黄土高原区SRTM1 DEM高程误差校正模型构建及对比分析

doi:10.12082/dqxxkx.2023.220677

摘要/Abstract

摘要：

对SRTM1 DEM高程误差进行校正可有效提高其应用精度。以具有典型地貌特征的黄土高原作为研究区域，以ICESat-2/ATL08陆地高程作为参考数据，引入主流机器学习算法建立SRTM1高程误差与影响因子之间的关系模型对高程值进行校正；通过分析模型性能指标、误差频数分布、校正误差空间格局以及典型剖面误差分布，以此得到不同地貌类型区的高程误差校正模型适用性。实验结果表明：在平原、风沙丘陵和黄土塬地貌区随机森林模型高程校正效果最佳，平均绝对误差分别降低0.49、0.82和1.2 m，同时校正误差在空间分布上异常值较少，低起伏度的平原和风沙丘陵地貌区典型剖面误差与原误差较为贴合；山地区支持向量机模型适用性更强，均方根误差和平均绝对误差分布降低了6.79 m和5.43 m，可大幅提升误差绝对值较小的点位频数，同时在空间格局和典型剖面验证效果最佳；黄土丘陵地貌区弹性反馈神经网络模型效果最优，均方根误差和平均绝对误差分别降低了2.3 m和2.04 m，空间分布上误差降低效果显著，典型剖面误差异常值较少；土石丘陵地貌区卷积神经网络模型效果更理想，均方根误差与平均绝对误差分别降低4.14 m和3.5 m，空间分布与地形剖面误差降低效果良好。研究结果可为该区域不同地貌类型区选用SRTM1 DEM高程校正模型提供了参考。

关键词: 数字高程模型, 误差模型, 机器学习, 随机森林, 支持向量机, 弹性反馈神经网络, 卷积神经网络, SRTM

Abstract:

Correcting the error of SRTM1 DEM can effectively improve its application accuracy. Taking the Loess Plateau with typical landform characteristics as the research area and using the ICESat-2/ATL08 land laser elevation point as the reference data, the mainstream machine learning algorithms are introduced to establish the relationship between SRTM1 DEM error and its influencing factors to correct the elevation value. By analyzing the model performance indicators, error frequency distribution curve, spatial pattern of correction error, and the comparison of original and corrected distributions of error in typical profile, we can obtain the applicability of elevation error correction models for different landform types. Results show that the elevation correction effect using the random forest regression model is the best in the plain landform, wind-sand landform, and loess tableland landform, and the average absolute errors are reduced by 0.49 m, 0.82 m and 1.2 m, respectively. The errors in typical profile of the low relief plain and wind-sand landforms are more consistent with the original errors. The support vector machine model in mountainous areas is more applicable, and the root mean square error and mean absolute error is reduced by 6.79 m and 5.43 m, respectively. The frequency of small absolute errors is increased in the spatial pattern and typical profile. The resilient backpropagation model is the best in the loess hilly landform area, and the root mean square error and mean absolute error are reduced by 2.3 m and 2.04 m, respectively. The error reduction effect is significant at the spatial scale, and there are few outliers of error in typical profile. The convolutional neural network model in the soil-rocky hilly landform area is more ideal compared to other models, and the root mean square error and the mean absolute error are reduced by 4.14 m and 3.5 m, respectively. The error reduction effect of terrain profile is significant. The study provides a reference for the selection of the SRTM1 DEM elevation correction model for different landform types in this region.

Key words: digital elevation model, error model, machine learning, random forest, support vector, resilient backpropagation, convolutional neural network, SRTM

黄帅元, 董有福, 李海鹏. 黄土高原区SRTM1 DEM高程误差校正模型构建及对比分析[J]. 地球信息科学学报, 2023, 25(3): 669-681.DOI:10.12082/dqxxkx.2023.220677

HUANG Shuaiyuan, DONG Youfu, LI Haipeng. Establishment and Comparative Analysis of SRTM1 DEM Error Correction Model in the Loess Plateau[J]. Journal of Geo-information Science, 2023, 25(3): 669-681.DOI:10.12082/dqxxkx.2023.220677

图/表 12

图1

表1

表2

图2

表3

表4

图3

表5

表6

图4

图5

图6

参考文献 51

[1]	Farr T G, Rosen P A, et al. The shuttle radar topography mission[J]. Reviews of Geophysics, 2007, 45(2): RG2004. DOI:10.1029/2005rg000183 doi: 10.1029/2005rg000183
[2]	Zandbergen P. Applications of shuttle radar topography mission elevation data[J]. Geography Compass, 2008, 2(5):1404-1431. DOI:10.1111/j.1749-8198.2008.00154.x. doi: 10.1111/j.1749-8198.2008.00154.x
[3]	Hengl T, Mendes de Jesus J, Heuvelink G B M, et al. SoilGrids250m: Global gridded soil information based on machine learning[J]. PLoS One, 2017, 12(2):e0169748. DOI: 10.1371/journal.pone.0169748 doi: 10.1371/journal.pone.0169748
[4]	Drăguţ L, Blaschke T. Automated classification of landform elements using object-based image analysis[J]. Geomorphology, 2006, 81(3/4):330-344. DOI:10.1016/j.geomorph.2006.04.013 doi: 10.1016/j.geomorph.2006.04.013
[5]	Sun G, Ranson K J, Kharuk V I, et al. Validation of surface height from shuttle radar topography mission using shuttle laser altimeter[J]. Remote Sensing of Environment, 2003, 88(4):401-411. DOI:10.1016/j.rse.2003.09.001 doi: 10.1016/j.rse.2003.09.001
[6]	陈学兄, 常庆瑞, 郭碧云, 等. 基于SRTM DEM数据的中国地形起伏度分析研究[J]. 应用基础与工程科学学报, 2013, 21(4):670-678.
	[Chen X X, Chang Q R, Guo B Y, et al. Analytical study of the relief amplitude in China based on SRTM DEM data[J]. Journal of Basic Science and Engineering, 2013, 21(4):670-678.] DOI:10.3969/j.issn.1005-0930.2013.04.009 doi: 10.3969/j.issn.1005-0930.2013.04.009
[7]	Grohman G, Kroenung G, Strebeck J. Filling SRTM voids: The delta surface fill method[J]. Photogrammetr. Eng. Remote Sens, 2006(72):213-216.
[8]	沈焕锋, 刘露, 岳林蔚, 等. 多源DEM融合的高差拟合神经网络方法[J]. 测绘学报, 2018, 47(6):854-863.
	[Shen H F, Liu L, Yue L W, et al. A multi-source DEM fusion method based on elevation difference fitting neural network[J]. Acta Geodaetica et Cartographica Sinica, 2018, 47(6):854-863.] DOI:10.11947/j.AGCS.2018.20180135 doi: 10.11947/j.AGCS.2018.20180135
[9]	Li S J, Hu G H, Cheng X H, et al. Integrating topographic knowledge into deep learning for the void-filling of digital elevation models[J]. Remote Sensing of Environment, 2022, 269:112818. DOI:10.1016/j.rse.2021.112818 doi: 10.1016/j.rse.2021.112818
[10]	Dong G S, Chen F, Ren P. Filling SRTM void data via conditional adversarial networks[C]// IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2018:7441-7443. DOI:10.1109/IGARSS.2018.8518992 doi: 10.1109/IGARSS.2018.8518992
[11]	Reuter H I, Nelson A, Jarvis A. An evaluation of void-filling interpolation methods for SRTM data[J]. International Journal of Geographical Information Science, 2007, 21(9):983-1008. DOI:10.1080/13658810601169899 doi: 10.1080/13658810601169899
[12]	胡海, 游涟, 胡鹏, 等. 数字高程模型内插方法的分析和选择[J]. 武汉大学学报(信息科学版), 2011, 36(1):82-85.
	[Hu H, You L, Hu P, et al. Research on the choice ofinterpolation methods for generating DEM[J]. Geomatics and Information Science of Wuhan University, 2011, 36(1):82-85.] DOI:10.13203/j.whugis2011.01.013 doi: 10.13203/j.whugis2011.01.013
[13]	Rodríguez E, Morris C S, Belz J E. A global assessment of the SRTM performance[J]. Photogrammetric Engineering & Remote Sensing, 2006, 72(3):249-260. DOI: 10.14358/pers.72.3.249 doi: 10.14358/pers.72.3.249
[14]	Mouratidis A, Briole P, Katsambalos K. SRTM 3'' DEM (versions 1, 2, 3, 4) validation by means of extensive kinematic GPS measurements: A case study from North Greece[J]. International Journal of Remote Sensing, 2010, 31(23):6205-6222. DOI:10.1080/01431160903401403 doi: 10.1080/01431160903401403
[15]	Suwandana E, Kawamura K, Sakuno Y, et al. Evaluation of aster gdem2 in comparison with gdem1, srtm Dem and topographic-map-derived Dem using inundation area analysis and rtk-dgps data[J]. Remote Sensing, 2012, 4(8):2419-2431. DOI: 10.3390/rs4082419. doi: 10.3390/rs4082419
[16]	Zhao G S, Xue H, Ling F. Assessment of ASTER GDEM performance by comparing with SRTM and ICESat/GLAS data in Central China[C] // 2010 18th International Conference on Geoinformatics. IEEE, 2010:1-5. DOI: 10.1109/GEOINFORMATICS.2010.5567970 doi: 10.1109/GEOINFORMATICS.2010.5567970
[17]	李鹏, 李振洪, 施闯, 等. 中国地区30 m分辨率SRTM质量评估[J]. 测绘通报, 2016(9):24-28.
	[Li P, Li Z H, Shi C, et al. Quality evaluation of 1 arc second version SRTM DEM in China[J]. Bulletin of Surveying and Mapping, 2016(9):24-28.] DOI:10.13474/j.cnki.11-2246.2016.0285 doi: 10.13474/j.cnki.11-2246.2016.0285
[18]	张朝忙, 刘庆生, 刘高焕, 等. 太湖地区SRTM DEM高程精度质量评价[J]. 测绘科学, 2014, 39(3):139-142,123.
	[Zhang C M, Liu Q S, Liu G H, et al. Evaluation of SRTM data elevation quality in Taihu drainage basin[J]. Science of Surveying and Mapping, 2014, 39(3):139-142.] DOI: 10.16251/j.cnki.1009-2307.2014.03.031 doi: 10.16251/j.cnki.1009-2307.2014.03.031
[19]	Grohmann C H. Evaluation of TanDEM-X DEMs on selected Brazilian sites: Comparison with SRTM, ASTER GDEM and ALOS AW3D30[J]. Remote Sensing of Environment, 2018, 212:121-133. DOI:10.1016/j.rse.2018.04.043 doi: 10.1016/j.rse.2018.04.043
[20]	Graf L, Moreno-de-las-Heras M, Ruiz M, et al. Accuracy assessment of digital terrain model dataset sources for hydrogeomorphological modelling in small Mediterranean Catchments[J]. Remote Sensing, 2018, 10(12): 2014. DOI: 10.3390/rs10122014 doi: 10.3390/rs10122014
[21]	赵尚民, 何维灿, 王莉. DEM数据在黄土高原典型地貌区的误差分布[J]. 测绘科学, 2016, 41(2):67-70.
	[Zhao S M, He W C, Wang L. Error distribution analysis of SRTM3DEM V4 data in the typical geomorphologic area of Loess Plateau[J]. Science of Surveying and Mapping, 2016, 41(2):67-70,102.] DOI:10.16251/j.cnki.1009-2307.2016.02.014 doi: 10.16251/j.cnki.1009-2307.2016.02.014
[22]	Smith B. Accuracy and resolution of shuttle radar topography mission data[J]. Geophysical Research Letters, 2003, 30(9):1467. DOI:10.1029/2002gl016643 doi: 10.1029/2002gl016643
[23]	Hirt C, Filmer M S, Featherstone W E. Comparison and validation of the recent freely available ASTER-GDEM ver1, SRTM ver4.1 and GEODATA DEM-9S ver3 digital elevation models over Australia[J]. Australian Journal of Earth Sciences, 2010, 57(3):337-347. DOI:10.1080/08120091003677553 doi: 10.1080/08120091003677553
[24]	Shortridge A, Messina J. Spatial structure and landscape associations of SRTM error[J]. Remote Sensing of Environment, 2011, 115(6):1576-1587. DOI:10.1016/j.rse.2011.02.017 doi: 10.1016/j.rse.2011.02.017
[25]	Miliaresis G C. The landcover impact on the aspect/slope accuracy dependence of the SRTM-1 elevation data for the Humboldt range[J]. Sensors (Basel, Switzerland), 2008, 8(5):3134-3149. DOI:10.3390/s8053134 doi: 10.3390/s8053134
[26]	Bhang K J, Schwartz F W, Braun A. Verification of the vertical error in C-band SRTM DEM using ICESat and landsat-7, otter tail County, MN[J]. IEEE Transactions on Geoscience and Remote Sensing, 2007, 45(1):36-44. DOI: 10.1109/TGRS.2006.885401 doi: 10.1109/TGRS.2006.885401
[27]	Braun A, Fotopoulos G. Assessment of SRTM, ICESat, and survey control monument elevations in Canada[J]. Photogrammetric Engineering & Remote Sensing, 2007, 73(12):1333-1342. DOI:10.14358/pers.73.12.1333 doi: 10.14358/pers.73.12.1333
[28]	Magruder L, Neuenschwander A, Klotz B. Digital terrain model elevation corrections using space-based imagery and ICESat-2 laser altimetry[J]. Remote Sensing of Environment, 2021, 264:112621. DOI: 10.1016/j.rse.2021.112621 doi: 10.1016/j.rse.2021.112621
[29]	O'Loughlin F E, Paiva R C D, Durand M, et al. A multi-sensor approach towards a global vegetation corrected SRTM DEM product[J]. Remote Sensing of Environment, 2016, 182:49-59. DOI:10.1016/j.rse.2016.04.018 doi: 10.1016/j.rse.2016.04.018
[30]	Bulatović V, Sušić Z, Ninkov T. Estimate of the ASTER GDEM regional systematic errors and their removal[J]. International Journal of Remote Sensing, 2012, 33(18):5915-5926. DOI:10.1080/01431161.2012.676744 doi: 10.1080/01431161.2012.676744
[31]	Tran T A, Raghavan V, Masumoto S, et al. A geomorphology-based approach for digital elevation model fusion - case study in Danang city, Vietnam[J]. Earth Surface Dynamics, 2014, 2(2):403-417. DOI:10.5194/esurf-2-403-2014 doi: 10.5194/esurf-2-403-2014
[32]	Pham, H T, Marshall L, Johnson F, et al. A method for combining SRTM DEM and ASTER GDEM2 to improve topography estimation in regions without reference data[J]. Remote Sensing of Environment, 2018, 210:229-241. DOI:10.1016/j.rse.2018.03.026 doi: 10.1016/j.rse.2018.03.026
[33]	秦臣臣, 陈传法, 杨娜, 等. 基于ICESat/GLAS的山东省SRTM与ASTER GDEM高程精度评价与修正[J]. 地球信息科学学报, 2020, 22(3):351-360. doi: 10.12082/dqxxkx.2020.190411
	[Qin C C, Chen C F, Yang N, et al. Elevation accuracy evaluation and correction of SRTM and ASTER GDEM in Shandong Province based on ICESat/GLAS[J]. Journal of Geo-information Science, 2020, 22(3):351-360.] DOI:10.12082/dqxxkx.2020.190411 doi: 10.12082/dqxxkx.2020.190411
[34]	Yue L W, Shen H F, Zhang L P, et al. High-quality seamless DEM generation blending SRTM-1, ASTER GDEM v2 and ICESat/GLAS observations[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2017, 123:20-34. DOI:10.1016/j.isprsjprs.2016.11.002 doi: 10.1016/j.isprsjprs.2016.11.002
[35]	张丽, 董有福, 厉鸣. 利用弹性反馈神经网络融合ASTER GDEM和SRTM1高程数据[J]. 武汉大学学报·信息科学版, 2022, 47(1):1-9.
	[Zhang L, Dong Y F, Li M. ASTER GDEM and SRTM1 elevation data integration with rprop neural network[J]. Geomatics and Information Science of Wuhan University, 2022, 47(1):1-9.] DOI:10.13203/j.whugis20200679 doi: 10.13203/j.whugis20200679
[36]	Kasi V, Yeditha P K, Rathinasamy M, et al. A novel method to improve vertical accuracy of CARTOSAT DEM using machine learning models[J]. Earth Science Informatics, 2020, 13(4):1139-1150. DOI:10.1007/s12145-020-00494-1 doi: 10.1007/s12145-020-00494-1
[37]	余婷婷, 董有福. 利用随机森林回归算法校正ASTER GDEM高程误差[J]. 武汉大学学报·信息科学版, 2021, 46(7):1098-1105.
	[Yu T T, Dong Y F. Correcting elevation error of ASTER GDEM using random forest regression algorithm[J]. Geomatics and Information Science of Wuhan University, 2021, 46(7):1098-1105.] DOI:10.13203/j.whugis20190245 doi: 10.13203/j.whugis20190245
[38]	NASA's Distributed Active Archive Centers[DB/OL]. https://search.earthdata.nasa.gov/.
[39]	Rabus B, Eineder M, Roth A, et al. The shuttle radar topography mission—a new class of digital elevation models acquired by spaceborne radar[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2003, 57(4):241-262. DOI:10.1016/S0924-2716(02)00124-7 doi: 10.1016/S0924-2716(02)00124-7
[40]	National Snow&Ice Data Center. Data at NSIDC[DB/OL]. https://nsidc.org/data/icesat-2/.
[41]	王密, 韦钰, 杨博, 等. ICESat-2/ATLAS全球高程控制点提取与分析[J]. 武汉大学学报·信息科学版, 2021, 46(2):184-192.
	[Wang M, Wei Y, Yang B, et al. Extraction and analysis of global elevation control points from ICESat-2/ATLAS data[J]. Geomatics and Information Science of Wuhan University, 2021, 46(2):184-192.] DOI:10.13203/j.whugis20200531 doi: 10.13203/j.whugis20200531
[42]	Tian X X, Shan J. Comprehensive evaluation of the ICESat-2 ATL08 terrain product[J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 59(10):8195-8209. DOI: 10.1109/TGRS.2021.3051086 doi: 10.1109/TGRS.2021.3051086
[43]	Zhu J, Yang P F, Li Y, et al. Accuracy assessment of ICESat-2 ATL08 terrain estimates: A case study in Spain[J]. Journal of Central South University, 2022, 29(1):226-238. DOI:10.1007/s11771-022-4896-x doi: 10.1007/s11771-022-4896-x
[44]	Globeland30官方网站,2020版30米全球土地覆盖数据集[DB/OL]. http://globallandcover.com/.
[45]	杨勤科. 黄土高原地区地理分区图(2000年). 地球系统科学数据共享平台黄土高原科学数据共享平台, 2013.
	[Yang Q K. Graphs of Geographical Division on the Loess Plateau (2000). Data Sharing Infrastructure of Earth System Science_Data Sharing Infrastructure of Loess Plateau, 2013.]
[46]	Srivastava N, Hinton G E, Krizhevsky A, et al. Dropout: A simple way to prevent neural networks from overfitting[J]. Journal of Machine Learning Research, 2014, 15(1):1929-1958.
[47]	Wang Y Y, Wu G L, Deng L, et al. Prediction of aboveground grassland biomass on the Loess Plateau, China, using a random forest algorithm[J]. Scientific Reports, 2017, 7:6940. DOI:10.1038/s41598-017-07197-6 doi: 10.1038/s41598-017-07197-6 pmid: 28761059
[48]	邢晓语, 杨秀春, 徐斌, 等. 基于随机森林算法的草原地上生物量遥感估算方法研究[J]. 地球信息科学学报, 2021, 23(7):1312-1324. doi: 10.12082/dqxxkx.2021.200605
	[Xing X Y, Yang X C, Xu B, et al. Remote sensing estimation of grassland aboveground biomass based on random forest[J]. Journal of Geo-Information Science, 2021, 23(7):1312-1324.] DOI:10.12082/dqxxkx.2021.200605 doi: 10.12082/dqxxkx.2021.200605
[49]	毛亚萍, 房世峰. 基于机器学习的参考作物蒸散量估算研究[J]. 地球信息科学学报, 2020, 22(8):1692-1701. doi: 10.12082/dqxxkx.2020.200085
	[Mao Y P, Fang S F. Research of reference evapotranspiration’s simulation based on machine learning[J]. Journal of Geo-Information Science, 2020, 22(8):1692-1701.] DOI: 10.12082/dqxxkx.2020.200085 doi: 10.12082/dqxxkx.2020.200085
[50]	周飞燕, 金林鹏, 董军. 卷积神经网络研究综述[J]. 计算机学报, 2017, 40(6):1229-1251.
	[Zhou F Y, Jin L P, Dong J. Review of convolutional neural network[J]. Chinese Journal of Computers, 2017, 40(6):1229-1251.] DOI: 10.11897/SP.J.1016.2017.01229 doi: 10.11897/SP.J.1016.2017.01229
[51]	厉鸣, 董有福, 张丽. 统计回归分析算法校正AW3D30高程误差适用性研究[J]. 测绘科学, 2022, 47(2):23-30.
	[Li M, Dong Y F, Zhang L. Research on the applicability of statistical regression analysis algorithm correcting AW3D30 elevation error[J]. Science of Surveying and Mapping, 2022, 47(2):23-30,61.] DOI:10.16251/j.cnki.1009-2307.2022.02.004 doi: 10.16251/j.cnki.1009-2307.2022.02.004

数据名称	发布时间	来源	分辨率	地理坐标系
SRTM GL1 v003	2015年7月	NASA数据档案中心 https://search.earthdata.nasa.gov/	30 m	WGS84
ICESat-2/ATL08	2018年10月至2021年7月	NASA冰雪数据中心 https://nsidc.org/data/icesat-2	沿轨100 m激光高程点	WGS84
GlobeLand30全球地表覆盖数据	2020年	Globallandcover官网 http://www.globallandcover.com	30 m	WGS84
黄土高原地区地理地貌分区图	2013年12月	国家地球系统科学数据中心 http://www.geodata.cn/	矢量数据	GCS_Krasovsky_1940

地形统计信息	整体	平原	风沙丘陵	黄土塬	山地	黄土丘陵	土石丘陵
样点数/个	108 387	17 335	24 817	6803	26 762	28 548	4092
高程差	4766	1719	2190	1964	4766	3602	1320
平均/最大坡度（°）	12.22/66.68	3.54/43.77	4.86/47.66	13.56/55.39	17.49/66.68	13.89/54.28	17.84/46.06
TRI范围	0~54.63	0~33.25	0~26.88	0~30.87	0~54.63	0.38~37.75	0.88~26.75
TPI范围	-26.00~30.75	-19.38~14.75	-13.25~14.50	-16.00~18.25	-26.00~30.75	-17.38~18.50	-17.88~20.00
平均绝对误差	7.70	3.49	4.37	9.65	10.96	9.58	12.89
误差标准差	10.48	4.89	5.48	12.03	13.53	11.49	13.79
误差均方根	11.65	5.55	6.46	13.65	15.04	12.89	16.59

模型	测试集整体	平原		风沙丘陵		黄土塬		山地		黄土丘陵		土石丘陵
模型	测试集整体	RMSE	差	RMSE	差	RMSE	差	RMSE	差	RMSE	差	RMSE	差
原始	12.08	7.37		9.04		14.18		15.15		13.59		17.05
RFR	11.15	7.51	-0.14	9.24	-0.2	12.84	1.34	13.62	1.53	11.92	1.67	14.39	2.66
SVR	11.19	10.51	-3.14	10.05	-1.01	11.06	3.12	8.36	6.79	11.74	1.85	13.33	3.72
RProp	11.33	7.34	0.03	9.09	-0.05	13.73	0.45	14.46	0.69	11.29	2.3	13.19	3.86
CNN	11.37	8.48	-1.11	9.70	-0.66	12.96	1.22	13.31	1.84	11.63	1.96	12.91	4.14

模型	测试集整体	平原		风沙丘陵		黄土塬		山地		黄土丘陵		土石丘陵
模型	测试集整体	MAE	差	MAE	差	MAE	差	MAE	差	MAE	差	MAE	差
原始	8.53	4.22		5.44		10.27		11.45		10.06		13.35
RFR	7.26	3.73	0.49	4.62	0.82	9.07	1.20	9.68	1.77	8.49	1.57	11.15	2.20
SVR	7.96	8.02	-3.80	7.58	-2.14	8.12	2.15	6.02	5.43	9.32	0.74	10.15	3.20
RProp	7.59	3.91	0.31	4.99	0.45	9.56	0.71	10.49	0.96	8.02	2.04	10.14	3.21
CNN	7.66	5.38	-1.16	5.68	-0.24	9.31	0.96	9.53	1.92	8.33	1.73	9.85	3.50

模型	平原		风沙丘陵		黄土塬		山地		黄土丘陵		土石丘陵
模型	RMSE	差值	RMSE	差值	RMSE	差值	RMSE	差值	RMSE	差值	RMSE	差值
原始	4.67		5.39		14.29		15.21		13.66		16.22
RFR	4.19	0.48	4.99	0.40	12.63	1.66	10.72	4.49	12.21	1.45	13.32	2.90
SVR	6.13	-1.46	6.53	-1.14	12.69	1.60	9.08	6.13	12.75	0.91	12.64	3.58
RProp	4.38	0.29	5.50	-0.11	13.17	1.12	11.92	3.29	11.03	2.63	12.57	3.65
CNN	5.79	-1.12	5.56	-0.17	13.06	1.23	12.02	3.19	11.05	2.61	12.61	3.61