UML Sourcing Domain Predictions

<strong>WHAT: </strong>This data accompanies Glick, H.B., Ament, J.M., Dallinga, J.S., Torres-Batlló, J., Verma, M., Clinton, N., and Wilcox, A. (2023). Model-based prediction and ascription of deforestation risk within commodity sourcing domains: Improving traceability in the palm oil supply chain.<br> <br> The data represents a collection of geographically referenced raster images (GeoTIFF) capturing the predicted sourcing domains of 1,570 palm oil processing facilities (mills) in Indonesia and Malaysia. There is one image per facility, delivered in WGS84 (EPSG:4326) with a nominal equatorial spatial resolution of 250 m. With respect to the models discussed in Glick et al (2023), each image file captures the results of our most accurate model, which was an ensemble of predictions from MaxEnt, random forest, and gradient boosted regression tree-based machine learning models, trained on passive geolocational traceability data (n = 3,355,437 cellular pings). The individual pixel values in each image are the probability of containing aggregated cellular ping data from individuals that have been spatio-temporally linked to the given processing facility. Functionally, these values represent the probability that a given pixelated location is part of a facility's sourcing domain, where a sourcing domain encompasses both harvesting locations and the intermediary transportation and sub-processing space. Please refer to the parent manuscript for details. Please note that our predictions were derived from a processing chain that used a modified World Mollweide projected coordinate system (essentially ESRI:54009), where the central meridian (longitude of origin) was set to 109.5 degrees. The images are delivered here in WGS84 (EPSG:4326). Users can access the data in its original coordinate reference system using a Google Earth Engine ImageCollection: ee.ImageCollection('projects/ul-gs-d-901791-09-prj/assets/users/hglick/Glick_et_al_2023/Ensemble). <strong>WHEN:</strong> Passive geolocational training data was gathered in 2020 and 2021. Palm oil processing facilities were derived from the Universal Mill List in November 2021. <strong>WHERE:</strong> All images contain predictions for palm oil processing facilities located in Indonesia and Malaysia, with predictions made to a maximum Euclidean distance of 100 km from each facility. <strong>WHY:</strong> Palm oil accounts for approximately 50% of global vegetable oil production, and trends in consumption have driven large-scale expansion of oil palm (<em>Elaeis guineensis</em>) plantations in Southeast Asia. This expansion has led to deforestation and other socio-environmental concerns that challenge consumer goods companies to meet no deforestation and sustainability commitments. In support of these commitments and supply chain traceability, we seek to improve on the current industry standard sourcing model for ascribing social and environmental risks to particular actors. Among other uses, this data can support the ascription or allocation of deforestation, carbon loss, and biodiversity risk to relevant actors, permitting targeted outreach, contract negotiation, and mitigation of large-scale resource degradation. <strong>HOW:</strong> Passive geolocational training data was gathered by Orbital Insights. All modeling was conducted in Python; Google Earth Engine served as the primary distributed computing platform. Details are presented in Glick et al (2023).**【数据内容（WHAT）】** 本数据集配套Glick, H.B.、Ament, J.M.、Dallinga, J.S.、Torres-Batlló, J.、Verma, M.、Clinton, N.及Wilcox, A.于2023年发表的研究论文《基于模型的商品采购域内森林砍伐风险预测与归因：提升棕榈油供应链溯源能力》（原标题：*Model-based prediction and ascription of deforestation risk within commodity sourcing domains: Improving traceability in the palm oil supply chain*）。 本数据集包含一系列地理参考栅格图像（GeoTIFF），涵盖印度尼西亚与马来西亚境内1570座棕榈油加工厂（榨油厂）的预测采购域。每座工厂对应一幅图像，采用WGS84（EPSG:4326）坐标系，标称赤道空间分辨率为250米。 针对Glick等人2023年研究中提及的模型，每幅图像均存储了我们最优模型的预测结果：该模型为基于最大熵模型（MaxEnt）、随机森林（random forest）及梯度提升回归树（gradient boosted regression tree）的集成机器学习模型，训练数据为被动地理溯源数据集，包含3,355,437条蜂窝网络ping（cellular pings）记录。图像中各像素的数值代表该像素包含与对应加工厂经时空关联的个体聚合蜂窝ping数据的概率，即该像素位置属于该加工厂采购域的概率。采购域涵盖收获点位、中间运输环节及次级加工空间。详细信息请参见主论文。 需注意，本研究的预测结果基于经修改的World Mollweide投影坐标系（即ESRI:54009）生成，其中央子午线（原点经度）设为109.5°。本次发布的图像采用WGS84坐标系。用户可通过谷歌地球引擎（Google Earth Engine）的图像集获取原始投影坐标系下的数据集：`ee.ImageCollection('projects/ul-gs-d-901791-09-prj/assets/users/hglick/Glick_et_al_2023/Ensemble')`。 **【数据采集时间（WHEN）】** 被动地理定位训练数据采集于2020年与2021年。棕榈油加工厂列表源自2021年11月发布的通用榨油厂清单（Universal Mill List）。 **【数据覆盖范围（WHERE）】** 所有图像均针对位于印度尼西亚与马来西亚的棕榈油加工厂生成，预测范围覆盖各加工厂周边最大100公里的欧几里得距离区域。 **【数据应用场景（WHY）】** 棕榈油约占全球植物油产量的50%，消费需求增长推动东南亚地区油棕（*Elaeis guineensis*）种植园大规模扩张。此类扩张引发了森林砍伐及其他社会环境问题，对消费品企业落实“零砍伐”与可持续发展承诺构成挑战。为支撑上述承诺及供应链溯源工作，本研究旨在优化当前行业通用的采购域风险归因模型，将社会与环境风险精准分配至特定市场主体。本数据集可用于将森林砍伐、碳损失及生物多样性风险归因或分配至相关责任主体，助力开展定向触达、合同谈判及大规模资源退化的减缓行动。 **【数据生成方法（HOW）】** 被动地理定位训练数据由Orbital Insights公司采集。所有建模工作均通过Python完成，谷歌地球引擎（Google Earth Engine）为核心分布式计算平台。详细方法请参见Glick等人2023年的研究论文。