基于置信学习互导框架的小样本条件下森林扰动类型遥感分类
doi: 10.11728/cjss2025.02.2024-0146 cstr: 32142.14.cjss.2024-0146
-
严燕 女, 1998年10月出生, 现为中国地质大学硕士研究生, 主要研究方向为机器学习、时空数据挖掘与森林扰动分类. E-mail: 2004220026@email.cugb.edu.cn -
叶昕 男, 1992年5月出生, 现为中国农业大学信息与电气工程学院副教授, 硕士生导师, 主要研究方向为热红外遥感与深度学习. E-mail: yexin@cau.edu.cn
作者简介:
通讯作者:
Forest Disturbance Attribution under Small Sample Conditions Based on Confidence Learning Mutual Guidance Framework
-
摘要: 遥感技术作为地球系统和空间信息科学研究的核心手段, 提高了人类对地球系统变化的理解, 而开展地表重要组成部分——森林生态系统的扰动监测与分类研究可更精准高效. 但受限于样本标注繁重和变化稀疏限制标注数量. 本研究提出基于置信学习互导框架的小样本条件下森林扰动类型分类方法. 利用2000－2021年Landsat数据与连续变化检测与分类(Continuous Change Detection and Classification, CCDC)算法检测森林扰动获得大量未标注数据, 结合少量样本及基于随机森林(Random Forest, RF)和梯度提升(Categorical Boosting, CatBoost)分类器构建的互导框架, 以迭代方式通过置信学习从未标注数据中筛选高置信度数据, 扩充对方标注样本集, 互相指导分类, 进而提升分类精度. 结果表明, 该方法总体分类精度达91.4%, 较单一分类器提升5%, 在小样本条件下表现出优异性能, 为森林扰动类型分类研究提供了高效、可靠的解决方案.Abstract: As the core technical means for human beings to carry out scientific research on the Earth system and the application of spatial information in multiple fields, remote sensing technology has comprehensively improved human beings’ understanding of the complex process of changes in the Earth system, and the use of remote sensing technology with spatial and temporal continuity to carry out disturbance monitoring and classification research on forest ecosystems, which are important components of the Earth’s land surface, can be more accurate and efficient. The use of remote sensing technology with spatial and temporal continuity to monitor and classify disturbances in forest ecosystems, an important component of the Earth’s land surface, can be more accurate and efficient. However, Forest disturbance attribution requires a large number of disturbance type samples, and the laborious manual labeling of high-quality samples and the sparseness of the change region itself limit the number of labelable disturbance samples. In this study, we propose a forest disturbance attribution method under small sample conditions based on the confidence learning mutual guidance framework. In this study, we first used Landsat long time-series remote sensing data to detect forest disturbances between 2000 and 2021 based on the Continuous Change Detection and Classification (CCDC) algorithm to obtain a large amount of unlabeled disturbance data, and then used a small number of manually labeled samples and the mutual guidance framework constructed with Random Forest (RF) and Categorical Boosting (CatBoost) classifiers to iteratively filter high-confidence data from the unlabeled disturbance data through confidence learning, expanding the labeled sample size of each other’s classifiers, and then guided each other's classifications to improve the classification accuracy of forest disturbance attribution. The results show that the overall accuracy of forest disturbance attribution based on the confidence learning mutual guidance framework is 91.4%. Compared with results using only a single classifier, the accuracy is improved by 5%. The method demonstrates excellent performance under small sample conditions and provides an efficient and reliable solution for forest disturbance type classification research.
-
图 1 研究区地理位置(1987年10月26日的Landsat5真彩色图像)
Figure 1. Geographic location of the study area (Natural color image of Landsat5 acquired on 26 October 1987)
图 4 基于RF和CatBoost分类器构建的置信学习互导框架
Figure 4. Schematic diagram of the confidence learning mutual guidance framework constructed based on RF and CatBoost classifiers
图 5 迭代过程中指标的变化. (a)训练样本数量的变化, (b)总体精度的变化
Figure 5. Changes in metrics during iteration. (a) Changes in sample size, (b) changes in overall accuracy
图 6 RF和CatBoost分类器的样本数量及总体精度变化(初始训练样本量: 100, 200, 300)
Figure 6. Variation in the number of samples, overall accuracy of RF and CatBoost classifiers (Initial training samples: 100, 200, 300)
图 7 RF和CatBoost分类器的样本数量及总体精度变化(初始训练样本量: 400, 500)
Figure 7. Variation in the number of samples, overall accuracy of RF and CatBoost classifiers (Initial training samples: 400, 500)
图 8 不同初始训练样本量下单一分类器与互导框架分类性能变化
Figure 8. Changes in classification performance of single classifier and mutual guidance framework with different initial training sample sizes
图 9 基于置信学习互导框架在森林扰动类型分类的错分. (a) (i)均来源于GEP, (b) (j)为扰动分类, (c)～(h)为森林砍伐的Landsat时间序列, 真彩色显示(波段321组合), (k)～(p)为开垦
Figure 9. Misclassification of forest disturbance attribution based on the confidence learning mutual guidance framework. (a) and (i) are from GEP, (b) and (j) are disturbance classification maps, (c)～(h) are Landsat time series of deforestation, shown in true color (band 321 combinations), and (k)～(p) are reclamation
图 10 森林扰动. (a)扰动类型, (b)扰动年度
Figure 10. Forest disturbance maps. (a) Disturbance by causal agents, (b) the disturbance map at an annual step
图 11 基于置信学习互导框架的扰动类型分类. (a) 2010年Landsat8影像真彩色合成——道路建设, (b) 2010年分类结果——道路建设, (c) 2013年Landsat8影像真彩色合成——砍伐, (d) 2013年分类结果——砍伐, (e) 2009年Landsat8影像真彩色合成——火灾, (f) 2013年分类结果——火灾, (g) 2009年Landsat8影像真彩色合成——开垦, (h) 2013年分类结果——开垦
Figure 11. Classification of disturbance types based on the confidence learning mutual conductivity framework. (a) 2010 Landsat8 imagery true color synthesis – roads construction, (b) 2010 classification results – roads construction, (c) 2013 Landsat8 imagery true color composite – logging, (d) 2013 classification results – logging, (e) 2009 Landsat8 imagery true color composite – fire, (f) 2013 classification results – fire, (g) 2009 Landsat8 image true color composite – clearing, (h) 2013 classification results – clearing
图 12 2000－2021年不同扰动类型导致的森林损失面积统计
Figure 12. Statistics on the area of forest loss due to different disturbance types from 2000 to 2021
表 1 研究区所有可用Landsat数据的月统计
Table 1. All available Landsat data for the study area by month
月份 1月 2月 3月 4月 5月 6月 7月 8月 9月 10月 11月 12月 影像数量 28 24 19 21 28 14 28 35 39 34 47 37 
下载: 导出CSV
表 2 本文定义的森林扰动类型
Table 2. Type of forest disturbance defined in this paper
扰动类型 描述 来源 样本数量 (以像元为单位) 道路建设 在森林中修建道路及道路旁的服务设施 GEP 214 砍伐 在不转变为其他土地利用类型的情况下, 砍伐造成的森林干扰 GEP和砍伐记录 214 开垦 在森林中开垦, 用于农作物或经济林种植 GEP 214 火灾 火导致的森林燃烧, 包括野火和人为火灾 GEP和Landsat8年度
火点数据214
下载: 导出CSV
表 3 用于森林扰动类型分类的分类特征
Table 3. Categorical features for forest disturbance attribution
类型 特征(简称) 光谱指数
(Spectrum Index, SI)描述 时间特征 变化前(PreC_SI)、变化后(PostC_SI)、变化幅度(MagC_SI) NBR, SWIR1,
TCW, Tasseled
TCG变化前、变化后指数或波段光谱值以及变化后与变化前的光谱差值 空间特征 滑动窗口内像元标准差(STD_SI)、
均值(Mean_SI)NBR, SWIR1, TCW, TCG 3×3, 5×5, 7×7, 9×9滑动窗口内指数或波段光谱值的标准差、均值 相异性变化幅度(MagDiss_SI)、对比度变化幅度(MagCon_SI)、方差变化幅度(MagVar_SI) NBR, SWIR1, TCG 3×3 滑动窗口内变化后与变化前纹理特征(相异性、对比度、方差)的差值 光谱特征 拟合“片段”的截距变化幅度(Mag_INTP)、斜率变化幅度(Mag_SLP) 截距(INTP)、斜率(SLP) CCDC拟合的变化后与变化前片段的趋势线的截距或斜率差值 地形特征 海拔、坡度 数字高程模型(Digital Elevation Model, DEM) 高程、坡度
下载: 导出CSV
表 4 CCDC扰动检测的空间域精度评估
Table 4. Accuracy assessment of disturbance detection in the spatial domain with different parameters
扰动像元(真实) 稳定像元(真实) 总数 用户精度/(%) 扰动像元(预测) 233 17 250 93.2 稳定像元(预测) 31 219 250 87.6 总数 264 236 500 — 生产者精度/(%) 88.3 92.8 — — 总体精度/(%) 90.4
下载: 导出CSV
表 5 经过后处理的扰动检测空间域精度评估
Table 5. Post-processing accuracy assessment of disturbance detection in the spatial domain
扰动像元(真实) 稳定像元(真实) 总数 用户精度/(%) 扰动像元(预测) 233 17 250 93.2 稳定像元(预测) 19 231 250 90.2 总数 252 248 500 — 生产者精度/(%) 92.5 93.1 — — 总体精度/(%) 92.8
下载: 导出CSV
表 6 经过后处理的扰动检测时间域精度评估
Table 6. Post-processing accuracy assessment of disturbance detection in the temporal domain
相同时间 延迟≤2个时间步长 延迟>2个时间步长 总数 扰动像元 185 5 43 233 比例/(%) 79.4 2.1 18.5 100.0
下载: 导出CSV
表 7 基于置信学习互导框架的森林扰动类型分类混淆矩阵
Table 7. Confusion matrix for forest disturbance attribution based on confidence learning mutual guidance framework
道路建设(预测) 砍伐(预测) 开垦(预测) 火灾(预测) 用户精度/(%) 道路建设(真实) 57 5 0 1 90.4 砍伐(真实) 2 53 3 1 88.3 开垦(真实) 4 1 61 2 92.4 火灾(真实) 0 1 2 64 94.1 生产者精度/(%) 90.5 89.8 89.7 94.1 — 总体精度/(%) 91.4
下载: 导出CSV
-
[1] 施建成, 赵天杰, 杨晓峰. 空间地球科学视角下的全球水循环研究[J]. 遥感学报, 2021, 25(4): 847-855 doi: 10.11834/jrs.20219467SHI Jiancheng, ZHAO Tianjie, YANG Xiaofeng. Global water cycle studies from the perspective of space earth science[J]. National Remote Sensing Bulletin, 2021, 25(4): 847-855 doi: 10.11834/jrs.20219467 [2] 田镇朋, 周维, 袁敬毅, 等. 基于多源遥感数据的植被冠层高度估算[J]. 空间科学学报, 2023, 43(6): 1176-1193 doi: 10.11728/cjss2023.06.2023-0074TIAN Zhenpeng, ZHOU Wei, YUAN Jingyi, et al. Forest canopy height mapping based on multi-source remote sensing data[J]. Chinese Journal of Space Science, 2023, 43(6): 1176-1193 doi: 10.11728/cjss2023.06.2023-0074 [3] XU R, LI Y, TEULING A J, et al. Contrasting impacts of forests on cloud cover based on satellite observations[J]. Nature Communications, 2022, 13(1): 670 doi: 10.1038/s41467-022-28161-7 [4] FEDERICI S, TUBIELLO F N, SALVATORE M, et al. New estimates of CO2 forest emissions and removals: 1990-2015[J]. Forest Ecology and Management, 2015, 352: 89-98 doi: 10.1016/j.foreco.2015.04.022 [5] MACDICKEN K G. Global forest resources assessment 2015: What, why and how?[J]. Forest Ecology and Management, 2015, 352: 3-8 doi: 10.1016/j.foreco.2015.02.006 [6] STAHL A T, ANDRUS R, HICKE J A, et al. Automated attribution of forest disturbance types from remote sensing data: a synthesis[J]. Remote Sensing of Environment, 2023, 285: 113416 doi: 10.1016/j.rse.2022.113416 [7] SENF C, SEIDL R. Mapping the forest disturbance regimes of Europe[J]. Nature Sustainability, 2021, 4(1): 63-70 [8] MCDOWELL N G, ALLEN C D, ANDERSON-TEIXEIRA K, et al. Pervasive shifts in forest dynamics in a changing world[J]. Science, 2020, 368(6494): 964 [9] OESER J, PFLUGMACHER D, SENF C, et al. Using intra-annual Landsat time series for attributing forest disturbance agents in Central Europe[J]. Forests, 2017, 8(7): 251 doi: 10.3390/f8070251 [10] SEBALD J, SENF C, SEIDL R. Human or Natural? Landscape context improves the attribution of forest disturbances mapped from Landsat in central Europe[J]. Remote Sensing of Environment, 2021, 262: 112502 doi: 10.1016/j.rse.2021.112502 [11] PICKETT S T A, WHITE P S. The Ecology of Natural Disturbance and Patch Dynamics[M]. New York: Academic Press, 1985 [12] BONAN G B. Forests and climate change: forcings, feedbacks, and the climate benefits of forests[J]. Science, 2008, 320(5882): 1444-1449 doi: 10.1126/science.1155121 [13] GOLDSTEIN G, SANTIAGO L S. Tropical Tree Physiology: Adaptations and Responses in A Changing Environment[M]. Cham: Springer, 2016: 337-355 [14] MASCORRO V S, COOPS N C, KURZ W A, et al. Choice of satellite imagery and attribution of changes to disturbance type strongly affects forest carbon balance estimates[J]. Carbon Balance and Management, 2015, 10(1): 30 doi: 10.1186/s13021-015-0041-6 [15] LEHNERT L W, BäSSLER C, BRANDL R, et al. Conservation value of forests attacked by bark beetles: highest number of indicator species is found in early successional stages[J]. Journal for Nature Conservation, 2013, 21(2): 97-104 doi: 10.1016/j.jnc.2012.11.003 [16] SEIDL R, RAMMER W, JÄGER D, et al. Impact of bark beetle (Ips typographus L. ) disturbance on timber production and carbon sequestration in different management strategies under climate change[J]. Forest Ecology and Management, 2008, 256(3): 209-220 doi: 10.1016/j.foreco.2008.04.002 [17] DE SY V, HEROLD M, ACHARD F, et al. Land use patterns and related carbon losses following deforestation in South America[J]. Environmental Research Letters, 2015, 10(12): 124004 doi: 10.1088/1748-9326/10/12/124004 [18] CURTIS P G, SLAY C M, HARRIS N L, et al. Classifying drivers of global forest loss[J]. Science, 2018, 361(6407): 1108-1111 doi: 10.1126/science.aau3445 [19] CHEN X, ZHAO W Z, CHEN J G, et al. Mapping large-scale forest disturbance types with multi-temporal CNN framework[J]. Remote Sensing, 2021, 13(24): 5177 doi: 10.3390/rs13245177 [20] LI Y T, XU X, WU Z Z, et al. A forest type-specific threshold method for improving forest disturbance and agent attribution mapping[J]. GIScience :Times New Roman;">& Remote Sensing, 2022, 59(1): 1624-1642 [21] DE MARZO T, GASPARRI N I, LAMBIN E F, et al. Agents of forest disturbance in the argentine dry Chaco[J]. Remote Sensing, 2022, 14(7): 1758 doi: 10.3390/rs14071758 [22] ZHANG Y T, WOODCOCK C E, CHEN S J, et al. Mapping causal agents of disturbance in boreal and arctic ecosystems of North America using time series of Landsat data[J]. Remote Sensing of Environment, 2022, 272: 112935 doi: 10.1016/j.rse.2022.112935 [23] SHIMIZU K, OTA T, MIZOUE N, et al. A comprehensive evaluation of disturbance agent classification approaches: strengths of ensemble classification, multiple indices, Spatio-temporal variables, and direct prediction[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2019, 158: 99-112 doi: 10.1016/j.isprsjprs.2019.10.004 [24] 吴伶, 刘湘南, 刘美玲, 等. 融合遥感时间序列时空谱信息的森林扰动检测与归因研究进展[J]. 遥感学报, 2024, 28(3): 558-575WU Ling, LIU Xiangnan, LIU Meiling, et al. Review of the detection and attribution of multi-type forest disturbances using an ensemble of Spatio-temporal-spectral information from remote sensing images[J]. National Remote Sensing Bulletin, 2024, 28(3): 558-575 [25] XIAO P F, SHENG G W, ZHANG X L, et al. Direction-dominated change vector analysis for forest change detection[J]. International Journal of Applied Earth Observation and Geoinformation, 2021, 103: 102492 doi: 10.1016/j.jag.2021.102492 [26] FRANKLIN S, ROBITAILLE S. Forest insect defoliation and mortality classification using annual Landsat time series composites: a case study in Northwestern Ontario, Canada[J]. Remote Sensing Letters, 2020, 11(12): 1175-1180 doi: 10.1080/2150704X.2020.1828659 [27] LI Y T, WU Z Z, XU X, et al. Forest disturbances and the attribution derived from yearly Landsat time series over 1990-2020 in the Hengduan mountains region of Southwest China[J]. Forest Ecosystems, 2021, 8(1): 73 doi: 10.1186/s40663-021-00352-6 [28] VOGELER J C, SLESAK R A, FEKETY P A, et al. Characterizing over four decades of forest disturbance in Minnesota, USA[J]. Forests, 2020, 11(3): 362 doi: 10.3390/f11030362 [29] MURILLO-SANDOVAL P J, HILKER T, KRAWCHUK M A, et al. Detecting and attributing drivers of forest disturbance in the Colombian Andes using Landsat Time-Series[J]. Forests, 2018, 9(5): 269 doi: 10.3390/f9050269 [30] DAS P, MUDI S, BEHERA M D, et al. Automated mapping for long-term analysis of shifting cultivation in Northeast India[J]. Remote Sensing, 2021, 13(6): 1066 doi: 10.3390/rs13061066 [31] GAO B C. NDWI-a normalized difference water index for remote sensing of vegetation liquid water from Space[J]. Remote Sensing of Environment, 1996, 58(3): 257-266 doi: 10.1016/S0034-4257(96)00067-3 [32] CHEN S J, OLOFSSON P, SAPHANGTHONG T, et al. Monitoring shifting cultivation in Laos with Landsat Time series[J]. Remote Sensing of Environment, 2023, 288: 113507 doi: 10.1016/j.rse.2023.113507 [33] CARDILLE J A, PEREZ E, CROWLEY M A, et al. Multi-sensor change detection for within-year capture and labelling of forest disturbance[J]. Remote Sensing of Environment, 2022, 268: 112741 doi: 10.1016/j.rse.2021.112741 [34] FRAZIER R J, COOPS N C, WULDER M A. Boreal shield forest disturbance and recovery trends using Landsat time series[J]. Remote Sensing of Environment, 2015, 170: 317-327 doi: 10.1016/j.rse.2015.09.015 [35] SENF C, PFLUGMACHER D, HOSTERT P, et al. Using Landsat time series for characterizing forest disturbance dynamics in the coupled human and natural systems of Central Europe[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2017, 130: 453-463 doi: 10.1016/j.isprsjprs.2017.07.004 [36] 高珺烨. 基于多源遥感数据的森林类型动态变化研究[D]. 长沙: 中南林业科技大学, 2021GAO Junye. Dynamic Change of Forest Types Based on Multi-Source Remote Sensing Data[D]. Changsha: Central South University of Forestry & Technology, 2021 [37] ZHANG T W, WU L, LIU X N, et al. Detection of forest disturbances with different intensities using Landsat time series based on adaptive exponentially weighted moving average charts[J]. Forests, 2023, 15(1): 19 doi: 10.3390/f15010019 [38] 王开新, 夏昕. 基于GIS技术的通道县森林资源特征分析[J]. 南方农业, 2024, 18(13): 258-261WANG Kaixin, XIA Xin. Analysis on forest resource characteristics in Tongdao County based on GIS technology[J]. South China Agriculture, 2024, 18(13): 258-261 [39] 吴子斌, 龙丁秀, 陆志旺, 等. 湖南省通道县侗族聚居村寨环境生态系统调查[J]. 绿色科技, 2016(6): 70-73,75 doi: 10.3969/j.issn.1674-9944.2016.06.032WU Zibin, LONG Dingxiu, LU Zhiwang, et al. Investigation on habitation village ecosytems dominated by dong people in Tongdao County, Hu'nan province[J]. Journal of Green Science and Technology, 2016(6): 70-73,75 doi: 10.3969/j.issn.1674-9944.2016.06.032 [40] ZHU Z, WOODCOCK C E. Continuous change detection and classification of land cover using all available Landsat Data[J]. Remote Sensing of Environment, 2014, 144: 152-171 doi: 10.1016/j.rse.2014.01.011 [41] MASEK J G, VERMOTE E F, SALEOUS N E, et al. A Landsat surface reflectance data set for North America, 1990-2000[J]. Geoscience and Remote Sensing Letters, 2006, 3(1): 68-72 doi: 10.1109/LGRS.2005.857030 [42] OLOFSSON P, FOODY G M, HEROLD M, et al. Good practices for estimating area and assessing accuracy of land change[J]. Remote Sensing of Environment, 2014, 148: 42-57 doi: 10.1016/j.rse.2014.02.015 [43] FCPF. Guidelines on the application of the Methodological framework Number 2-On Technical Corrections to GHG emissions and removals reported in the reference period[EB/OL]. (2020-11-01)[2024-10-29]. https://www.forestcarbonpartnership.org/sites/fcp/files/FCPF%20Guidelines%20on%20the%20Application%20of%20the%20Methodological%20Framework%20Number%202_2020_0.pdf [44] CAVALHEIRO L P, BERNARD S, BARDDAL J P, et al. Random forest kernel for high-dimension low sample size classification[J]. Statistics and Computing, 2024, 34(1): 9 doi: 10.1007/s11222-023-10309-0 [45] DONG J X, YU Z S, ZHANG X K, et al. Data-driven predictive prognostic model for power batteries based on machine learning[J]. Process Safety and Environmental Protection, 2023, 172: 894-907 doi: 10.1016/j.psep.2023.02.081 [46] NORTHCUTT C G, JIANG L, CHUANG I. Confident learning: estimating uncertainty in dataset labels[J]. Journal of Artificial Intelligence Research, 2021, 70: 1373-1411 doi: 10.1613/jair.1.12125 [47] TAI X X, LI M J, XIANG M, et al. A mutual guide framework for training hyperspectral image classifiers with small data[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 5510417 -
-
下载:
计量
- 文章访问数: 268
- HTML全文浏览量: 92
- PDF下载量: 19
-
被引次数:
0(来源:Crossref)
0(来源:其他)
