融入土壤湿度指标的青藏高原近地表土壤冻融机器学习监测算法

Using Machine Learning Algorithms to Monitor Near-surface Freeze/Thaw State by Considering Soil Moisture in Tibetan Plateau

青藏高原作为中低纬度地区最大的高山冻土区,多年冻土和季节冻土广泛分布。高精度的地表冻融监测结果对研究该区域的水热交换、碳氮循环和土壤冻融侵蚀非常重要。本文基于4个青藏高原典型地区的土壤温湿度观测网数据,开展利用LightGBM算法和随机森林算法进行土壤冻融循环监测的研究。在构建土壤冻融监测模型的过程中,发现土壤湿度是影响冻融判别的一个关键因子。使用AMSR2亮温数据和ERA5-Land土壤湿度数据,基于两种机器学习算法判别地表冻融状态,将结果与传统冻融判别式算法进行对比分析。结果表明:相比冻融判别式算法,LightGBM算法在白天和夜间的总体判对率提高了12.09%;14.45%,随机森林算法在白天和夜间的总体判对率提高了13.23%和14.96%。近80%的错分样本分布在-4.0℃~4.0℃之间,说明2个机器学习算法能够识别出稳定的土壤冻结状态和融化状态。另外,LightGBM算法和随机森林算法得到的日冻融转换天数的平均RMSE降低了112.82和117.00;冻结天数的平均RMSE降低了47.87和53.96;融化天数的平均RMSE降低了37.10和39.80。同时,基于随机森林算法计算了2014年7月—2015年6月青藏高原冻结天数、融化天数、日冻融转换天数。得到的青藏高原冻结天数图,以中国冻土区划图为参考进行精度评价,总体分类精度为96.78%。

As the largest alpine permafrost area in the middle and low latitudes,permafrost and seasonally frozen soil are widely distributed in the Tibetan Plateau(TP).Accurate spatiotemporal observation of surface freeze/thaw state in the TP is important for quantifying surface energy balance,carbon and nitrogen exchange,and soil freeze-thaw erosion.However,land surface freeze/thaw state can hardly be detected in this area because of its harsh and complex geographical environment.This study aimed to employ the LightGBM algorithm and random forest algorithm to identify near-surface freeze/thaw state,based on four soil temperature&moisture observational networks.Previous studies have shown that soil moisture could significantly affect the seasonal variation characteristics of near-surface soil freeze-thaw cycles.In this study,soil moisture was introduced as a discriminant feature.In order to illustrate the contribution of microwave brightness temperature,discriminant index,and soil moisture,four different feature combination schemes were designed.We utilized AMSR2brightness temperature data and ERA5-Land soil moisture data to identify the surface freeze/thaw state using these two machine learning algorithms.By evaluating the importance of different features based on the training set,we found that the importance score of soil moisture was high in both LightGBM and random forest algorithms,which indicates that soil moisture is a very important feature that affects freeze-thaw discrimination.To evaluate the performance of our algorithms,we compared LightGBM and random forest algorithms with a traditional freeze-thaw discriminant algorithm.Results show that the accuracy of the two machine learning algorithms was higher than that of the traditional method,and the overall rate of correct classification for daytime and nighttime was increased by 12.09%,14.45%,respectively using LightGBM,and 13.23%,14.96%,respectively using random forest.Nearly 80%of the misclassification occurred when the surface soil temperature was between-4.0℃and 4.0℃.So the two machine learning algorithms are able to identify stable soil freeze/thaw state.In addition,the average RMSE of the freeze-thaw conversion days obtained by the LightGBM algorithm and the random forest algorithm decreased by 112.82,117.00,respectively;the average RMSE of the frozen days decreased by 47.87,53.96,respectively;and the average RMSE of the thawed days decreased by 37.10,39.80,respectively.Based on random forest algorithm,we calculated the number of frozen days,number of thawed days,and number of freeze-thaw conversion days from July 2014 to June 2015.The accuracy assessment was carried out using the map of permafrost classification as the reference,and the total classification accuracy of frozen days within the permafrost zone was 96.78%.