基于RF和SGT算法的子区优先建模对绿洲尺度 土壤盐度预测精度的影响

doi:10.3864/j.issn.0578-1752.2018.24.007

摘要/Abstract

摘要：

目的试图通过优先在干旱区绿洲的子区构建模型以提高绿洲全局土壤盐度的预测精度。同时量化全局模型和子区模型之间精度的差异性和不确定性。方法利用随机森林（Random Forest,RF）和随机梯度增进算法（Stochastic Gradient Treeboost,SGT）定量化上述不确定性,同时,对比本地尺度多个情景(景观)优先建立模型再合并预测值对于模拟全局土壤盐度的精度影响。基于驱动因子（土地利用和地貌）,响应因子（Normalized Difference Vegetation Index, NDVI和土壤电导率,EC）,研究设计了27个能够相对覆盖典型绿洲不同土壤盐度变异性的环境情景。结果 70.37%（19/27）的情景证明SGT的预测精度高于RF。单独建模的10个情景的预测精度高于全局模型下10个再分类情景（根据情景设定规则将全局模型预测值再分类）的精度。特别是,EC≤4 dS·m ^-1 和 2 dS·m ^-1＜ EC＜16 dS·m ^-1两个情景应该单独进行建模预测。4个情景（两两合并）预测值合并后的精度高于全局模型再分类后的精度。需要指出的是,用于绿洲尺度子区情景构建的首选分割变量是EC,其次是地貌和土地利用。结论研究推荐基于SGT在绿洲内部不同景观尺度上优先建模,再将各景观尺度的预测值进行合并,以提高绿洲土壤盐度的推理精度。

关键词: 土壤盐分, 机器学习, 干旱区, Landsat OLI, 空间异质性, 随机森林算法, 随机梯度增进算法

Abstract:

【Objective】 This study attempts to improve the prediction accuracy of soil salinity in arid oasis by building models preferentially in the sub-area of oasis. At the same time, the difference and uncertainty of accuracy between global model and sub-region model are quantified. 【Method】 Therefore, to investigate the above differences, this study used two machine learning methods (Random Forest, RF and Stochastic Gradient Treeboost, SGT) to quantify the above effects and to prove the necessity of the building model in the sub-region compared with the full-sample model with respect to the simulation precision under the complex background of an arid region. Twenty-seven environmental scenarios (twelve original and fifteen derivatives) were designed based on the driving factors (land use and landform) and response factors (Normalized Difference Vegetation Index, NDVI and electrical conductivity, EC), which reflected variety of variabilities in soil salinity. After analyzing the results, the following preliminary conclusions were drawn. 【Result】 The simulation results from 70.37% (19/27) of the scenarios showed that the predicted value of soil salinity from SGT was closer to the observed value from RF. Ten original sub-regions were modeled individually and compared with the full-sample model under the oasis scale (according to the 10 partition rules to reclassify the simulated values), and the result showed that the prediction accuracy of the former 70% scenario was higher than that of the latter. In particular, the regions of EC≤4 dS·m ^-1 and 2 ddS·m ^-1＜EC＜16 dS·m ^-1 should be modeled separately to predict the spatial variability of regional salinity. By combining the predictions of sub-regions and comparing them with the predicted values of the full-sample model, the former (all four different combination modes) showed a higher prediction accuracy than the latter. In addition, this result also indicated that the preferred medium for partitioning the sub-regions was soil electrical conductivity, followed by landform and land use. 【Conclusion】 The study proposes to establish a soil salinity model based on SGT preferentially on different landscape scales within the oasis, and then combine the predicted values of each landscape scale to improve the prediction accuracy of oasis soil salinity.

Key words: soil salinity, machine learning, arid regions, Landsat OLI, spatial heterogeneity, Random Forest, Stochastic Gradient Treeboost

王飞,杨胜天,魏阳,杨晓东,丁建丽. 基于RF和SGT算法的子区优先建模对绿洲尺度土壤盐度预测精度的影响[J]. 中国农业科学, 2018, 51(24): 4659-4676.

WANG Fei,YANG ShengTian,WEI Yang,YANG XiaoDong,DING JianLi. Influence of Sub-Region Priority Modeling Constructed by Random Forest and Stochastic Gradient Treeboost on the Accuracy of Soil Salinity Prediction in Oasis Scale[J]. Scientia Agricultura Sinica, 2018, 51(24): 4659-4676.

0
/ / 推荐

导出引用管理器 EndNote|Reference Manager|ProCite|BibTeX|RefWorks

链接本文: https://www.chinaagrisci.com/CN/10.3864/j.issn.0578-1752.2018.24.007

https://www.chinaagrisci.com/CN/Y2018/V51/I24/4659

图/表 16

图1

图2

图3

表1

基于Landsat OLI（30 m）和DEM（8个空间分辨率）衍生的环境变量"

变量组 Environment variable group	指数 Index
波段及其衍生变量 Band & derivatives	全波段;缨帽变换（亮度（TC1）,绿度（TC2）,湿度（TC3）,主成分分析（前三个波段） Bands;Tasseled Cap (brightness,TC1;Greenness,TC2; Wetness,TC3 ), Principal Component Analysis(PC1,PC2,PC3)
植被指数 Vegetation indices	归一化植被指数;土壤调节植被指数;增强植被指数;广义植被指数;冠层响应盐度指数^[3];比值植被指数;两波段增强植被指数^[30];扩展的归一化植被指数^[31];扩展的增强植被指数^[31] Normalized Difference Vegetation Index, NDVI; Soil Adjusted Vegetation Index, SAVI; Enhanced Vegetation Index, EVI; Generalized Difference Vegetation Index, GDVI^[5]; Canopy Response Salinity Index, CRSI^[3]; Simple Ratio Vegetation Index, SR; Two-Band Enhanced Vegetation Index, EVI2^[30]; Extended NDVI, ENDVI^[31]; Extended EVI, EEVI^[31]
土壤相关指数 Soil-related indices	盐度指数（Salinity Index, SIT）^[1];盐度指数（Salinity Index, SI）^[1];盐度指数（Salinity Index（SI1）^[1];盐度指数（Salinity Index, SI2）^[1];盐度指数（Salinity Index, SI3）^[1];盐度指数（Salinity Index, SIA）^[1];盐度指数（Salinity Index, SIB）;盐度比值指数（Salinity Ratio, SAIO）^[32];黏土指数（Clay Index, CLEX）^[4] ;石膏指数（Gypsum Index, GYEX）^[4];亮度指数（Brightness Index, BREX）^[4];碳酸盐指数（Carbonate Index, CAEX）^[4];FSEN= (B5-B7）/(B5+B7）^[33];颜色指数（Color Indices -色相Hue, 饱和度 Saturation, 色调Value）;归一化热感指数（Normalized Difference Infrared Index, NDII）^[4];全球植被湿度指数（Global Vegetation Moisture Index, GVMI）^[34];地表温度（Temperature）
DEM衍生因子 Dem derivatives	河道相关：谷深（Valley Depth ,VD）;与河网的垂直距离（Vertical Distance To Channel Network ,VDCN）;水文学相关：坡长因子（LS-Factor, LSF）;地形指数指数（Topographic Wetness Index, TWI）;坡长（Slope Length, SL）;通视度（Sky View Factor, SVF）;地形位置指数（Topographic Position Index, TPI）;多尺度谷底平整度（Multiresolution Index Of Valley Bottom Flatness（MRVBF And MRRTF））,坡高（Slope Height, SH）,归一化高度（Normalized Height, NH）,标准化高度（Standardized Height, STH）,坡度中值位置Mid-Slope Position（MSP）,地表纹理（Terrain Surface Texture, TEX）,汇流累积量（Flow Accumulation, FA）;形态：截面曲率（Cross-Section Curvature, CSC）,纵向曲率（Longitudinal Curvature, LC）,相对坡度位置（Relative Slope Position, RSP）,流域坡度（Catchment Slope, CS）（SAGA Development Team, 2011）

表1

图4

图5

图6

图7

表2

表3

表4

表5

图8

图9

图10

图11

参考文献 43

[1]	ALLBED A, KUMAR L, ALDAKHEEL Y Y . Assessing soil salinity using soil salinity and vegetation indices derived from IKONOS high-spatial resolution imageries: Applications in a date palm dominated region. Geoderma, 2014,230:1-8. doi: 10.1016/j.geoderma.2014.03.025
[2]	SCUDIERO E, SKAGGS T H, CORWIN D L . Regional scale soil salinity evaluation using Landsat 7, western San Joaquin Valley, California, USA.Geoderma Regional, 2014(2/3):82-90. doi: 10.1016/j.geodrs.2014.10.004
[3]	SCUDIERO E, SKAGGS T H, CORWIN D L . Regional-scale soil salinity assessment using Landsat ETM + canopy reflectance. Remote Sensing of Environment, 2015,169:335-343. doi: 10.1016/j.rse.2015.08.026
[4]	TAGHIZADEH-MEHRJARDI R, MINASNY B, SARMADIAN F, MALONE B P . Digital mapping of soil salinity in Ardakan region, central Iran. Geoderma, 2014,213:15-28. doi: 10.1016/j.geoderma.2013.07.020
[5]	WU W, MHAIMEED A S , AL-SHAFIE W M,ZIADAT F,DHEHIBI B,NANGIA V,PAUW E D.Mapping soil salinity changes using remote sensing in Central Iraq. Geoderma Regional, 2014(2/3):21-31. doi: 10.1016/j.geodrs.2014.09.002
[6]	YAHIAOUI I, DOUAOUI A, QIANG Z, ZIANE A . Soil salinity prediction in the Lower Cheliff plain(Algeria) based on remote sensing and topographic feature analysis. Journal of Arid Land, 2015,7(6):794-805. doi: 10.1007/s40333-015-0053-9
[7]	ZHANG T T, QI J G, GAO Y, OUYANG Z T, ZENG S L, ZHAO B . Detecting soil salinity with MODIS time series VI data. Ecological Indicators, 2015,52:480-489. doi: 10.1016/j.ecolind.2015.01.004
[8]	ZHANG T-T, ZENG S-L, GAO Y, OUYANG Z-T, LI B, FANG C-M, ZHAO B . Using hyperspectral vegetation indices as a proxy to monitor soil salinity. Ecological Indicators, 2011,11(6):1552-1562. doi: 10.1016/j.ecolind.2011.03.025
[9]	GARCíA M, OYONARTE C, VILLAGARCíA L, CONTRERAS S, DOMINGO F, PUIGDEFáBREGAS J . Monitoring land degradation risk using ASTER data: The non-evaporative fraction as an indicator of ecosystem function. Remote Sensing of Environment, 2008,112(9):3720-3736. doi: 10.1016/j.rse.2008.05.011
[10]	LI X, YANG K, ZHOU Y. Progress in the study of oasis-desert interactions.Agricultural and Forest Meteorology 2016,230-231:1-7. DOI: 10.1016/j.agrformet.2016.08.022 doi: 10.1016/j.agrformet.2016.08.022
[11]	GONG L, RAN Q, HE G, TIYIP T . A soil quality assessment under different land use types in Keriya river basin, Southern Xinjiang, China. Soil and Tillage Research, 2015,146:223-229. doi: 10.1016/j.still.2014.11.001
[12]	TUTEJA N K, BEALE G, DAWES W, VAZE J . Predicting the effects of landuse change on water and salt balance-a case study of a catchment affected by dryland salinity in NSW, Australia. Journal of Hydrology, 2003,283(1):67-90. doi: 10.1016/S0022-1694(03)00236-1
[13]	WANG Y, LI Y . Land exploitation resulting in soil salinization in a desert-oasis ecotone. Catena, 2013,100:50-56. doi: 10.1016/j.catena.2012.08.005
[14]	HENGL T , MENDES D J J, HEUVELINK G B, RUIPEREZ G M, KILIBARDA M, BLAGOTIĆ A, SHANGGUAN W, WRIGHT M N, GENG X, BAUERMARSCHALLINGER B. SoilGrids250m: Global gridded soil information based on machine learning. Plos One, 2017,12(2):e0169748. doi: 10.1371/journal.pone.0169748 pmid: 5313206
[15]	SCHILLACI C, LOMBARDO L, SAIA S, FANTAPPIè M ,MäRKER M,ACUTIS M. Modelling the topsoil carbon stock of agricultural lands with the Stochastic Gradient Treeboost in a semi-arid Mediterranean region. Geoderma, 2017,286:35-45. doi: 10.1016/j.geoderma.2016.10.019
[16]	HASTIE T, TIBSHIRANI R, FRIEDMAN J H, FRANKLIN J . The elements of statistical learning, second edition: data mining, inference, and prediction. Mathematical Intelligencer, 2009,27(2):83-85.
[17]	ANGILERI S E, CONOSCENTI C, HOCHSCHILD V ,MäRKER M, ROTIGLIANO E, AGNESI V. Water erosion susceptibility mapping by applying Stochastic Gradient Treeboost to the Imera Meridionale River Basin (Sicily, Italy). Geomorphology, 2016,262:61-76. doi: 10.1016/j.geomorph.2016.03.018
[18]	HEUNG B, BULMER C E, SCHMIDT M G . Predictive soil parent material mapping at a regional-scale: A Random Forest approach. Geoderma, 2014,214:141-154. doi: 10.1016/j.geoderma.2013.09.016
[19]	HEUNG B, HO H C, ZHANG J, KNUDBY A, BULMER C E, SCHMIDT M G . An overview and comparison of machine-learning techniques for classification purposes in digital soil mapping. Geoderma, 2016,265:62-77. doi: 10.1016/j.geoderma.2015.11.014
[20]	LIEß M, GLASER B, HUWE B . Uncertainty in the spatial prediction of soil texture: Comparison of regression tree and Random Forest models. Geoderma, 2012,170:70-79. doi: 10.1016/j.geoderma.2011.10.010
[21]	BREIMAN L. . Random Forests Machine Learning, 2001,45(1):5-32.
[22]	FRIEDMAN J H. Stochastic Gradient Boosting. Computational Statistics & Data Analysis, 2002,38(4):367-378.
[23]	DING J, YU D. Monitoring and evaluating spatial variability of soil salinity in dry and wet seasons in the Werigan-Kuqa Oasis,China,using remote sensing and electromagnetic induction instruments.. Geoderma ,2014,235 236:316-322. doi: 10.1016/j.geoderma.2014.07.028
[24]	BENNETT S J, BARRETTLENNARD E G, COLMER T D . Salinity and waterlogging as constraints to saltland pasture production: a review. Agriculture Ecosystems & Environment, 2009,129(4):349-360. doi: 10.1016/j.agee.2008.10.013
[25]	WANG F, CHEN X, LUO G, HAN Q . Mapping of regional soil salinities in Xinjiang and strategies for amelioration and management. Chinese Geographical Science, 2015,25(3):321-336. doi: 10.1007/s11769-014-0718-x
[26]	MINASNY B, MCBRATNEY A B . A conditioned Latin hypercube method for sampling in the presence of ancillary information Computers & Geosciences, 2006,32(9):1378-1388. doi: 10.1016/j.cageo.2005.12.009
[27]	鲁如坤 . 土壤农业化学分析方法.北京: 中国农业科技出版社, 1999.
	LU R K. Methods for Soil Agrochemistry Analysis. Beijing: China Agricultural Science and Technology Press, 1999. (in Chinese).
[28]	SONG W, MU X, RUAN G, GAO Z, LI L, YAN G . Estimating fractional vegetation cover and the vegetation index of bare soil and highly dense vegetation with a physically based method. International Journal of Applied Earth Observations & Geoinformation, 2017,58:168-176. doi: 10.1016/j.jag.2017.01.015
[29]	RICHARDS L A . Diagnosis and Improvement of Saline and Alkali Soils. Usda Agriculture Handbook, 1954,60(3):290.
[30]	MONDAL P . Quantifying surface gradients with a 2-band Enhanced Vegetation Index (EVI2). Ecological Indicators, 2011,11(3):918-924. doi: 10.1016/j.ecolind.2010.10.006
[31]	陈红艳, 赵庚星, 陈敬春, 王瑞燕, 高明秀 . 基于改进植被指数的黄河口区盐渍土盐分遥感反演. 农业工程学报, 2015,31(5):107-114.
	CHEN H Y, ZHAO G X, CHEN J C, WANG R Y, GAO M X . Remote sensing inversion of saline soil salinity based on modified vegetation index in estuary area of Yellow River. Transactions of the Chinese Society of Agricultural Engineering, 2015,31(5):107-114. (in Chinese)
[32]	METTERNICHT G I, ZINCK J A . Remote sensing of soil salinity: potentials and constraints. Remote Sensing of Environment, 2003,85(1):1-20. doi: 10.1016/S0034-4257(02)00188-8
[33]	YU R, LIU T, XU Y, ZHU C, ZHANG Q, QU Z, LIU X, LI C . Analysis of salinization dynamics by remote sensing in Hetao Irrigation District of North China. Agricultural Water Management, 2010,97(12):1952-1960. doi: 10.1016/j.agwat.2010.03.009
[34]	CECCATO P, GOBRON N, FLASSE S, PINTY B, TARANTOLA S . Designing a spectral index to estimate vegetation water content from remote sensing data: Part 1: Theoretical approach. Remote Sensing of Environment, 2002,82(2):188-197. doi: 10.1016/S0034-4257(02)00037-8
[35]	LIAW A, WIENER M . Classification and Regression by RandomForest.R News, 2002(2/3):18-22.
[36]	ELITH J, LEATHWICK J R, HASTIE T . A working guide to boosted regression trees. Journal of Animal Ecology, 2008,77(4):802-813. doi: 10.1111/j.1365-2656.2008.01390.x pmid: 18397250
[37]	SVETNIK V, LIAW A, TONG C, CULBERSON J C, SHERIDAN R P, FEUSTON B P . Random Forest:A Classification and Regression Tool for Compound Classification and QSAR Modeling. Journal of Chemical Information and Computer Sciences, 2003,43(6):1947-1958. doi: 10.1021/ci034160g pmid: 14632445
[38]	NAGHIBI S A, POURGHASEMI H R, DIXON B . GIS-based groundwater potential mapping using boosted regression tree, classification and regression tree, and random forest machine learning models in Iran. Environmental Monitoring & Assessment, 2016,188(1):44. doi: 10.1007/s10661-015-5049-6 pmid: 26687087
[39]	YOUSSEF A M, POURGHASEMI H R, POURTAGHI Z S ,AL-KATHEERI M M.Landslide susceptibility mapping using random forest, boosted regression tree, classification and regression tree, and general linear models and comparison of their performance at Wadi Tayyah Basin, Asir Region, Saudi Arabia. Landslides, 2016,13(5):839-856.
[40]	YANG R M, ZHANG G L, LIU F, LU Y Y, YANG F, YANG F, YANG M, ZHAO Y G, LI D C . Comparison of boosted regression tree and random forest models for mapping topsoil organic carbon concentration in an alpine ecosystem. Ecological Indicators, 2016,60:870-878. doi: 10.1016/j.ecolind.2015.08.036
[41]	BLANCO C M G, GOMEZ V M B, CRESPO P, LIEß M . Spatial prediction of soil water retention in a Páramo landscape: Methodological insight into machine learning using random forest. Geoderma, 2018,316:100-114.
[42]	MULDER V L, LACOSTE M , RICHER-DE-FORGES A C, MARTIN M P, ARROUAYS D . National versus global modelling the 3D distribution of soil organic carbon in mainland France. Geoderma, 2016,263:16-34. doi: 10.1016/j.geoderma.2015.08.035
[43]	王玉刚, 李彦, 肖笃宁 . 土地利用对天山北麓土壤盐渍化的影响. 水土保持学报, 2009,23(5):179-183.
	WANG Y G, LI Y, XIAO D N . Effects of land use type on soil salinization at Northern Slope of Tianshan Mountain. Journal of Soil and Water Conservation, 2009,23(5):179-183. (in Chinese)

情景 Scenario	最小值 Min	最大值 Max	平均值 Average	变异系数 CV	Q25	Q50	Q75
S1	0.14	184.5	31.32	1.29	1.42	13.11	48.00
S2	0.14	128.50	11.25	2.02	0.42	1.50	9.90
S3	0.15	184.50	47.14	0.95	11.84	26.00	79.95
S4	0.15	103.9	5.57	2.97	0.38	0.90	3.36
S5	0.14	184.50	47.33	0.95	11.59	26.00	88.00
S6	0.23	150.50	34.38	1.10	6.76	17.33	60.90
S7	0.27	184.50	50.40	0.92	11.77	35.40	78.82
S8	0.14	147.40	28.20	1.37	1.34	10.69	33.89
S9	0.14	3.99	0.99	0.95	0.33	0.55	1.41
S10	2.01	15.92	8.47	0.50	4.54	7.97	12.04
S11	4.06	184.50	46.43	0.90	13.11	25.70	75.60
S12	16.28	184.5	64.10	0.63	24.85	61.10	94.15

情景 Scenario	随机森林 Random Forest			随机梯度增进算法 Stochastic Gradient Treeboost
情景 Scenario	R²	RMSE	RPD	R²	RMSE	RPD
S2	0.33**	18.52	1.23	0.35**	18.11	1.26
S3	0.34**	36.487	1.23	0.34**	36.485	1.23
S4	0.39**	35.36	1.28	0.41**	34.84	1.30
S5	0.47**	27.36	1.39	0.58**	24.15	1.58
S6	0.50**	33.15	1.17	0.48**	27.80	1.39
S7	0.33**	37.88	1.23	0.50**	33.15	1.40
S8	0.29**	0.79	1.20	0.23**	0.83	1.14
S9	0.51**	2.94	1.44	0.38**	3.31	1.28
S10	0.32**	34.57	1.22	0.38**	33.26	1.27
S11	0.30**	34.08	1.20	0.35**	32.99	1.24

情景 Scenario	随机森林 Random Forest			随机梯度增进算法 Stochastic Gradient Treeboost
情景 Scenario	R²	RMSE	RPD	R²	RMSE	RPD
S1	0.46**	29.65	1.37	0.48**	29.17	1.39
S2	0.30**	19.20	1.18	0.44**	16.97	1.34
S3	0.38**	35.50	1.27	0.37**	35.71	1.26
S4	0.35**	36.46	1.24	0.34**	36.67	1.24
S5	0.51**	27.94	1.36	0.54**	26.06	1.46
S6	0.46**	28.42	1.36	0.50**	27.31	1.42
S7	0.39**	36.78	1.26	0.36**	37.09	1.25
S8	ns	15.19	0.06	ns	18.52	0.05
S9	0.20**	21.68	0.20	0.18**	21.68	0.20
S10	0.37**	33.73	1.25	0.39**	27.17	1.27
S11	0.28**	37.88	1.08	0.31**	37.12	1.10

情景合并 Scenario combination	随机森林 Random Forest			随机梯度增进算法 Stochastic Gradient Treeboost
情景合并 Scenario combination	R²	RMSE	RPD	R²	RMSE	RPD
S2 & S 3	0.49**	29.73	1.37	0.47**	29.54	1.38
S9 & S11	0.51**	28.25	1.43	0.55**	27.23	1.48
S5 & S6	0.43**	32.86	1.32	0.47**	31.63	1.37
S5 & S6 (RSOSM)	0.40**	33.77	1.29	0.40**	33.47	1.30
S7 & S8	0.43**	32.25	1.32	0.51**	30.01	1.43
S7 & S8 (RSOSM)	0.46**	31.55	1.36	0.47**	31.03	1.38