KMS XINJIANG INSTITUTE OF ECOLOGY AND GEOGRAPHY，CAS
| 天山区域典型灌区生长季土壤盐分含量时空差异分析 | |
| Alternative Title | Spatial and temporal differences of soil salt content in typical irrigation areas of Tianshan region during growing season |
| 徐红涛 | |
| Subtype | 硕士 |
| Thesis Advisor | 郑宏伟 |
| 2020-06-30 | |
| Degree Grantor | 中国科学院大学 |
| Place of Conferral | 北京 |
| Degree Discipline | 理学硕士 |
| Keyword | 机器学习 支持向量回归 土壤盐渍化 建模变量和模型参数同步优选 时空变化 Machine learning Support vector regression Soil salinization Simultaneous optimization of feature subset and model parameters temporal and spatial changes |
| Abstract | 土壤盐渍化是干旱区土地退化的主要表现形式之一， 遥感结合机器学习算法已成为大尺度土壤盐渍化监测的主要手段之一。对于机器学习算法而言，建模变量和模型参数对于模型精度提高至关重要，以往研究较少关注两者的同步优化。本研究依托机器学习、遥感等技术手段，基于 Landsat 5 TM/8 OLI、 DEM 数据提取 7 类 40 个环境变量，经相关分析，分别带入格网搜索算法（Gride Search,GS）、遗传算法（Genetic Algorithm, GA）和粒子群（Particle Swarm Optimization,PSO）算法分别同步优选支持向量回归（Support Vector Regression, SVR）的建模变量和模型参数，分别建立三工河流域和玛纳斯灌区的盐渍化监测模型（GS-SVR、GA-SVR、 PSO-SVR），并分别分析不同土地利用类型的盐渍化分布和生长季内土壤盐分含量的动态变化特征；而后选择最优的优化算法，结合历史采样数据，分别建立天山南北子区（渭-库绿洲、玛纳斯河水流域）盐渍化监测模型并反演2008~2019 年的两个子区的土壤盐分含量。结果表明：(1)在三工河流域， 相较于 GS-SVR， GA-SVR 的 R2/RMSE 提高了 44.65%。该区非、轻度、中度、重度盐渍地和盐土的面积占比分别为 42.83%、 11.02%、15.88%、 9.22%、 21.05%； 草地和未利用地主要以非盐渍地和盐土为主，耕地和林地中非盐渍地分布比例均为最大； 不同土地利用类型的均值和土壤盐分含量标准差均呈现未利用地>草地>耕地>林地的规律。（2） 在玛纳斯灌区，与 GS-SVR 相比， GA-SVR 和 PSO-SVR 减少了建模变量，适应度值分别提高了 53.87%、 69.96%；生长季内，春秋季积盐，夏季脱盐，土壤盐分含量均值变化趋势：整个研究区、中部和南部为降低-增加；北部为增加-降低-增加；依据生长季土壤盐分含量小提琴图表明土壤盐分含量数值区间变化趋势为：整个研究区，中部和北部 SSC 数值区间变化趋势为扩张-收缩-扩张，南部为扩张-收缩-平稳。（3）提取的植被指数和特征空间对土壤盐分含量的敏感性优于其他环境变量。（4）基于 PSO 和 SVR 建立的渭-库绿洲耕地盐渍化监测模型 PSO-SVR（R2=0.722， RMSE=14.931 g/kg）并反演了该区 2008-2019 年的土壤盐分含量。在 2008~2011 年期间， 盐分含量高值主要分布在绿洲边缘的耕地中， 盐分含量的均值和标准差均呈减小的趋势；在 2013~2019 年期间， 盐分含量高值分布较少且呈向下游地势低洼处转移的趋势， 盐分含量的均值和标准差波动较大。 在2008~2019 年期间，非盐渍地的面积最大且以 50.002×km2/a 的速度增加，中度和轻度盐渍地的面积增加速率分别为 42.243 km2/a、 29.239 km2/a，重度盐渍地面积的减少速率为 8.347km2/a；盐土面积的减少速率在 2008~2011 年（160.33km2/a）高于 2013~2019 年（7.974km2/a）。（5） 基于 PSO 和 SVR 建立了玛纳斯河流域的耕地盐渍化监测模型 PSOSVR (R2=0.812， RMSE=4.748 g/kg)并反演了 2008-2019 年的土壤盐分含量。在2008~2019 年，整个流域非盐渍地的分布面积最大且以 110.07 km2/a 的速度增长，中度盐渍地、盐土次之且分别以 35.403km2/a、 43.957km2/a 的速度减少，轻度、重度盐渍地面积较小且分别以 5.731km2/a、 24.983km2/a 的速度减少；流域南部的盐渍化最为严重，以中度盐渍地为主且呈 5.687 km2/a 的速度减小，非盐渍地次之，增长速率为 24.62 km2/a，轻度盐渍地、重度盐渍地和盐土三者的面积相当，增加速率分别为 0.822 km2/a、 12.731 km2/a 和 7.023 km2/a；中部次之，以非盐渍地为主，增长速率约为 60.341 km2/a，其次为中度盐渍地，减少速率约为 24.136km2/a，轻度盐渍地、重度盐渍地和盐土的面积相当，其增加、减少、减少速率分别为 0.230 km2/a、 13.842 km2/a、 22.592 km2/a。北部较轻，以非盐渍地为主，增加速率为 25.462 km2/a，中度盐渍地次之，减少速率约为 5.645 km2/a，盐土和重度盐渍地的分积相当，其减少速率分别为 8.835 km2/a 和 4.238 km2/a，轻度盐渍地分布面积较少，减少速率约为 6.745 km2/a。（6）渭-库绿洲盐渍化程度轻于玛纳斯河流域、南部和中部，重于北部。 自然因素和人为因素均会导致土壤盐渍化，温度，人口数量对渭-库绿洲盐渍化的影响更大，而土壤质地和地下水埋深对玛纳斯河流域的盐渍化作用更为明显。本研究的建模变量和模型参数同步优选的方法提高了盐渍化监测的准确度，为干旱区高精度土壤盐渍化定量反演提供了技术支持。 |
| Other Abstract | Soil salinization is one of the main manifestations of land degradation in aridregions. Remote sensing combined with machine learning algorithm has become oneof the main means of large-scale soil salinization monitoring. For machine learningalgorithms, feature subset and model parameters are critical to improving modelaccuracy, while previous studies have paid less attention to the simultaneousoptimization of the above two.Relied on machine learning, remote sensing and other technical methods, thisstudy extracts 40 environmental variables of 7 categories based on Landsat 5 TM / 8OLI and DEM data. After correlation analysis, environmental fators of significtantrelated to measured soil salt content were brought into Grid Search algorithm(GS),Genetic Algorithm(GA) and Particle Swarm Optimization(PSO) to select feature subsetand model parameters of Support Vector Regression(SVR), based on which thesalinization monitoring models(GS-SVR, GA-SVR, PSO-SVR) of Sangong RiverBasin and Manasi River Basin were established and the distribution of salinization inthe above two regions were retrieved , respectively. Moreover, the distribution ofsalinization in different land use in Sangong River Basin, 2016 and the dynamic chageof soil salt content in growing season in Manasi Irrigation Distric, 2016 were analyed,respectively.(1) In Sangong River Basin, compared with GS-SVR, the R2/RMSE of AGA-SVRincreased by 44.65%. In terms of the different types of salinized soil, the proportion ofnon-salinized soil, slightly salinized soil, moderately salinized soil, severely salinizedsoil, saline soil in Sangong River Basin was 42.83%, 11.02%, 15.88%, 9.22%, 21.05%,respectively. In terms of the distribution of SSC in different land use types, the unusedland and grassland were mainly comprised of non-salinized soil and saline soil, whilethe distribution proportion of non-salinized soil were the largest in farmland and forestland. Moreover, the mean and standard deviation of SSC of different land use typeswere in the order of unused land > grassland > farmland > forest land.（ 2） In Manasi Irrigation District, compared with GS-SVR, the GA-SVR andPSO-SVR improved the accuracy of the salinization monitoring while reducing thenumber of feature subset, and the fitness value increased by 53.87% and 69.96%,respectively. During the growing season, salt accumulates in spring and autumn andfades in summer. The trend of average SSC of the whole study area and the central part and the southern part was decreasing-increasing, while the northern part was increasingdecreasing-increasing. According to the SSC violin plots in the growing season, it wasfound that the trend of SSC range of the whole study area and the central part and thenorthern part was expansion-contraction-expansion, while it was expansioncontraction-stability in southern part of the north slope of Tianshan Mountain.(3) The extracted vegetation index and feature space were more sensitive to soilsalt content than other environmental factors.(4) Based on PSO and SVR, a monitoring model PSO-SVR (R2 = 0.722, RMSE =14.931 g / kg) for salinization of cultivated land in Wei Ku oasis was established, andthe soil salt content in 2008-2019 was retrieved. From 2008 to 2011, the high value ofsalt content is mainly distributed in the cultivated land at the edge of oasis, while themean value and standard deviation of salt content are decreasing; from 2013 to 2019,the high value of salt content is less distributed and tends to transfer to the low-lyingareas in the downstream, while the mean value and standard deviation of salt contentfluctuate greatly. From 2008 to 2019, the area of non saline land was the largest andincreased at the rate of 50.002 × km2/a, and the area of moderate and light saline landincreased at 42.243 km2/a and 29.239 km2/a, respectively, the area of heavy saline landdecreased at 8.347 km2/a. The area of saline soil decreased at 160.33km2/a in 2008-2011, which was higher than that in 2013-2019 (7.974km2/a).（5） Based on PSO and SVR, a monitoring model PSO-SVR (R2 = 0.812, RMSE= 4.748 g / kg) for farmland salinization in Manas River Basin was established, and thesoil salt content from 2008 to 2019 was retrieved. From 2008 to 2019, the distributionarea of non saline land in the whole basin was the largest and increased at the rate of110.07 km2/a, followed by moderate saline land and saline land, which decreased at therates of 35.403km2/a and 43.957km2/a respectively, while the area of light and heavysaline land was smaller and decreased at the rates of 5.731km2/a and 24.983km2/arespectively。 In the south of the basin, the salinization is the most serious, the moderatesalinization land is the main and the decreased at the speed of 5.687 km2 /a, the nonsalinization land is the second and grows at the rate of 24.62 km2 /a. The areas of lightsalinization land, heavy salinization land and salinization land are of less difference andthe increase rate is 0.822 km2/a, 12.731 km2/a and 7.023 km2/a respectively. In themiddle part of the basin, non saline land is main and increases at the rate of 60.341km2/a, follwed by moderate saline land and decrease at the rate of 24.136 km2/a; theareas of light saline land, heavy saline land and saline soil are of little difference, while were increased, decreased and decreased at the rate of 0.230 km2/a, 13.842 km2/a and22.592 km2/a respectively. In the northern part of the basin, the non saline land is themain with an increase rate of 25.462 km2/a, followed by the moderate saline land witha decrease rate of about 5.645 km2/a. The areas of saline land and heavy saline land areof little difference, decreasing at the rate of 8.835 km2/a and 4.238 km2/a, respectively.The distribution area of light saline land is less, with a decrease rate of about 6.745km2/a.(6) The salinization in Wei-Ku oasis was less serious than Manasi IrrigationDistrict, southern part and middle part, while more serious than northern part. Naturaland human factors will lead to soil salinization. Temperature and population have agreater impact on the salinization of Wei Ku oasis, while soil texture and groundwaterdepth have a more obvious effect on the salinization of Manas River Basin.The method of synchronous optimization of feature subset and model parametersin this study improves the accuracy of salinization monitoring which provides technicalsupport for high-precision quantitative inversion of soil salinization in arid areas. |
| Subject Area | 地图学与地理信息系统 |
| Language | 中文 |
| Document Type | 学位论文 |
| Identifier | http://ir.xjlas.org/handle/365004/15455 |
| Collection | 中国科学院新疆生态与地理研究所 研究系统 |
| Affiliation | 中国科学院新疆生态与地理研究所 |
| First Author Affilication | 中国科学院新疆生态与地理研究所 |
| Recommended Citation GB/T 7714 | 徐红涛. 天山区域典型灌区生长季土壤盐分含量时空差异分析[D]. 北京. 中国科学院大学,2020. |
| Files in This Item: | There are no files associated with this item. | |||||
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.
Edit Comment