机可以满足需要快速计算情形。由于机器学习缺少物理约束,应用于土壤温度预估时需要重视模型检验和结果对比。针对当前研究的挑战对下一步工作进行展望,认为未来利用机器学习方法对土壤温度进行预估可以从在学习模型中融入先验科学知识、结合遥感资料和多层观测进行土壤温度立体空间建模以及利用卷积循环神经网络进行时空建模三方面进行。

Soil temperature is an important physical quantity in earth science research. In the study of Land-atmosphere interaction,soil temperature not only affects the physical,biological and chemical processes of the underlying surfaces,but also plays an important role in the energy and material exchange between land and atmosphere. With the availability of more and more relevant data,machine learning methods have been introduced into soil temperature prediction by more and more researchers,and have surpassed the performance of statistical models and physical models in many tasks. This paper compares three common methods of soil temperature calculation:statistical model,physical model and machine learning method,and briefly introduces the principles and characteristics of various machine learning models applied to soil temperature research. Based on domestic and foreign literatures,the research progress of traditional machine learning and deep learning in three aspects of soil temperature is summarized. In the study of the spatial distribution of soil temperature,traditional machine learning methods can learn spatial characteristics through the spatial heterogeneity of influencing factors,and use site observation data to calculate the temperature in the depth of the soil,but the model effect weakens as the soil depth increases. While deep learning model has a structure that can extract spatial features,it has high requirements on the amount of data,and is only used for remote sensing inversion of surface temperature in current research. In the study of soil temperature time series,the traditional machine learning method with periodic information has better performance,the sequence learning model in deep learning can automatically capture the law of soil temperature changes,and the hybrid model combined with the non-stationary sequence analysis method can fully consider the continuity and periodicity of soil temperature changes. Due to the complexity of land surface processes,there are few studies on the temporal and spatial variation of soil temperature. Based on model characteristics and research status,this paper summarizes the applicability of machine learning in soil temperature prediction,and discusses data selection,model selection,and model evaluation methods. Different data conditions and research purposes determine the choice of data and models. Decision Tree methods can provide a certain degree of interpretability through visualization. Support Vector Machines can be applied to situations with a small amount of data,and Extreme Learning Machines can meet the needs of fast computing. Due to the lack of physical constraints in machine learning,model testing and comparison of results should be emphasized when applied to soil temperature prediction. In view of the challenges of current research,the future work is prospected.The use of machine learning methods to predict soil temperature can be carried out from three aspects in the future:integrating prior scientific knowledge into the learning model,combining remote sensing data and multilayer observations for soil temperature three-dimensional spatial modeling,and using convolutional recurrent neural networks for spatio-temporal modeling.