DroughtScope dataset for AI4EO IMAGIN-e Challenge

Climate change poses new challenges to the food industry, which is required to optimise production in terms of quantity and quality of products against reduced water and land availability due to increased periods of drought and urbanisation. The availability of accurate and real-time information about ET is expected to improve decision-making concerning the optimisation of agricultural practice by properly allocating the resources needed. In this regard, the idea of using real-time ET estimates has a significant societal value, as it contributes to <strong>sustainable agriculture</strong>, <strong>food security</strong>, <strong>water conservation</strong>, and <strong>climate change adaptation</strong>. The envisioned business model includes data-driven agricultural consulting services, a subscription-based platform and partnerships with agritech companies.<br> Evapotranspiration (ET) data helps irrigation planning and is a powerful tool for managing land and water resources. Water scarcity makes knowledge of crop water consumption essential for water budgeting of connected ecosystems like agriculture, industry and cities. In a context in which severe droughts induce farmers to change crop typology to face reduced water availability, even in areas historically characterised by water abundance, offline processing of satellite data can result in delayed or sub-optimal decision-making. Timely information about the status of the cultivations can dramatically reduce water consumption, leading to cost savings, increased quality and quantity of food production and profitability. In this context, exploiting pre-trained models is an enabling factor for onboard processing, which is not feasible for direct retrieval of ET from measured brightness temperature via computationally expensive radiative transfer models.<br> This project aims to develop a decision support system helping public and private agrifood decision-makers optimise water resource management. Our solution provides <strong>early warnings</strong> about <strong>crop water stress</strong> based on the estimate of the evaporative stress index (<strong>ESI</strong>), which is one of the most important indicators of stress conditions identified in the literature.<br> <strong>Main innovations</strong> Original deep learning architecture to infer information typically contained in the thermal domain from the near-infrared one Real-time onboard estimates thanks to lightweight processing Multi-task architecture for ESI estimation and LULC mapping <strong>Benefits for users/stakeholders</strong> Optimization of water resources management to tackle increasing water scarcity Possibility to implement local actions based on daily warning maps Optimization of food production Reduction of costs and resources employed for field monitoring <strong>Benefits for the scientific community</strong> Availability of new data with better resolution of the ones today available provided by the ECOSTRESS mission Daily LULC maps to tackle with high-dynamic phenomena (river mobility, sedimentation, deforestation) <strong>Data</strong><br> This research exploits L1C data from the <strong>IMAGIN-e hyperspectral</strong> (HS) sensor, which features 50 spectral bands within the range [450,950] nm with a spatial resolution of 45 metres. The primary data sources are integrated with auxiliary data useful for classifying agricultural lands and estimating Evapotranspiration (ET).<br> To this end, we use the <strong>ESA WorldCover 2021</strong> dataset (https://worldcover2021.esa.int/) as the ground truth for Crop Classification; the original dataset at 10m spatial resolution is resampled at 45 meters with a max-voting strategy. Finally, a binary map is obtained in the code, considering only the Cropland values as 1 and 0 otherwise.<br> For ET training, we have considered <strong>ECOSTRESS</strong> data as a reference. ECOSTRESS provides ET data based on the Priestley-Taylor Jet Propulsion Laboratory (PT-JPL) method, with a spatial resolution of approximately 70 metres, which will be rescaled to 45 metres for consistency with input HS data.<br> The patches (14) included in the dataset are a stack of all these data. So we have 54 channels in this order: 1:50: hyperspectral bands 51: ESA WorldCover 2021 at 45m 52: Evaporative Stress Index (ESI) 53: Quality Control Band 54: Evapotranspiration In the data folder are also available original ESI and LULC data.气候变化给食品工业带来了全新挑战：由于干旱周期延长与城市化进程加剧，可用水资源与土地资源持续缩减，食品工业亟需在产品产量与质量层面优化生产。精准且实时的蒸散发（Evapotranspiration, ET）信息，可通过合理配置所需资源，助力农业生产优化相关决策。就此而言，实时ET估算方案具备显著社会价值，可为**可持续农业**、**粮食安全**、**水资源保护**以及**气候变化适应**提供支撑。该项目拟采用的商业模式涵盖数据驱动型农业咨询服务、订阅制平台，以及与农业科技企业的合作关系。 蒸散发（ET）数据可为灌溉规划提供支撑，同时也是管理土地与水资源的有力工具。水资源短缺背景下，掌握作物需水量，对农业、工业、城市等关联生态系统的水资源预算至关重要。在极端干旱迫使农民调整作物类型以应对水资源缩减的当下——即使在历史上水资源充沛的地区亦是如此——通过离线处理卫星数据往往会导致决策延迟或不够优化。及时获取作物生长状态信息，可大幅降低水资源消耗，实现成本节约、粮食产量与质量双提升，以及盈利能力增强。在此场景下，利用预训练模型开展星上处理成为可行方案，而通过计算成本高昂的辐射传输模型直接从实测亮温反演ET则难以实现。 本项目旨在开发一套决策支持系统，助力公共与私营农业食品领域决策者优化水资源管理。我们的方案基于蒸散发胁迫指数（Evaporative Stress Index, ESI）的估算结果，提供**作物水分胁迫**相关的**早期预警**，而ESI是学界公认的最重要胁迫状态指示指标之一。 **核心创新点** 原创深度学习架构：可从近红外波段信息中提取通常仅能从热红外波段获取的相关信息 轻量化处理实现实时星上估算 支持ESI估算与土地利用/土地覆盖（Land Use and Land Cover, LULC）制图的多任务架构 **用户与利益相关方收益** 优化水资源管理，应对日益严峻的水资源短缺问题 可基于每日预警地图实施本地化应对措施 优化粮食生产 减少田间监测所需的成本与资源投入 **科研社区收益** 获得分辨率优于现有ECOSTRESS任务数据的全新数据集 获取每日LULC地图，以应对高动态现象（如河流迁移、泥沙沉积、森林砍伐） **数据集说明** 本研究采用来自**IMAGIN-e高光谱（Hyperspectral, HS）**传感器的L1C级数据，该传感器拥有50个光谱波段，波段范围为[450,950] nm，空间分辨率为45米。本研究的核心数据源将与辅助数据结合，用于农业用地分类与蒸散发（ET）估算。 为此，我们采用**欧空局WorldCover 2021**数据集（https://worldcover2021.esa.int/）作为作物分类的真值标签：原始数据集空间分辨率为10米，通过最大投票策略重采样至45米。最终在代码中生成二值地图，仅将农田类别标记为1，其余类别标记为0。 在ET模型训练阶段，我们采用**ECOSTRESS**任务数据作为参考基准。ECOSTRESS基于普里斯特利-泰勒喷气推进实验室（Priestley-Taylor Jet Propulsion Laboratory, PT-JPL）方法生成ET数据，原始空间分辨率约为70米，为与输入的高光谱数据保持一致，将其重采样至45米。 本数据集包含14个数据块，为上述各类数据的堆叠组合，总通道数为54，通道顺序如下： 1至50通道：高光谱波段 51通道：重采样至45米分辨率的欧空局WorldCover 2021数据 52通道：蒸散发胁迫指数（ESI） 53通道：质量控制波段 54通道：蒸散发（ET） 数据文件夹中同时提供原始ESI与LULC数据。