基于自组织映射与随机森林耦合模型的流域水质空间差异性评估

Self-organizing map random forest coupling model based spatial heterogeneity evaluation of water quality in the watershed

流域水环境质量空间分布特征分析是推进流域精细化管理的基础.本研究基于流域特征指标与水质的关联性,以子流域为分析单元,利用自组织映射人工神经网络模型(SOM)对苕溪流域水质数据聚类分析为3类后与随机森林模型(RF)进行耦合,对全流域水质进行了空间差异性评估.研究结果显示,上游山地区域水质较好,而平原河网人口集聚区的CODMn、NH3-N及TP浓度较高,山地与平原过渡地带水质则主要受到CODMn和TN的影响.采用自然环境、社会经济及土地利用/覆盖指标作为流域特征进行水质分级模式识别,SOM与RF模型耦合模型的准确率稳定在80%左右;在对强相关性特征进行筛选识别后,将蒸发蒸腾量、坡度、人口密度、大于10℃积温、旱地占比、城镇用地占比及景观多样性指数为作为输入特征,准确率可达83%,可以有效地开展全流域水质分级评估.

A systematic understanding of the spatial heterogeneity of water quality within watershed is very important for precision watershed management.In this study,sub-watershed scaled water monitoring data in Tiaoxi watershed was classified into 3 clusters by self-organizing map(SOM),and the cluster results was used as input features of water quality characteristics to train random forest model(RF).The trained RF model was then applied to the spatial heterogeneity assessment of watershed water quality.Results showed that water quality was better in the upstream mountainous region,and the concentrations of CODMn,NH3-N and TP were higher the plain area with high population density.The water quality of mid-stream area was mainly influenced by CODMn.Water quality pattern recognition was conducted by using watershed features including natural environmental and socio-economic features and land use/cover,and the generalization accuracy of SOM-RF coupling model was approximately 80%.After the identification of strongly correlated features,the accuracy of the model can reach 83%by using evapotranspiration,slope,population density,accumulated temperature greater than 10 degrees Celsius,dry field,urban are proportion,and the Shannon-Ville index of for landscape diversity as input features.The results indicate SOM and RF coupling model can provide decision support for water quality management and spatial control in the watershed.