Journal of Geo-information Science >
Accuracy Assessment Strategy based on Pseudo-pure Pixels and its Application
Received date: 2022-01-25
Revised date: 2022-02-22
Online published: 2022-10-25
Supported by
The Second Tibetan Plateau Scientific Expedition and Research Program(2019QZKK0402)
National Natural Science Foundation of China(41971239)
National Natural Science Foundation of China(41461017)
Scientific Research Foundation of Yunnan Provincial Education Department(2022Y061)
Land cover is an important parameter in geoscience research, and the accuracy assessment of land cover products is a prerequisite to ensure reasonable application of land cover products. In this study, an accuracy assessment strategy based on pseudo-pure pixels (i.e., pseudo-pure pixel strategy) is proposed. That is, calculating the area of land cover types of the high-resolution pixels within a pixel window of coarse spatial resolution data and defining the land cover type with the largest area as the advantage type, and then generating pseudo-pure pixels of advantage type based on the pixel window when the area proportion of the advantage type is higher than a pseudo-pure pixel purity threshold (ranging from 35% to 100%, with a step length of 5%). We take the Lancang-Mekong (Lanmei) basin as the study area and select the GlobeLand30 as the reference data. The confusion matrix accuracy assessment method was used to compare the difference in the accuracy of two sets of global land cover data, i.e., CCI-LC (300 m) and MCD12Q1 (500 m), using different assessment methods, i.e., the pseudo-pure pixel strategy and the resampling method (Nearest and Majority). Our results show that: (1) The accuracy using the pseudo-pure pixel strategy for CCI-LC and MCD12Q1 in the Lanmei Basin under the purity thresholds of 35%~100% are 72.76%~55.26% and 71.44%~45.41%, respectively, and it can better reflect the influence of pixel purity on the accuracy of land cover data than the single accuracy obtained by resampling method (Nearest: 71.21% and 70.54%; Majority: 71.48 and 69.87%); (2) The overall accuracy of CCI-LC is higher than that of MCD12Q1. The accuracy difference of the two datasets increases with the increase of purity threshold, and it is 1.32% and 9.85% respectively for purity threshold of 35% and 100%, respectively; (3) In both datasets, the classification accuracy of cropland, forest, grassland, and water is relatively high, and the classification accuracy of shrub land and bare land is relatively low; (4) The spatial inconsistency between the two datasets and the GlobeLand30 mainly occur in mixed pixel regions with highly heterogeneous land cover types. And the assessment sample grids are purer with the increase of pseudo-pure pixel threshold, which reduces the effects of mixed pixels on accuracy assessment. The pseudo-pure pixel strategy has the potential to compare the mapping accuracy of land cover data with different spatial resolutions, and provides a promising validation method for determining the applicability and application scope of global land cover products at regional scales.
XU Xiao , LI Yating , FAN Hui . Accuracy Assessment Strategy based on Pseudo-pure Pixels and its Application[J]. Journal of Geo-information Science, 2022 , 24(8) : 1617 -1630 . DOI: 10.12082/dqxxkx.2022.220048
图2 澜沧江-湄公河流域区位及地形注:水文流域边界来源于HydroSHEDS数据集,由FAO GeoNetwork(http://www.fao.org/geonetwork)提供;数字高程模型数据来源于SRTM 90 m数据库(http://srtm.csi.cgiar.org);本图基于自然资源部标准地图服务网站中审图号为GS(2021)5443号的标准地图绘制,底图无修改。 Fig. 2 Location and topography of the Lancang-Mekong River Basin |
表1 3套全球土地覆被产品参数Tab. 1 Product parameters of three global land cover datasets |
数据集 | GlobleLand30 | CCI-LC | MCD12Q1 |
---|---|---|---|
传感器 | Landsat TM5/ETM+、HJ-1 | ENVISAT MERIS、SPOT | Terra MODIS |
分类方法 | 像元-对象-知识(POK方法) | 非监督分类 | 监督分类/决策树/神经网络 |
分辨率/m | 30 | 300 | 500 |
时相 | 2000年、2010年 | 1992—2020年 | 2001—2015年 |
分类体系 | 自分类体系 (10类) | LCCS (22类) | IGBP (17类) |
总体精度/% | 83.50 | 74.40 | 75.00 |
验证方法 | 交叉验证 | 样本验证 | 交叉验证 |
表2 不同土地覆被分类系统间的类别对应关系Tab. 2 Corresponding relationship of land cover classes among the three global land cover datasets |
编号 | 类别 | GlobleLand30 | CCI-LC | MCD12Q1 |
---|---|---|---|---|
1 | 耕地 | 10用于种植农作物的土地,包括水田、灌溉旱地、大棚用地,以及果园茶园等灌木类经济作物种植地 | 10/11/12农田旱地/草本覆被的耕地/树或灌丛覆被的耕地 20灌溉旱地或洪泛耕地 30耕地(>50%)与树、灌、草等自然植被(<50%)的混交地 | 12农用地 14农用地与自然植被混合:40%~60%为天然乔木、灌木林地或草本植被 |
2 | 有林地 | 20树冠覆被度>30%的土地,包括落叶阔叶林、常绿阔叶林、落叶针叶林、常绿针叶林、混交林,以及树冠覆被度为10%~30%的疏林地 | 50常绿阔叶林(>15%) 60/61/62落叶阔叶林(>15%)/(>40%)/(15%~40%) 70/71/72常绿针叶林(>15%)/(>40%)/(15%~40%) 80/81/82落叶针叶林(>15%)/(>40%)/(15%~40%) 90针阔混交林 100乔木和灌木林地(>50%)与草本植物(<50%)的混交地 | 1常绿针叶林:冠幅>4 m2,覆被度>60% 2常绿阔叶林:冠幅>4 m2,覆被度>60% 3落叶针叶林:冠幅>4 m2,覆被度>60% 4落叶阔叶林:冠幅>4 m2,覆被度>60% 5针阔混交林:落叶常绿为主,冠幅>4 m2,覆被度>60% 8木本稀树草原:冠幅>4 m2,覆被度30%~60% |
3 | 草地 | 30天然草本覆被度>10%的土地,包括草原、草甸、稀疏草原、荒漠草原、人工草地 | 110草本植被(>50%)与树(<50%)的混交地 130草地 | 10草地:以年生草本植物为主,高度<2 m |
4 | 灌木林地 | 40灌木覆被度>30%的土地,包括山地灌丛、落叶和常绿灌丛,以及荒漠地区覆被度>10%的荒漠灌丛 | 120/121/122灌木林地/常绿灌木林地/落叶灌木林地 40树、灌、草等自然植被(>50%)与耕地(<50%)的混交地 | 6稠密灌丛:以木本多年生植物为主,高度1~2 m,覆被度>60% 7稀疏灌丛:以木本多年生植物为主,高度1~2 m,覆被度10%~60% 9稀树草原:树木覆被度10-30%,冠幅>4 m2 |
5 | 湿地 | 50有浅层积水或土壤过湿的土地,包括内陆沼泽、湖泊沼泽、河流洪泛湿地、森林/灌木湿地、泥炭沼泽、盐沼等 | 160被水淹没的有林地(淡水或咸水) 170被水淹没的有林地(海水) 180被水淹没的灌木林地和草本植物(淡水、海水或咸水) | 11永久湿地:永久淹水面积30%~60%,植被覆被度>10%的土地 |
6 | 水体 | 60 陆地范围液态水覆被的区域 | 210水体 | 17水体:至少60%的区域被永久水体覆被 |
7 | 建设用地 | 80由人工建造活动形成的地表 | 190城镇地区 | 13城市和建筑区:至少30%的不透水表面,包括建筑材料、沥青道路 |
8 | 裸地 | 90植被覆被度<10%的自然覆被土地,包括荒漠、沙地、砾石地、裸岩、盐碱地等 | 200/201/202裸地/坚固的裸地/松散的裸地 150/152/153树、灌、草等稀疏植被(<15%) | 16裸地:至少60%的区域是没有植被覆被的裸露地区(沙、岩石、土壤),植被覆被<10% |
9 | 永久冰雪 | 100由永久积雪、冰川和冰盖覆被的土地 | 220 永久冰雪 | 15雪和冰:全年至少有60%的地区被冰雪覆被10个月 |
图3 分类系统规整后的3套澜湄流域土地覆被图Fig. 3 Maps of land covers in the Lancang-Mekong River basin extracted by the three global land cover datasets with a unified land cover classification system |
[1] |
|
[2] |
|
[3] |
吴英迪, 蒙吉军. 中国自然资源生态服务重要性评价与空间格局分析[J]. 自然资源学报, 2022, 37(1):17-33.
[
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
刘琼欢, 张镱锂, 刘林山, 等. 七套土地覆被数据在羌塘高原的精度评价[J]. 地理研究, 2017, 36(11):2061-2074.
[
|
[9] |
|
[10] |
|
[11] |
|
[12] |
马京振, 孙群, 肖强, 等. 河南省GlobeLand30数据精度评价及对比分析[J]. 地球信息科学学报, 2016, 18(11):1563-1572.
[
|
[13] |
|
[14] |
|
[15] |
|
[16] |
陈逸聪, 邵华, 李杨. 多源土地覆被产品在长三角地区的一致性分析与精度评价[J]. 农业工程学报, 2021, 37(6):142-150.
[
|
[17] |
|
[18] |
张永梅, 滑瑞敏, 马健喆, 等. 基于深度学习与超分辨率重建的遥感高时空融合方法[J]. 计算机工程与科学, 2020, 42(9):1578-1586.
[
|
[19] |
张周威, 余涛, 孟庆岩, 等. 空间重采样方法对遥感影像信息影响研究[J]. 华中师范大学学报(自然科学版), 2013, 47(3):426-430.
[
|
[20] |
赵英时. 遥感应用分析原理与方法[M]. 北京: 科学出版社, 2003:34-35.
[
|
[21] |
|
[22] |
MRC. Overview of the hydrology of the Mekong Basin[R]. Vientiane: Mekong River commission, 2005:73-75.
|
[23] |
何大明. 澜沧江──湄公河水文特征分析[J]. 云南地理环境研究, 1995,(1):58-74.
[
|
[24] |
陈丽晖, 曾尊固, 何大明. 国际河流流域开发中的利益冲突及其关系协调——以澜沧江-湄公河为例[J]. 世界地理研究, 2003, 12(1):71-78.
[
|
[25] |
|
[26] |
ESA. Land Cover CCI: Product User Guide Version 2 (2017). Available online: https://www.esa-landcover-cci.org/?q=webfm_send/84.
|
[27] |
|
[28] |
|
[29] |
陈军, 廖安平, 陈晋, 等. 全球30m地表覆盖遥感数据产品-Globe Land30[J]. 地理信息世界, 2017, 24(1):1-8.
[
|
[30] |
黄亚博, 廖顺宝. 首套全球30m分辨率土地覆被产品区域尺度精度评价——以河南省为例[J]. 地理研究, 2016, 35(8):1433-1446.
[
|
[31] |
|
[32] |
姚远, 陈曦, 钱静. 定量遥感尺度转换方法研究进展[J]. 地理科学, 2019, 39(3):367-376.
[
|
[33] |
栾海军, 田庆久, 章欣欣, 等. 定量遥感地表参数尺度转换研究趋势探讨[J]. 地球科学进展, 2018, 33(5):483-492.
[
|
[34] |
郝大磊, 肖青, 闻建光, 等. 定量遥感升尺度转换方法研究进展[J]. 遥感学报, 2018, 22(3):408-423.
[
|
[35] |
|
[36] |
周觅, 张杰林. 遥感影像尺度转换及最优尺度选择探讨[J]. 世界核地质科学, 2011, 28(2):94-98.
[
|
[37] |
郭紫甜, 王春梅, 刘欣, 等. 基于小流域抽样单元的中国FROM-GLC30数据精度评价[J]. 地球信息科学学报, 2021, 23(3):524-535.
[
|
[38] |
廖顺宝, 葛乐玮, 王艳萍, 等. 利用地形参数提高土地覆被分类精度方法的改进[J]. 遥感信息, 2021, 36(3):10-16.
[
|
[39] |
|
[40] |
|
[41] |
许光明, 杨雅萍, 杨飞, 等. 多分辨率遥感土地覆被数据质量综合评价——以湖南省桃源县为例[J]. 地球信息科学学报, 2016, 18(4):553-563.
[
|
[42] |
|
[43] |
|
[44] |
|
[45] |
|
/
〈 | 〉 |