Wetland Vegetation of the Lachlan - Great Cumbung Swamp 2023

This wetland vegetation map of the Great Cumbung Swamp was produced using a machine learning-based classification framework that integrates multi-source satellite and terrain with a cluster-guided training approach (Wen et al., 2025).\r\n \r\nInputs and training data\r\n\r\nInputs included Sentinel-1 synthetic aperture radar (SAR) time series, Sentinel-2 optical time series, and hydro-morphological variables derived from a gap-filled 5 m LiDAR digital elevation model (DEM) and hydrologically enforced shuttle radar topography mission (SRTM) DEM. \r\nTo capture the high spatial and seasonal variability of wetland vegetation, K-means clustering was used to guide sample selection. Clusters were reviewed by an expert vegetation ecologist against high-resolution aerial and drone imagery, topographic context, and existing field data, and then assigned to plant community types (PCTs) where appropriate. The verified clusters formed the basis of the training dataset for a Random Forest classifier which used 48 predictors (spectral, temporal, structural, terrain). Model outputs were produced at three hierarchical class levels: NSW Vegetation Formations (L1: 9 classes), Functional (L2: 14 classes) and PCTs (L3: 23 classes).\r\n\r\nPost-processing and manual edits\r\n\r\nFollowing classification, model outputs were post-processed to enhance spatial coherence while preserving hydrologically meaningful patches. Steps included edge-aware smoothing and progressive gap-filling/merging with class-specific minimum mapping units (MMU): < 0.1 ha for non-woody wetland PCTs and < 0.2 ha for woody wetland PCTs. Outputs were then manually edited by an expert vegetation ecologist to resolve any residual artifacts and boundary issues.\r\n\r\nModel accuracy assessment\r\n\r\nThe following metrics are the raw model output (before post-processing and editing) performance for each class level in Wen et al. (2025) (reported on internal independent test set). Metrics include Overall Accuracy (OA), Cohens Kappa (κ) and Matthews Correlation Coefficient (MCC):\r\n\r\n- NSW Vegetation Formations (L1): OA ≈ 97 %, κ ≈ 0.96, MCC ≈ 0.96;\r\n\r\n- Functional (L2): OA ≈ 94 %, κ ≈ 0.93, MCC ≈ 0.93;\r\n\r\n- PCTs (L3): OA ≈ 93 %, κ ≈ 0.91, MCC ≈ 0.89\r\n\r\nClass hierarchy\r\n\r\nLabels were assigned at PCT level using the NSW BioNet Vegetation Classification (https://vegetation.bionet.nsw.gov.au/) and then aligned to the NSW framework’s Vegetation Class and Formation levels (https://www.environment.nsw.gov.au/topics/animals-and-plants/biodiversity/nsw-bionet/the-nsw-vegetation-classification-framework). For water management reporting, each wetland PCT was aligned to a Monitoring, Evaluation and Reporting (MER) Functional Group consistent with the Lachlan Long-Term Water Plan (LTWP) (https://www.environment.nsw.gov.au/sites/default/files/lachlan-long-term-water-plan).\r\n\r\nKey fields dictionary\r\n\r\n‘PCT_ID’ (PCT Code); ‘PCT_Desc’ (PCT Name); ‘Veg_Class’ (NSW Vegetation Class); ‘Veg_Format’ (NSW Vegetation Formation); ‘MER_FG’ (MER Functional Group for LTWP reporting); ‘Hectares’ (polygon area); ‘DN’ (classifier code) and ‘Functional’ (model specific functional group per Wen et al., 2025). Context classes (‘Bare ground’, ‘Cleared/Disturbed’, ‘Open water’, ‘Dam') are included for completeness and accuracy assessment.\r\n\r\nIntended use\r\n\r\nBaseline for environmental water planning, MER reporting under the LTWP, conservation management, and long-term monitoring at landscape and site scales. Not intended for statutory site assessment without targeted field verification.\r\n\r\nInput data limitations\r\n\r\nCloud, inundation state and sensor geometry may influence satellite image quality and contribute to classification error; LiDAR and ancillary datasets may differ in acquisition date from satellite inputs. Localised errors in source DEMs/orthophoto errors can propagate to terrain-derived predictors. \r\n\r\nValidation scope\r\n\r\nThe above model accuracy metrics are from internal hold-out testing (80/20 train-test split) and repeated cross-validation of the expert-labelled dataset in Wen et al. (2025). A withheld, ground-based validation dataset collected independent of model training will be used to validate the final post-processed and edited map product; those results will be provided in future versions to supplement the raw model accuracy values for reporting purposes. Users requiring statutory-grade evidence should conduct targeted field verification.\r\n\r\nVersioning\r\n\r\nThis version is v1.0 (release date: 2025-10-30). Results are versioned; Identified errors will be corrected in subsequent releases with an accompanying changelog. \r\n\r\nAcknowledgements\r\n\r\nThis mapping project was funded by the NSW Water for the Environment Program.\r\n\r\nRelated publication\r\n\r\nWen, L., Ryan, S., Powell, M., and Ling, J.E. (2025). From Clusters to Communities: Enhancing Wetland Vegetation Mapping Using Unsupervised and Supervised Synergy. Remote Sensing, 17(13): 2279. https://doi.org/10.3390/rs17132279\r\n

大坎邦沼泽（Great Cumbung Swamp）湿地植被图基于集成多源卫星与地形数据的机器学习分类框架制作，该框架采用聚类引导的训练方法（Wen等，2025）。

输入数据与训练数据集

输入数据包括哨兵-1合成孔径雷达（Sentinel-1 Synthetic Aperture Radar, SAR）时间序列、哨兵-2（Sentinel-2）光学时间序列，以及源自间隙填充的5米分辨率激光雷达数字高程模型（LiDAR Digital Elevation Model, DEM）和经水文修正的航天飞机雷达地形测绘任务（Shuttle Radar Topography Mission, SRTM）DEM的水文地貌变量。
为捕捉湿地植被的高空间与季节变异性，研究采用K均值聚类（K-means clustering）指导样本选择。资深植被生态学家结合高分辨率航空与无人机影像、地形背景及现有野外调查数据对聚类结果进行复核，随后将符合条件的聚类结果分配至植物群落类型（Plant Community Types, PCTs）。经验证的聚类结果构成随机森林分类器（Random Forest Classifier）的训练数据集基础，该分类器共使用48个预测变量（涵盖光谱、时序、结构与地形特征）。模型输出分为三级分类层级：新南威尔士州（New South Wales, NSW）植被群系（层级1：9类）、功能组（层级2：14类）及植物群落类型（层级3：23类）。

后处理与人工编辑

分类完成后，对模型输出进行后处理以提升空间连贯性，同时保留具有水文意义的斑块。处理步骤包括边缘感知平滑算法，以及结合分类专属最小制图单元（Minimum Mapping Units, MMU）的渐进式间隙填充与合并：非木本湿地PCT的最小制图单元小于0.1公顷，木本湿地PCT的最小制图单元小于0.2公顷。随后由资深植被生态学家对输出结果进行人工编辑，以修正残留伪影与边界问题。

模型精度评估

以下指标为Wen等（2025）中各分类层级的原始模型输出（后处理与人工编辑前）性能（基于内部独立测试集报告）。评估指标包括总体精度（Overall Accuracy, OA）、科恩卡帕系数（Cohens Kappa, κ）与马修斯相关系数（Matthews Correlation Coefficient, MCC）：
- 新南威尔士州植被群系（层级1）：OA≈97%，κ≈0.96，MCC≈0.96；
- 功能组（层级2）：OA≈94%，κ≈0.93，MCC≈0.93；
- 植物群落类型（层级3）：OA≈93%，κ≈0.91，MCC≈0.89

分类层级

标签基于植物群落类型层级赋值，采用新南威尔士州生物网植被分类系统（NSW BioNet Vegetation Classification，https://vegetation.bionet.nsw.gov.au/），随后对齐至新南威尔士州植被分类框架的植被类别与群系层级（https://www.environment.nsw.gov.au/topics/animals-and-plants/biodiversity/nsw-bionet/the-nsw-vegetation-classification-framework）。为满足水管理报告需求，每个湿地PCT均对齐至适配拉克兰长期水资源计划（Lachlan Long-Term Water Plan, LTWP）的监测、评估与报告（Monitoring, Evaluation and Reporting, MER）功能组（https://www.environment.nsw.gov.au/sites/default/files/lachlan-long-term-water-plan）。

关键字段字典

"PCT_ID"（PCT代码）；"PCT_Desc"（PCT名称）；"Veg_Class"（新南威尔士州植被类别）；"Veg_Formation"（新南威尔士州植被群系，原文字段名为Veg_Format）；"MER_FG"（适配拉克兰长期水资源计划的MER功能组）；"Hectares"（多边形面积，单位：公顷）；"DN"（分类器编码）；"Functional"（Wen等，2025中定义的模型专属功能组）。为保证完整性与精度评估，同时纳入上下文类别：裸地（Bare ground）、已清理/受干扰区域（Cleared/Disturbed）、开阔水体（Open water）、水坝（Dam）。

预期用途

本数据集可作为环境水规划、拉克兰长期水资源计划下的MER报告、保护管理以及景观与场地尺度长期监测的基线数据。未经针对性野外验证，不得用于法定场地评估。

输入数据局限性

云覆盖、淹没状态以及传感器几何姿态可能影响卫星影像质量，进而引发分类误差；激光雷达与辅助数据集的获取时间可能与卫星输入数据不一致。源DEM/正射影像的局部误差可能传播至地形衍生预测变量中。

验证范围

上述模型精度指标源自Wen等（2025）中专家标注数据集的内部预留测试（训练测试划分比例80/20）与重复交叉验证。将采用独立于模型训练的预留地面验证数据集，对最终经过后处理与人工编辑的地图产品进行验证；相关结果将在后续版本中发布，以补充原始模型精度值用于报告需求。有法定级证据需求的用户应开展针对性野外验证。

版本管理

本版本为v1.0（发布日期：2025-10-30）。结果已进行版本管理；识别出的错误将在后续版本中修正，并附带变更日志。

致谢

本制图项目由新南威尔士州环境水计划资助。

相关出版物

Wen, L., Ryan, S., Powell, M., and Ling, J.E. (2025). 从聚类到群落：基于无监督与监督学习协同的湿地植被制图精度提升. 《遥感》（Remote Sensing）, 17(13): 2279. https://doi.org/10.3390/rs17132279