收藏设为首页 广告服务联系我们在线留言

干旱区地理 ›› 2023, Vol. 46 ›› Issue (1): 103-114.doi: 10.12118/j.issn.1000-6060.2022.277

• 生物与土壤 • 上一篇    下一篇

基于机器学习和多光谱遥感的银川平原土壤盐分预测

魏慧敏1(),贾科利1(),张旭1,张俊华2   

  1. 1.宁夏大学地理科学与规划学院,宁夏 银川 750021
    2.宁夏大学生态环境学院西北土地退化与生态恢复国家重点实验室培育基地,宁夏 银川 750021
  • 收稿日期:2022-06-11 修回日期:2022-07-17 出版日期:2023-01-25 发布日期:2023-02-21
  • 通讯作者: 贾科利(1975-),男,博士,教授,主要从事3S与土地利用研究. E-mail: jiakl@nxu.edu.cn
  • 作者简介:魏慧敏(1998-),女,硕士研究生,主要从事遥感监测与分析研究. E-mail: weihm09@163.com
  • 基金资助:
    国家自然科学基金项目(42061047);国家自然科学基金项目(42067003);宁夏回族自治区重点研发计划项目(2021BEG03002)

Prediction of soil salinity based on machine learning and multispectral remote sensing in Yinchuan Plain

WEI Huimin1(),JIA Keli1(),ZHANG Xu1,ZHANG Junhua2   

  1. 1. College of Geographical Sciences and Planning, Ningxia University, Yinchuan 750021, Ningxia, China
    2. Breeding Base for State Key Laboratory of Land Degradation and Ecological Restoration in Northwestern China, School of Ecology and Environment, Ningxia University, Yinchuan 750021, Ningxia, China
  • Received:2022-06-11 Revised:2022-07-17 Online:2023-01-25 Published:2023-02-21

摘要:

快速获取区域土壤盐渍化程度信息,对于盐渍化治理与生态环境保护具有重要意义。以银川平原为研究区,以盐分影响因子和盐分指数分别作为输入参数,建立支持向量机(SVM),BP神经网络(BPNN)和贝叶斯神经网络(BNN)3种土壤盐分预测模型,选取最佳模型进行研究区不同深度的土壤盐渍化预测。结果表明:(1) 0~20 cm土壤盐分预测模型中基于影响因子变量组的BNN模型效果最佳,决定系数(R2)为0.618,均方根误差(RMSE)为2.986;20~40 cm土壤盐分预测模型中基于盐分指数变量组的BNN模型效果最佳,R2为0.651,RMSE为1.947;综合对比下,BNN模型的预测效果最好,可用于研究区土壤盐渍化预测。(2) 银川平原主要是以非盐渍化和轻度盐渍化为主,0~20 cm土壤重度盐渍化及盐土共占总面积的11.59%,20~40 cm土壤重度盐渍化及盐土共占总面积的7.04%,20~40 cm土壤盐渍化程度较0~20 cm土壤盐渍化轻。

关键词: 机器学习, 土壤盐分预测, 贝叶斯神经网络, 银川平原

Abstract:

Soil salinization can hinder agricultural development. In this study, the degree of regional soil salinization was obtained to provide a theoretical reference for improving agricultural land quality. Using Yinchuan Plain of China as the study area with a grid size of 5 km×5 km, the soil salinity data of 166 sampling points at different depths were obtained. Combined with the Landsat 8 OLI image corresponding to the sampling time, the salt influence factor and salt index were used as input parameters, respectively, and soil salinity at field sampling points was used as output layer parameters. Support vector machine, back propagation neural network, and Bayesian neural network (BNN) were established as soil salinity inversion models. The determination coefficient and root mean square error of the different models were compared to screen the best model. Finally, soil salinization inversion at different depths was performed in the study area. The following results were obtained: (1) In the 0-20 cm soil salinity inversion model, the BNN model based on the influence factor variable group of salinization was the best, with a coefficient of determination (R2) and root mean square error (RMSE) of 0.618 and 2.986, respectively; the best inversion result of 20-40 cm soil salinity was the BNN model based on the salt index variable group (R2=0.651; RMSE=1.947); the comparative analysis of the modeling and verification effects of different variables of the selected algorithms revealed that the BNN model was the best inversion model with a better fitting degree than the other two models, and the introduction of a neural network had certain advantages in the model construction. (2) Non-salinized and mildly salinized soils were the main soil types in Yinchuan Plain. Soil salinization showed a low trend in the south and a high trend in the north. The 20-40 cm soil salinization was found to be lighter than the 0-20 cm soil salinization.

Key words: machine learning, soil salinity prediction, Bayesian neural network, Yinchuan Plain