地球信息科学学报 ›› 2017, Vol. 19 ›› Issue (7): 983-993.doi: 10.3724/SP.J.1047.2017.00983

• 遥感科学与应用技术 • 上一篇    下一篇

基于神经网络模型的干旱区绿洲土壤盐渍化评价分析

姜红1(), 玉素甫江·如素力1,2,*(), 热伊莱·卡得尔1, 阿迪来·乌甫1   

  1. 1. 新疆师范大学 地理科学与旅游学院 流域信息集成与生态安全实验室,乌鲁木齐 830054
    2. 新疆干旱区湖泊环境与资源重点实验室,乌鲁木齐 830054
  • 收稿日期:2017-01-03 修回日期:2017-02-23 出版日期:2017-07-10 发布日期:2017-07-10
  • 通讯作者: 玉素甫江·如素力 E-mail:xjnujh@163.com;Yusupjan@xjnu.edu.cn
  • 作者简介:

    作者简介:姜 红(1991-),男,重庆人,硕士生,主要从事资源环境遥感研究。E-mail:xjnujh@163.com

  • 基金资助:
    新疆师范大学地理学博士点支撑学科开放课题基金项目(XJNU-DL-201605);国家自然科学基金项目(41161007、41661047);新疆维吾尔族自治区青年科技创新人才培养工程项目(QN2015YX009)

Evaluation and Analysis of Soil Salinization in the Arid Zones based on Neural Network Model

JIANG Hong1(), YUSUFUJIANG Rusuli1,2,*(), REYILAI Kadeer1, ADILAI Wufu1   

  1. 1. Institute of Geographical Science and Tourism / Laboratory of Information Integration and Eco-Security, Xinjiang Normal University, Urumqi 830054, China
    2. Xinjiang Key Laboratory of lake Environment and Resources in Arid Zone, Urumqi 830054, China
  • Received:2017-01-03 Revised:2017-02-23 Online:2017-07-10 Published:2017-07-10
  • Contact: YUSUFUJIANG Rusuli E-mail:xjnujh@163.com;Yusupjan@xjnu.edu.cn

摘要:

土壤盐渍化严重制约了农业可持续发展和生态安全,土壤盐渍化的精确评价分析,对土壤盐渍化的改善和治理具有重要的意义。本文以新疆焉耆盆地为研究对象,Landsat8 OLI遥感影像和实测采样数据相结合,提取地下水埋深(GD)、盐分指数(SI)、地表蒸散量(SET)和改进型温度植被干旱指数(MTVDI)建立了土壤盐渍化评价模型。结果表明:①结合野外实测土壤盐分数据,对BP神经网络模型进行训练。最终以最优的4-4-1结构的3层BP神经网模型对研究区土壤盐渍化进行了预测(R2=0.864,RMSE=0.569)。相比传统多元线性回归模型(R2=0.741,RMSE=0.767),神经网络模型对土壤盐渍化的预测精度更高;②土壤盐渍化分布与GD、SI、SET和MTVDI等存在较强的关联性,不同等级的土壤盐渍化是不同影响因素不同程度上组合而引起的结果,盐渍化土地主要分布在地下水位较低以及土地开垦之后没有利用的荒地区域;③整个研究区大部分区域受到不同程度的盐渍化影响,耕地退化为盐渍地导致该区域土壤盐渍化以及土壤次生盐渍化进一步加剧。

关键词: 盐渍化, 神经网络, 遥感, 预测, 焉耆盆地

Abstract:

Soil salinization is the process of increasing the salt content in the soil. Salinization occurs when the groundwater table is between two and three meters from the surface of the soil in arid lands. The salts from the groundwater are raised by capillary action to the surface of the soil, and it affects human and natural resources, such as native vegetation and crops, animals, infrastructure, agricultural inputs, biodiversity, aquatic ecosystems and water supply quality in the environment. Factors such as climate, features of landscape, soils, drainage, aspect and the effects of human activities all impact on the severity and occurrence of salinization. Therefore, it is an important concern to evaluate and monitor soil salinity in order to take protective measures against further deterioration of the soil. Traditionally, soil salinity evaluation and monitoring are often carried out with intensive field work and sampling. Most previous studies have focused on differentiating salinized and non-salinized soil qualitatively by analyzing the salinity distribution and monitoring its dynamics. Remote Sensing (RS), Geographical Information Systems (GIS) and modeling have recently outperformed the traditional methods. Remotely sensed data has great potential for monitoring dynamic processes, including salinization. Remote sensing of surface features using aerial photography, videography, infrared thermometry, and multispectral scanners has been used intensively to identify and map salt-affected areas. Salinization has seriously restricted the sustainable development of agriculture and ecological security in the Yanqi Basin, Xinjiang, China. Therefore, accurate assessment and monitoring of soil salinization is particularly important. In this paper, based on the Landsat 8 OLI remote sensing data and measured data, the soil salinization evaluation model was established by using the four evaluation indexes of groundwater depth (GD), salinity index (SI), surface evapotranspiration (SET) and modified temperature vegetation dryness index (MTVDI) in Yanqi basin, Xinjiang. Results demonstrate that: (1) BP neural network model for training was combined with the field measured soil salinity data and the best performance was archived in 4-4-1 architecture (R2=0.864, RMSE=0.569) in the three networks. Compared with traditional multiple linear regression model (R2=0.741, RMSE=0.767), the artificial neural network can improve the predictive accuracy of soil salinization. (2) Soli salinization is strongly associated with GD、SI、SET and MTVDI, and the soil salinization are the results of different combinations of different combination effect factors. Salinization is mainly distributed in low groundwater level and unused area. (3) Most of the study area was salinized in different degrees of salinization, and the degradation of farmland led to further soil salinization and secondary soil salinization.

Key words: salinization, neural network, remote sensing, forecasting, Yanqi Basin