Issue #2 — Foundation Models, LLM Agents & Operational Flood Mapping第 2 期 — 基础模型、LLM 智能体与业务化洪水制图TEST ISSUE测试版
Curated, ranked reading on hydrology, remote sensing, natural hazards & AI.
围绕水文、遥感、自然灾害与人工智能的精选排序阅读。
Issue期号#2
Coverage window覆盖时间2026-07-11 → 2026-07-25
Pushed推送时间2026-07-25 22:26 CDT
Next issue下期预计2026-07-27
Frequency更新频率Every Mon 07:00每周一 07:00
All全部
Hydrology水文
Remote sensing遥感
Hazards灾害
Flood monitoring洪水监测
Foundation model基础模型
LLM大模型
AI agents智能体
Scientific ML科学机器学习
◆
Key Papers重要文献
5 reads, ranked by quality + relevance5 篇,按质量+相关性排序
★
Top Priority重点优先
#1
Overall rank · by quality + relevance综合排名 · 按质量+相关性
HydroAgent: Closing the Gap Between Frontier LLMs and Human Experts in Hydrologic Model Calibration via Simulator-Grounded RLHydroAgent:用"模拟器为锚"的强化学习,缩小前沿 LLM 与人类专家在水文模型率定上的差距
Preprint · Not peer-reviewed预印本 · 未同行评审
Open access开放获取
Type: Preprint · benchmark + method类型:预印本 · 基准评测 + 方法
~20 min read约 20 分钟
Keywordssimulator-grounded RLLLM model calibrationhydrologic calibration benchmarkfrontier-LLM vs expertagentic parameter tuning关键词模拟器为锚强化学习LLM模型率定水文率定基准前沿LLM对比专家智能体参数调优
Authors: Zhi Li, Songkun Yan, Jie Cao, Mofan Zhang, Anjiang Wei, Jinwoong Yoo, Yang Hong · Source: arXiv:2605.17792 · Published: 18 May 2026
作者:Zhi Li、Songkun Yan、Jie Cao、Mofan Zhang、Anjiang Wei、Jinwoong Yoo、Yang Hong · 来源:arXiv:2605.17792 · 发布:2026-05-18
Relevance 9.8/10相关性 9.8/10
Summary 简要摘要
Can an LLM agent replace a human expert at calibrating a distributed hydrologic model? The authors benchmark nine frontier agents (Claude, GPT-5, Gemini variants) on the operational CREST model, then fine-tune a small open model (Qwen3-4B) with supervised trajectories plus reinforcement learning that uses simulated NSE as a verifiable reward.
LLM 智能体能替代人类专家去率定一个分布式水文模型吗?作者在业务化的 CREST 模型上评测了九个前沿智能体(Claude、GPT-5、Gemini 等),再用 2576 条专家率定轨迹做监督微调,并以模拟得到的 NSE 作为可验证奖励做强化学习,训练出一个小型开源模型(Qwen3-4B)。
Key findings关键发现
- Frontier LLMs span NSE −0.16 to 0.75; only Opus-4.7 matched a human expert, and only on one of four gauges.
前沿 LLM 的 NSE 从 −0.16 到 0.75;只有 Opus-4.7 达到人类专家水平,且仅在 4 个站点中的 1 个。
- The gap is domain grounding, not raw model size — a small model tuned with simulator-based RL is more efficient and more physically faithful than scaling a generic LLM.
差距在于领域接地(domain grounding)而非模型规模 —— 用"模拟器强化学习"微调的小模型,比一味放大通用 LLM 更高效、更物理自洽。
- Verifiable reward (run the simulator, read the NSE) turns calibration into an RL problem the agent can actually learn.
可验证奖励(跑模拟器、读 NSE)把率定变成智能体真正学得会的 RL 问题。
- Gauges span 329–40,792 km², so results aren't a single-basin artifact.
站点面积跨 329–40,792 km²,结论不是单一流域的偶然。
Why it's worth reading为什么值得读
This is the sharpest, most quantitative answer yet to the reader request for "LLM agents for hydrologic modeling." It doesn't hype agents — it measures exactly where they fall short and shows a concrete recipe (SFT + simulator-grounded RL) to fix it. The CREST model and author list connect it directly to CREST/EF5 and HyDROS-style workflows.
这是迄今对"用 LLM 智能体做水文建模"这一读者请求最锐利、最量化的回答。它不吹捧智能体,而是精确度量其短板,并给出可复制的改进配方(监督微调 + 模拟器强化学习)。所用 CREST 模型与作者名单,直接对接 CREST/EF5 与 HyDROS 式工作流。
Relevance to your research与你研究的关联
Bullseye for AQUAH / HyDROS Lab: same CREST modeling engine, the natural-language-agent-drives-a-hydrologic-model paradigm, and a training recipe you could adapt for AQUAH's calibration step.
正中 AQUAH / HyDROS Lab:同样的 CREST 建模引擎、"自然语言智能体驱动水文模型"的范式,以及可迁移到 AQUAH 率定环节的训练配方。
Suggested reading path推荐阅读路径
Read the abstract + Fig. 1 (system diagram) first. Then the results table comparing the nine agents. Skim the RLSF training details unless you're implementing it. ~20 min.
先读摘要与图 1(系统框图),再看九个智能体对比的结果表;除非要复现,RLSF 训练细节可略读。约 20 分钟。
⚠ Risks / limitations风险 / 局限:
Preprint, not peer-reviewed. Four test gauges is still modest; "human-expert" is one reference point. Improvements are in NSE — confirm they translate to hydrologically meaningful hydrographs, not just a metric bump.
预印本,未经同行评审。4 个测试站点仍偏少;"人类专家"只是一个参照点。提升体现在 NSE —— 需确认它对应到水文上有意义的过程线,而非仅是指标上升。
Read original ↗阅读原文 ↗
·
Download PDF ↓下载 PDF ↓
#2
Overall rank · by quality + relevance综合排名 · 按质量+相关性
Operational SAR flood mapping as a full-stack systems problem: an AI-enabled perspective把业务化 SAR 洪水制图当作"全栈系统"问题:一个 AI 赋能的视角
Peer-reviewed同行评审
Open access开放获取
Type: Perspective (Hypothesis & Theory)类型:观点(Hypothesis & Theory)
~15 min read约 15 分钟
Keywordsoperational SAR flood mappingfull-stack systems viewdata-to-decision pipelineAI-enabled inundation detectionlatency and provenance关键词业务化SAR洪水制图全栈系统视角数据到决策流水线AI赋能淹没检测时效与数据溯源
Author: Qing Yang (CISESS / ESSIC, University of Maryland) · Source: Frontiers in Water · Published: 10 July 2026
作者:Qing Yang(CISESS / ESSIC,马里兰大学)· 来源:Frontiers in Water · 发布:2026-07-10
Relevance 9.5/10相关性 9.5/10
Summary简要摘要
This perspective argues that operational SAR flood mapping is no longer mainly an algorithm problem — it's a full-stack systems problem spanning data access, preprocessing, ancillary layers, compute, validation and maintenance. It presents SIENA, an AI-agent-assisted rebuild of a failing operational workflow now producing CONUS/Alaska flood products multiple times daily.
本文主张:业务化 SAR 洪水制图已不再主要是算法问题,而是一个横跨数据获取、预处理、辅助数据、算力、验证与维护的"全栈系统"问题。文中给出 SIENA —— 一个由 AI 智能体协助、对失效业务流程进行重建的系统,现已能对 CONUS/阿拉斯加多次/日出图。
Key findings关键发现
- The field's bottleneck has shifted from algorithmic novelty to system bridging: mature SAR archives + rising agency demand mean the missing piece is engineering, not a better classifier.
领域瓶颈已从"算法新颖性"转向系统衔接:SAR 存档成熟、机构需求上升,缺的是工程而非更强的分类器。
- AI/LLM agents help as builders — literature synthesis, code generation, workflow assembly — under expert supervision, not as replacement flood classifiers.
AI/LLM 智能体作为"建造者"发挥作用 —— 文献综合、代码生成、流程搭建 —— 在专家监督下进行,而非替代洪水分类器。
- SIENA cut memory needs from ~256 GB to 64 GB and reduced false alarms (high precision, lower recall) versus the legacy workflow.
SIENA 把内存需求从约 256 GB 降到 64 GB,并较旧流程降低误报(高精确率、较低召回)。
- Validated across four settings (atmospheric-river West Coast, Kentucky flood, Central Valley, a 3-event DSWX-S1 comparison).
在四种场景验证(西海岸大气河、肯塔基洪水、中央谷地、三事件 DSWX-S1 对比)。
Why it's worth reading 为什么值得读
You asked specifically for operational SAR flood mapping — not another method paper — and this is exactly that: a candid systems/deployment account of what actually breaks in production and how an AI-assisted rebuild fixed it. Rare to see the un-glamorous engineering (preprocessing placement, memory, maintenance) treated as first-class.
你明确想看"业务化 SAR 洪水制图"而非又一篇方法论文 —— 本文正是如此:坦诚讲述生产环境中真正会崩的环节,以及 AI 协助重建如何修复。把不"光鲜"的工程(预处理位置、内存、维护)当作一等公民,实属难得。
Relevance to your research与你研究的关联
Directly relevant to flood-monitoring operationalization and to any AQUAH-style "agent assembles the workflow" ambition. The precision-over-recall design choice and the validation-across-events framing are reusable for your flash-flood datasets.
与洪水监测业务化、以及任何"智能体搭建工作流"(AQUAH 式)的目标直接相关。"以精确率优先于召回"的设计取舍,以及"跨事件验证"的框架,可复用于你的山洪数据集。
Suggested reading path推荐阅读路径
Read the abstract and the three-phase history, then the "full-stack" framework figure and the SIENA case study; the design-tensions section is the payoff. ~15 min.
先读摘要与"三阶段历史",再看"全栈"框架图与 SIENA 案例;"设计张力"一节是精华。约 15 分钟。
⚠ Risks / limitations风险 / 局限:
A single-author perspective built around one system (SIENA) and mostly US scenes — persuasive but not a controlled benchmark. "Conservative" mapping means real floods can be missed (lower recall); weigh that for warning use.
单作者观点,围绕单一系统(SIENA)且多为美国场景 —— 有说服力但非受控基准。"保守"制图意味着可能漏报真实洪水(召回偏低);用于预警时需权衡。
Read original ↗阅读原文 ↗
·
Download PDF ↓下载 PDF ↓
+
Strong Selections优选推荐
#3
Overall rank综合排名
Zero-shot forecasting of streamflow using time series foundation models: are we there yet?用时间序列基础模型做径流的零样本预测:我们到那一步了吗?
Peer-reviewed同行评审
Open access开放获取
Type: Research article类型:研究论文
~18 min read约 18 分钟
Keywordstime-series foundation modelszero-shot streamflow forecastingMOIRAI Chronos TTM Sundial531 CAMELS basinssub-daily benchmark关键词时序基础模型零样本径流预测MOIRAI/Chronos/TTM/Sundial531个CAMELS流域次日尺度评测
Authors: Alexander Y. Sun, Albert A. Sun · Source: Machine Learning: Earth, 2(1) · Published: 20 March 2026
作者:Alexander Y. Sun、Albert A. Sun · 来源:Machine Learning: Earth, 2(1) · 发布:2026-03-20
Relevance 9.3/10相关性 9.3/10
Summary简要摘要
A careful test of whether general-purpose time-series foundation models (MOIRAI, Chronos, TTM, Sundial) can forecast streamflow with no hydrological training. Across 531 CAMELS basins at daily and 3-hourly steps, the best zero-shot model nearly matches an LSTM trained on CAMELS — for univariate forecasting.
一项严谨的检验:通用时间序列基础模型(MOIRAI、Chronos、TTM、Sundial)在完全不用水文数据训练的情况下,能否预测径流。在 531 个 CAMELS 流域、日与 3 小时步长上,最好的零样本模型在"单变量"预测上几乎追平了在 CAMELS 上训练的 LSTM。
Key findings关键发现
- Univariate zero-shot: Sundial reached median NSE 0.564 vs. a domain-trained LSTM's 0.593 — strikingly close for a model that never saw streamflow.
单变量零样本:Sundial 中位 NSE 达 0.564,对比领域训练 LSTM 的 0.593 —— 对一个从未见过径流的模型而言差距惊人地小。
- But multivariate zero-shot (with meteorological forcings) lags badly; fine-tuning TTM with static attributes recovers it (NSE 0.753).
但多变量零样本(含气象强迫)明显落后;用静态属性微调 TTM 可恢复(NSE 0.753)。
- Performance is hydroclimate-dependent: better in snowmelt-driven, highly seasonal basins; worse in low-seasonality regions.
表现依赖水文气候:融雪主导、强季节性流域更好,低季节性地区更差。
- TSFMs stayed robust under domain shift — encouraging for transfer to ungauged basins.
TSFM 在域漂移下仍稳健 —— 对迁移到无资料流域是利好信号。
Why it's worth reading为什么值得读
The most rigorous, honest read on the exact question behind your request: can foundation models solve ungauged/data-scarce prediction? The answer — "promising for univariate, not yet for multivariate" — is more useful than any hype piece, and the CAMELS setup is the field's gold standard.
针对你请求背后核心问题——基础模型能否解决无资料/资料稀缺预测——迄今最严谨、最诚实的解读。答案"单变量有希望、多变量尚未"比任何炒作都更有用;CAMELS 设置是该领域的金标准。
Relevance to your research与你研究的关联
Core to your remote-sensing-precipitation + streamflow work and to any planetary-scale hydrology foundation-model plan. The univariate-vs-multivariate gap tells you where forcing data (e.g. satellite precip) still buys you the most.
与你的遥感降水 + 径流工作、以及任何行星尺度水文基础模型计划高度相关。"单变量 vs 多变量"的差距,指明了强迫数据(如卫星降水)当前最能带来增益的地方。
Suggested reading path推荐阅读路径
Read the abstract, then jump to the NSE comparison table (zero-shot vs LSTM) and the hydroclimate-stratified results figure; the fine-tuning section shows the ceiling. ~18 min.
先读摘要,直接看 NSE 对比表(零样本 vs LSTM)与按水文气候分层的结果图;微调一节展示上限。约 18 分钟。
⚠ Risks / limitations风险 / 局限:
CAMELS is US-centric and relatively well-behaved — global/tropical generalization is untested here. "Comparable to LSTM" is for univariate only; don't over-read it as foundation models being ready for operational multivariate forecasting.
CAMELS 以美国为主、流域相对"规整" —— 全球/热带泛化未在此检验。"追平 LSTM"仅限单变量;勿据此高估基础模型已可用于业务化多变量预测。
Read original ↗阅读原文 ↗
·
Download PDF ↓下载 PDF ↓
#4
Overall rank综合排名
Land cover and flood type govern the detection limits of satellite-based flood mapping across diverse global flood events土地覆盖与洪水类型决定了卫星洪水制图的检测极限:面向全球多样洪水事件
Preprint · Not peer-reviewed预印本 · 未同行评审
Open access开放获取
Type: Evaluation类型:评测
~15 min read约 15 分钟
Keywordsland-cover detection limitsflood-type sensitivityglobal flood event evaluationsatellite flood-mapping stress-testgeneralization assessment关键词土地覆盖检测极限洪水类型敏感性全球洪水事件评测卫星洪水制图压力测试泛化能力评估
Authors: V. Kolluru, R. Shinde, A. Marouane, C. Helbling, D. Shah, O. Drew, I. Gurung, M. Maskey, R. Ramachandran (NASA IMPACT) · Source: arXiv:2606.07780 · Published: 5 June 2026
作者:V. Kolluru、R. Shinde、A. Marouane、C. Helbling、D. Shah、O. Drew、I. Gurung、M. Maskey、R. Ramachandran(NASA IMPACT)· 来源:arXiv:2606.07780 · 发布:2026-06-05
Relevance 8.6/10相关性 8.6/10
Summary简要摘要
A stress-test of a geospatial foundation model on 19 out-of-distribution flood events (2017–2025) across six continents, eight climate zones and six flood mechanisms. Detection accuracy depends jointly on land cover and flood type — and some apparent "failures" are really disagreements between reference products.
对一个地理空间基础模型的压力测试:覆盖 6 大洲、8 个气候带、6 种成因的 19 个分布外洪水事件(2017–2025)。检测精度同时取决于土地覆盖与洪水类型 —— 而部分表面上的"失败"其实是不同参考产品之间的口径不一致。
Key findings关键发现
- Cropland is easiest (IoU ≈ 52%) and riverine floods identify well (F1 = 0.69); tree cover and built-up areas are near-blind (IoU ≈ 4%).
农田最易(IoU≈52%),河流型洪水识别良好(F1=0.69);树冠与建成区近乎"失明"(IoU≈4%)。
- 23 distinct failure modes catalogued — and pipeline engineering, not model capacity, was the bigger early constraint (echoing article #2).
归纳出 23 种失败模式 —— 且早期主要瓶颈是流水线工程而非模型容量(与第 2 篇呼应)。
- "Definitional inconsistency between reference products" inflates apparent error — a warning about how flood benchmarks are scored.
不同参考产品之间的"定义不一致"会放大表面误差 —— 提醒我们洪水基准是如何被评分的。
- Out-of-distribution, multi-continent design makes this a real generalization test, not an in-sample demo.
分布外、跨大洲的设计使其成为真正的泛化检验,而非样本内演示。
Why it's worth reading为什么值得读
A refreshingly skeptical companion to the foundation-model hype: it says exactly where satellite flood mapping breaks (canopy, cities) and why benchmark scores can lie. Pairs naturally with article #2's systems view and #3's foundation-model evaluation.
对基础模型热潮而言,这是一份令人耳目一新的"怀疑派"补充:明确指出卫星洪水制图在哪里会失效(树冠、城市),以及为何基准分数会"说谎"。与第 2 篇的系统视角、第 3 篇的基础模型评测天然互补。
Relevance to your research与你研究的关联
Directly informs flash-flood mapping and any use of geospatial foundation models for water extent. The land-cover-stratified error profile tells you where to distrust an automated flood product before it reaches a warning.
直接服务于山洪制图,以及任何用地理空间基础模型做水体范围提取的工作。按土地覆盖分层的误差画像,能告诉你在自动洪水产品进入预警前,哪些地方不该轻信。
Suggested reading path推荐阅读路径
Read the abstract, then the per-land-cover / per-flood-type results table and the failure-mode taxonomy; skip model-architecture details. ~15 min.
先读摘要,再看按土地覆盖/洪水类型的结果表与失败模式分类;模型架构细节可略。约 15 分钟。
⚠ Risks / limitations风险 / 局限:
Preprint, not peer-reviewed. Evaluates one foundation model family, so "detection limits" are partly model-specific. Metrics (IoU/F1) depend on the very reference products the paper shows to be inconsistent — read the numbers as directional.
预印本,未经同行评审。仅评测一类基础模型,故"检测极限"部分是模型特定的。指标(IoU/F1)依赖于本文指出"口径不一致"的那些参考产品 —— 数值宜作方向性参考。
Read original ↗阅读原文 ↗
·
Download PDF ↓下载 PDF ↓
#5
Overall rank综合排名
Probabilistic Sub-Daily Streamflow Forecasting in Data-Scarce Catchments Through a Zero-Shot Inductive Transfer Learning Approach用零样本归纳式迁移学习做资料稀缺流域的概率性次日尺度径流预测
Peer-reviewed同行评审
Paywalled (not OA)付费墙(非开放获取)
Type: Research article类型:研究论文
~12 min read约 12 分钟
KeywordsTemporal Fusion Transformerzero-shot inductive transferdata-scarce catchmentsprobabilistic sub-daily forecastingungauged basin transfer关键词时序融合Transformer零样本归纳迁移资料稀缺流域概率性次日预测无资料流域迁移
Authors: Heidari et al. · Source: JAWRA (J. American Water Resources Association), 2026 · Published: 2026 (see article)
作者:Heidari 等 · 来源:JAWRA(J. American Water Resources Association),2026 · 发布:2026(以原文为准)
Relevance 8.3/10相关性 8.3/10
Summary简要摘要
Targets sub-daily streamflow in catchments with little or no discharge record. A Temporal Fusion Transformer is transferred zero-shot from data-rich basins and produces probabilistic (not just point) forecasts, so users get uncertainty bands where data are thin.
面向几乎没有流量记录的流域做次日尺度径流预测。用 Temporal Fusion Transformer 从资料丰富流域做零样本迁移,并输出概率性(而非仅点值)预报,使数据稀薄处也能给出不确定性区间。
Key findings关键发现
- A zero-shot transferred TFT can produce useful sub-daily forecasts in newly- or sparsely-gauged catchments.
零样本迁移的 TFT 可在新设站/稀疏站流域给出有用的次日尺度预报。
- Probabilistic outputs (uncertainty bands) are the point — arguably more honest than a single number when local data are scarce.
概率性输出(不确定性区间)是重点 —— 在本地数据稀缺时,比单一数值更诚实。
- Even a small amount of local data further sharpens predictions in nearby ungauged basins.
即使少量本地数据,也能进一步提升邻近无资料流域的预测。
Why it's worth reading为什么值得读
A concrete, transfer-learning-first answer to your ungauged-basins request, and a useful contrast to article #3: here the emphasis is sub-daily and probabilistic, which matters for flash-flood timing. Included for topical fit even though we could not access the full text.
针对你"无资料流域"请求、以迁移学习为核心的具体回答,也是对第 3 篇的有益对照:这里强调次日尺度与概率性,对山洪的时间性尤为重要。因切题而收录,尽管我们未能获取全文。
Relevance to your research与你研究的关联
Sub-daily + probabilistic is the right shape for flash-flood datasets and warning lead-time work; the donor-basin transfer idea is directly reusable.
次日尺度 + 概率性,正契合山洪数据集与预警提前量工作;"供体流域迁移"的思路可直接复用。
Suggested reading path推荐阅读路径
Read the abstract and the probabilistic-forecast figure (coverage of the uncertainty band). Skim TFT internals. ~12 min. Note: paywalled — the summary here is based on abstract/metadata only.
读摘要与概率预报图(不确定性区间的覆盖情况),TFT 内部可略读。约 12 分钟。注意:付费墙 —— 此处摘要仅据摘要/元数据整理。
⚠ Risks / limitations 风险 / 局限:
Paywalled (no open PDF); our summary rests on the abstract, so treat specifics cautiously. Data-scarce validation is inherently hard to trust — check how "ungauged" was simulated and whether it was tested across regions.
付费墙(无开放 PDF);本摘要基于摘要页,具体细节请谨慎对待。资料稀缺场景的验证本就难以取信 —— 需查看"无资料"是如何被模拟的,以及是否做了跨区域检验。
Read original ↗阅读原文 ↗
≡
Briefs关键资讯
6 short items — scan, don't study6 条短讯 —— 供快速扫读
AI
Moonshot AI releases Kimi K3, the largest open-weight LLM yet月之暗面发布 Kimi K3,迄今最大的开源权重大模型
A ~2.8T-parameter MoE (16 of 896 experts fire per token) with a 1M-token context and native multimodality; full weights follow on Jul 27 — a serious open base Earth/geo teams can fine-tune.
约 2.8 万亿参数的 MoE(每 token 仅激活 896 个专家中的 16 个),具备 100 万 token 上下文与原生多模态;7 月 27 日放出完整权重 —— 为地球/地学团队提供可微调的强开源底座。
AI
Alibaba previews Qwen3.8-Max, a 2.4T multimodal model阿里预览 Qwen3.8-Max,2.4 万亿参数多模态模型
Alibaba's first >1T multimodal Qwen (sparse MoE; text, image, video, documents; 1M context) landed days after Kimi K3 — though it shipped as a preview with no benchmarks or model card.
阿里首个 >1T 的多模态 Qwen(稀疏 MoE;文本、图像、视频、文档;100 万上下文),紧随 Kimi K3 —— 但以预览形式发布,暂无基准与模型卡。
Earth observation对地观测
NASA releases the final Landsat 10 spacecraft Request for ProposalNASA 发布 Landsat 10 卫星正式招标(RFP)
The RFP (proposals due Aug 13; launch readiness no later than Dec 2031) continues the ~50-year Landsat land-imaging record — vital for long-term continuity of the water, snow and land data streams the community relies on.
该招标(提案截止 8 月 13 日;不晚于 2031 年 12 月具备发射条件)延续约 50 年的 Landsat 陆地成像记录 —— 对水、积雪与陆地数据流的长期连续性至关重要。
Policy · funding政策 · 经费
WRDA 2026 advances with drought-resilience & flood-risk programsWRDA 2026 推进,新增抗旱与洪水风险管理项目
The House T&I Committee approved the Water Resources Development Act of 2026, adding two new continuing-authority programs for drought resilience and flood-risk management and reauthorizing EPA's State Revolving Funds — a major US water-infrastructure signal.
美国众议院交通与基础设施委员会通过《2026 水资源开发法案》,新设抗旱与洪水风险管理两项常设授权项目,并重新授权 EPA 的州循环基金(SRF)—— 美国水基础设施领域的重大信号。
Report · attribution报告 · 归因
World Weather Attribution: warming is driving Europe's drought世界天气归因:变暖正在加剧欧洲干旱
A rapid study found climate change made 2026 Europe's agricultural soil drought ~5× more likely in the west and ~11× in the east, with the extreme evaporative conditions driving it about 80× more probable — timely and quotable.
一项快速归因研究发现,气候变化使 2026 年欧洲农业土壤干旱在西部出现的可能性提高约 5 倍、东部约 11 倍,其背后的极端蒸发条件概率约提升 80 倍 —— 及时且可引用。
Report报告
ISciences Worldwide Water Watch List, July 2026ISciences 全球水情观察清单(2026 年 7 月)
The monthly global outlook flags a possibly record-strength building El Niño, a historic European water emergency, and a drought-to-flood reversal across Central Asia / the Middle East — a concise scan of global hydrological anomalies.
该月度全球展望提示可能创纪录的厄尔尼诺正在形成、欧洲历史性水危机,以及中亚/中东"旱转涝"反转 —— 全球水文异常的简明扫读来源。
✦
Frontier Concepts前沿概念
3 concepts, each with a reading order3 个概念,各附阅读顺序
1 · Time-Series Foundation Models (TSFMs) for streamflow1 · 用于径流的时间序列基础模型(TSFM)
Definition: A model pretrained on billions of generic time series that can forecast a new signal (e.g. streamflow) with little or no task-specific training ("zero-shot").
定义:在数十亿条通用时间序列上预训练的模型,无需(或几乎无需)针对具体任务再训练,即可预测一个"新"信号(如径流),即"零样本"。
Why it matters: it's a candidate answer to prediction in ungauged/data-scarce basins — the model brings knowledge from other domains.
为何重要:它是"无资料/资料稀缺流域预测"的候选解 —— 模型把其他领域的知识迁移过来。
Suggested order建议顺序
- Start with the plain-English intro to foundation models for Earth science — 先读地球科学基础模型的通俗综述 —— Foundation Models for Environmental Science: A Survey (arXiv:2504.04280) ↗
- Then read this issue's article #3 to see how TSFMs actually score against a trained LSTM on CAMELS.然后读本期第 3 篇,看 TSFM 在 CAMELS 上与训练好的 LSTM 究竟差多少。
- Finally, browse the Chronos time-series model to try one yourself — 最后浏览 Chronos 时间序列模型,动手试一个 —— amazon-science/chronos-forecasting ↗
2 · Prediction in Ungauged Basins (PUB) & regionalization2 · 无资料流域预测(PUB)与区域化
Definition: Estimating streamflow where no gauge exists by transferring information from gauged "donor" catchments (similarity-, physics-, or learning-based).
定义:在没有观测站的地方,通过从有观测的"供体"流域迁移信息来估计径流(基于相似性、物理或学习)。
Why it matters: most of the planet's rivers are ungauged; this is the core generalization test for any hydrologic AI.
为何重要:地球上大多数河流没有观测;这是任何水文 AI 的核心泛化能力检验。
Suggested order建议顺序
- Read the classic framing of the PUB problem and regionalization — 先读 PUB 问题与区域化的经典框架 —— AI in Hydrological Modeling for Ungauged Watersheds: A Review ↗
- Then this issue's #5 (transfer learning for data-scarce, sub-daily forecasting).然后读本期第 5 篇(面向资料稀缺、次日尺度预测的迁移学习)。
- Optional deep dive: the CAMELS large-sample dataset the whole field benchmarks on — 进阶(可选):全领域基准所用的 CAMELS 大样本数据集 —— CAMELS dataset ↗
3 · LLM agents for scientific workflows3 · 用于科学工作流的 LLM 智能体
Definition: An LLM that doesn't just answer questions but plans, calls tools/simulators, runs code, and iterates toward a goal (e.g. calibrating a model).
定义:不只回答问题,而是会规划、调用工具/模拟器、运行代码并朝目标迭代的 LLM(如率定一个模型)。
Why it matters: it's how "AQUAH-style" automation of hydrologic modeling gets built — and where it still fails.
为何重要:这正是"AQUAH 式"水文建模自动化的构建方式 —— 也是它目前仍会失败的地方。
Suggested order建议顺序
- Warm up with the survey of LLM-based scientific agents — 先用 LLM 科学智能体综述热身 —— A Survey of LLM-based Scientific Agents (arXiv:2503.24047) ↗
- Then read this issue's #1 (HydroAgent) to see agents on a real calibration task, and #2 for the deployment reality.然后读本期第 1 篇(HydroAgent)看智能体在真实率定任务上的表现,再读第 2 篇看落地现实。
- Hands-on: the DL4Water hydrology-agent tutorial — 上手:DL4Water 的水文智能体教程 —— LLM Agent in Hydrology (DL4Water) ↗
Appendix附录
Past issues: browse the full archive ↗往期内容:浏览全部往期 ↗
Feedback log反馈日志
-
Issue #2 (2026-07-25) — responding to @hydros-ou (Issue #1 notes): (1) Ungauged basins × foundation models / transfer learning → now the central thread (articles #3 & #5). (2) Operational SAR flood mapping, not just method papers → featured as a Top Priority (#2), a systems/deployment view. (3) LLM AI agents for hydrologic modeling → the #1 pick benchmarks frontier LLM agents on real model calibration, with an ordered beginner path in Frontier Concepts. (4) Shorter summaries — abstracts trimmed to 2–3 sentences, Key Findings kept. (5) Frontier Concepts now give an ordered "read this first, then this" path.
第 2 期(2026-07-25)—— 回应 @hydros-ou(第 1 期反馈):(1)无资料流域 × 基础模型 / 迁移学习 → 作为本期主线(第 3、5 篇)。(2)业务化 SAR 洪水制图(不只是方法论文)→ 列为 Top Priority(第 2 篇),侧重系统与落地视角。(3)用于水文建模的 LLM 智能体 → 第 1 篇即在真实模型率定上评测前沿 LLM 智能体,并在"前沿概念"给出入门顺序。(4)更短的摘要 —— 摘要压缩到 2–3 句,保留"关键发现"。(5)"前沿概念"给出"先读哪个、再读哪个"的顺序。
Coverage & method note覆盖范围与方法说明
Window: 2026-07-11 → 2026-07-25 (strict 14 days). A strict 14-day search surfaced too few genuinely strong, on-target items — and rather than pad the issue with weak matches, we widened the search to recent months so we could deliver strong material on exactly the topics requested. Every featured article shows its true publication date. Dates range from March to July 2026; two items are clearly-marked preprints. We'd rather be transparent than fake freshness.
窗口:2026-07-11 → 2026-07-25(严格 14 天)。严格的 14 天窗口内真正高质量且切题的内容偏少。为不以弱相关内容凑数(宁缺毋滥),我们把检索范围放宽到近几个月,以便就明确请求的主题给出扎实材料。下方每篇入选论文都标注了真实发布日期,时间跨度为 2026 年 3—7 月;其中两篇为明确标注的预印本。我们宁可如实说明,也不伪装"最新"。
Disclaimer. Content is curated and summarized by an automated agent — we summarize and interpret only; no article text is reproduced, and all links go to original sources whose authors and publishers hold all rights. Preprints are marked and are not peer-reviewed. Scores are editorial judgments, not the venues'. Please verify claims against the originals before citing.
免责声明。本刊内容由自动化 agent 整理与摘要 —— 我们只做摘要与解读,不复制文章正文;所有链接均指向原始来源,版权归原作者与出版方所有。预印本已标注且未经同行评审。评分为编辑判断,非期刊官方评分。引用前请对照原文核实。
Hydrology × AI Weekly Reading · Issue #2 · Coverage 2026-07-11 → 2026-07-25 · Pushed 2026-07-25 22:26 CDT · Next issue 2026-08-01 · Updated every Monday 07:00 America/Chicago.
水文 × AI 每周精读 · 第 2 期 · 覆盖 2026-07-11 → 2026-07-25 · 推送 2026-07-25 22:26 CDT · 下期 2026-08-01 · 每周一 07:00(America/Chicago)更新。