基于随机森林算法与农户视角的村落级农业面源污染风险分析  

Risk Analysis of Agricultural Non-point Source Pollution at Village Level Based on Random Forest Algorithm and Farmers’Perspective

作　　者：朱康文 张晟 陈玉成[2] 熊海灵[3] 雷波 黄昌前 段秋宴 ZHU Kangwen;ZHANG Sheng;CHEN Yucheng;XIONG Hailing;LEI Bo;HUANG Changqian;DUAN Qiuyan(Chongqing Academy of Eco-environmental Science,Chongqing 401147,China;College of Resources and Environment,South-west University,Chongqing 400716,China;College of Computer&Information Science,Southwest University,Chongqing 400716,China;Bureau of Ecological Environment of Qijiang District,Chongqing 400803,China;Ecological Environment Monitoring Station of Yubei District,Chongqing 401120,China)

机构地区：[1]重庆市生态环境科学研究院,重庆401147 [2]西南大学资源环境学院,重庆400716 [3]西南大学计算机与信息科学学院,重庆400716 [4]重庆市綦江区生态环境局,重庆400803 [5]重庆市渝北区生态环境监测站,重庆401120

出　　处：《三峡生态环境监测》2022年第1期61-72,共12页Ecology and Environmental Monitoring of Three Gorges

基　　金：重庆市技术创新与应用发展示范专项重点研发项目(cstc2019jscx-gksbX0103)。

摘　　要：当前农业面源污染问题依然严峻,村落级尺度下各地块的污染风险状态精细识别有待进一步研究。本文针对三峡库区典型村落重庆市涪陵区南沱镇睦和村进行研究,结合无人机多光谱技术、农户行为调研、随机森林算法等进行地物识别及泛地块网格划分,通过测算总氮(total nitrogen,TN)、总磷(total phosphorus,TP)施用水平达到识别各网格污染负荷与风险空间分异情况的目的。结果显示:(1)无人机多光谱数据技术结合随机森林算法可有效用于地物精细化分类,精度优于90%。(2)基于农户行为调研与泛地块网格划分框架下的污染负荷及风险测算,有利于识别村落尺度的污染负荷及风险差异。一级网格中,仅有1个网格属于无风险等级;存在种植的二级网格中,分别有124、285、63、1309个网格的TN施用强度处于无风险、低风险、高风险和极高风险状态,分别有409、205、1167个网格的TP施用强度处于无风险、中风险和极高风险状态。(3)区域已有人工湿地污染物的消减能力明显,悬浮物、TP、TN、铵态氮、硝态氮、生化需氧量分别消减了近86.67%、54.66%、81.11%、10.67%、83.85%、59.42%。劳动力下降、人口老龄化等问题引起作物复种指数高、高肥高药行为是导致区域污染风险较高的主要原因。本研究可为村落管理、指导农户行为提供有效的数据支撑和工程指导。At present,the problem of agricultural non-point source pollution is still serious,and the fine identification of pollution risk status of each plot at village level needs to be further studied.This paper focused on a typic village of the Three Gorges reser⁃voir area,namely,Muhe village,Nantuo Town,Fuling District of Chongqing city.The UAV multispectral technology,farmers’behav⁃ior survey and random forest algorithm were used to identify the features of each pixel,and the classification data were obtained by field verification.By calculating the application levels of total nitrogen(TN)and total phosphorus(TP),the pollution load and risk spatial differentiation of each grid were identified.The results show that:(1)the classification accuracy of UAV multispectral data combined with random forest algorithm was 90%.(2)The research framework based on the combination of farmers’behavior re⁃search and pan plot grid division was conducive to identify the pollution load and risk differences at the village scale.In the first level grid,only one grid belongs to no-risk level;among the secondary grids with planting,124,285,63 and 1309 grids in TN were in no-risk,low-risk,high-risk and extremely high-risk states,respectively,and 409,205 and 1167 grids in TP were in no-risk,medi⁃um risk and extremely high-risk states,respectively.(3)The pollutant reduction capacity of the constructed wetlands in the region was obvious,with suspended matter,TP,TN,ammonium nitrogen,nitrate nitrogen and chemical oxygen demand(COD)reduced by 86.67%,54.66%,81.11%,10.67%,83.85%and 59.42%,respectively.The high multiple-corp index and heavy utilization of pesticide and fertilizer caused by the decline of labor force and aging population were the main reasons for the high risk of regional pollution.The research can provide effective data support and engineering guidance for village management and guiding farmers’behavior.

关 键 词：农业面源污染 多光谱 随机森林算法 风险分析 

分 类 号：X87[环境科学与工程—环境工程]