Mapping built infrastructure in semi-arid systems using data integration and open-source approaches for image classification

Accurate land use land cover (LULC) maps that delineate built infrastructure are useful for numerous applications, from urban planning, humanitarian response, disaster management, to informing decision making for reducing human exposure to natural hazards, such as wildfire. Existing products lack sufficient spatial, temporal, and thematic resolution, omitting critical information needed to capture LULC trends accurately over time. Advancements in remote sensing imagery, open-source software and cloud computing offer opportunities to address these challenges. Using Google Earth Engine, we developed a novel built infrastructure detection method in semi-arid systems by applying a random forest classifier to a fusion of Sentinel-1 and Sentinel-2 time series. Our classifier performed well, differentiating three built environment types: residential, infrastructure, and paved, with overall accuracies ranging from 90 to 96%. Producer accuracies were highest for the infrastructure class (98–99%), followed by the residential class (91–96%). Sentinel-1 variables were important for differentiating built classes. We illustrated the utility of our mapped products by generating a time-series of change across southern Idaho spanning 2015 to 2024 and comparing this with publicly available products: National Land Cover Database (NLCD), Microsoft Building Footprints (MBF) and the global Dynamic World (DW). For 2024, our product estimated 5.88% of the study area as built, aligning closely with NLCD (6%) and DW (4.64%). Our mapped built infrastructure products offer enhancements over NLCD spatially and temporally, over DW thematically, and over MBF both temporally and thematically. We demonstrate the potential of fusing data sources to improve LULC mapping and present a case for regionally parameterized models that can more accurately capture built infrastructure change over time. We used open-source approaches for built infrastructure detection, aiming for broader adoption of this workflow across other ecosystems and environments to support decision-making.

精准描绘建筑基础设施的土地利用/土地覆盖（Land Use Land Cover, LULC）地图，可广泛应用于城市规划、应急响应、灾害管理，以及为降低人类暴露于野火等自然灾害的风险提供决策依据。现有相关产品在空间、时间与专题分辨率上存在不足，遗漏了精准捕捉长期LULC变化趋势所需的关键信息。遥感影像、开源软件与云计算技术的进步，为解决上述难题提供了契机。本研究借助谷歌地球引擎（Google Earth Engine），将哨兵1号（Sentinel-1）与哨兵2号（Sentinel-2）的时间序列影像融合后输入随机森林分类器，在半干旱区域开发了一种全新的建筑基础设施提取方法。该分类器表现优异，可有效区分住宅、公共基础设施与铺装路面三类建筑环境类型，总体精度达90%至96%。基础设施类别的生产者精度最高，达98%至99%，其次为住宅类别（91%至96%）。哨兵1号影像变量在区分建筑类别中发挥了关键作用。本研究通过生成2015年至2024年爱达荷州南部地区的建筑变化时间序列，并与公开可用的美国国家土地覆盖数据库（National Land Cover Database, NLCD）、微软建筑轮廓数据集（Microsoft Building Footprints, MBF）以及全球动态世界数据集（Dynamic World, DW）进行对比，验证了所生成制图产品的实用性。2024年，本研究产品估算研究区域内建筑用地占比为5.88%，与NLCD的6%、DW的4.64%结果较为接近。本研究所生成的建筑基础设施制图产品，在空间与时间维度上优于NLCD，在专题维度上优于DW，同时在时间与专题维度上均优于MBF。本研究证实了多源数据融合可有效提升LULC制图精度，并提出了可精准捕捉长期建筑基础设施变化的区域参数化模型方案。本研究采用开源方法开展建筑基础设施提取，旨在推动该流程在其他生态系统与区域的推广应用，为相关决策提供支撑。