Concept Graph & Resume using Claude 3 Opus | Chat GPT4o | Llama 3:

Resume:

1.- Early warning systems are limited in the global South compared to the developed world, like the US. More work is needed.

2.- UN Secretary General Guterres mandated the community to come up with better warning systems after recent extreme weather events.

3.- Global kilometer-scale weather models can now resolve features that drive flood events, which couldn't be resolved previously.

4.- Combination of advancements in physical models, AI surrogates, observations in a meaningful way is needed for future forecasting and warning systems.

5.- Facilitating discussions between different sectors, like academia and industry, to learn from each other and develop new collaborations is important.

6.- AI weather forecasting models like ForeCast-Net and PangU-Weather are approaching or surpassing skill of traditional numerical weather prediction models.

7.- Challenges remain for AI weather models, like physics constraints, uncertainty, and building trust. ECMWF is developing their own AI model.

8.- AI can help with many aspects of Earth observation, like cloud detection, optical-radar data fusion, resolution enhancement, and prediction.

9.- Onboard AI in satellites enables real-time, reprogrammable Earth observation. Inter-satellite communication allows an observing satellite to trigger a focused observation from another.

10.- ESA is working to make AI and Earth observation more accessible through challenges and open source. Finding experts is a challenge.

11.- AI can help climate adaptation and resilience through near real-time monitoring of floods, fires, hurricanes and improved short and long-term prediction.

12.- Seasonal and multi-year climate predictions have very low confidence. Machine learning techniques show potential for improvement but more work is needed.

13.- Improving representation of key physical processes like clouds and vegetation in climate models using AI and high-resolution simulations is promising.

14.- Post-processing climate model projections with AI provides better regional details and uncertainty estimates compared to raw climate model ensembles.

15.- Gaps remain in AI for climate adaptation on seasonal to multi-year scales, empirical risk indices, and training people in both AI and climate.

16.- Visual communication of risk, like photorealistic images of future flood or landscape change, can be more effective than numbers alone.

17.- AI should be integrated along the whole early warning chain from observation to decision making, with an emphasis on communication.

18.- World Food Programme uses real-time phone surveys and machine learning models for food security early warning, now-casting, and forecasting.

19.- WFP's HungerMap provides near real-time food security monitoring in over 90 countries by integrating various data sources.

20.- WFP uses XGBoost models to estimate food security in countries without real-time data, and recurrent neural networks for 30-90 day forecasting.

21.- Input data for WFP models includes conflict fatalities, market prices, economic indicators, rainfall, NDVI. Expanding data sources is an ongoing effort.

22.- Tailoring model outputs to end-user needs, like categorical change rather than precise numbers, is important for uptake and usability.

23.- Collaborations with country offices and external researchers is valuable for WFP to get feedback, improve models, and build capacity.

24.- Internal displacement from disasters is a global issue. IDMC maintains the world's most comprehensive database on internal displacement.

25.- IDMC data comes from governments, UN, NGOs, and news monitoring. It's cleaned, harmonized and validated into annual datasets and real-time updates.

26.- Challenges for internal displacement data include variations in collection methodologies, coverage, continuity, accessibility. IDMC is working to improve these issues.

27.- IDMC data has been used with weather forecasts to anticipate displacement from tropical cyclones, and in machine learning models.

28.- Planned improvements to IDMC data include standardization, interoperability with hazard datasets, disaggregation, automated quality checks, and enhanced metadata.

29.- Explainable AI techniques like SHAP values reveal compounding effects of precipitation, wealth, and other factors on global disaster displacement risk.

30.- In Somalia, causal discovery suggests complex interactions between drought, water prices, food prices, and displacement at varying time lags.

31.- Causal model forecasts of drought displacement perform similarly to non-causal baselines but use fewer features, potentially allowing earlier warning.

32.- More work is needed to validate causal graphs of displacement and include more variables like remittances and micro-level factors.

33.- Deep learning models trained on weather and vegetation data can predict drought impacts on vegetation at high resolution months in advance.

34.- Predictive skill decreases with longer forecast horizons but still outperforms baselines. More variables like soil moisture could further improve performance.

35.- Gridded global predictions of NDVI anomalies from the models capture spatial patterns of drought impact well, with some uncertainty.

36.- Potential applications include anticipatory action like distributing seeds before a drought. Scaling to new regions requires retraining on local data.

37.- Next steps are studying compound extremes, improving temporal forecasting skill, engaging stakeholders, and making data and code openly available.

38.- Traditional fire danger models consider weather but not differences in vegetation, land use or human activity. Machine learning can help.

39.- ML models are trained on historical satellite data of fire weather, land surface, vegetation, human factors to predict large fires.

40.- The temporal and spatiotemporal ML models perform best, outperforming traditional fire weather indices and simpler ML approaches.

41.- Interpreting the ML models reveals sensible relationships between variables like temperature, humidity, NDVI and fire probability that build trust.

42.- Applied in Greece operationally, the high resolution ML fire danger maps provided more actionable information compared to coarser regional maps.

43.- Expanding to a Mediterranean-wide ML model is in progress. The larger dataset will allow predicting fire spread and extremes.

44.- Sub-seasonal and seasonal fire prediction relies on weather anomalies from numerical models, but skill is limited beyond a few weeks.

45.- Treating long-term fire prediction as a segmentation problem with deep learning shows promise, capturing spatial patterns months ahead.

46.- However, the segmentation approach doesn't consider global teleconnections and memory effects important for fire weather and climate oscillations.

47.- Transformers using global climate oscillation data and multi-scale spatial convolutions further improve long-term fire prediction compared to local-only predictors.

48.- Next steps are improving evaluation, interpreting the models, disentangling climate modes from trends, and moving beyond burned area to emissions and severity.

49.- Anomaly detection identifies rare patterns or events in data, with applications across domains like climate, health, ecology, cybersecurity, etc.

50.- Key challenges include handling high dimensional data, uncertainty, online learning, concept drift, and incorporating domain knowledge like physics constraints.

51.- Successful techniques demonstrated include: autoencoders and Gaussian processes for unsupervised anomaly detection and active learning to improve over time.

52.- Spatiotemporal anomaly detectors like the Maximally Divergent Interval method can identify anomalous and extreme weather patterns in climate data.

53.- Causal discovery and inference is important for attribution of anomalies and improved forecasting by identifying governing physical relationships.

54.- Granger causality on vector autoregressive models reveals changes in information flow between climate variables during anomalies like heatwaves.

55.- Knockoffs and causal regularization in neural networks provide more robust causal discovery than traditional methods like PC algorithm in initial tests.

56.- Predicting anomalies remains very challenging due to their rarity, nonstationarity and high-dimensionality. Some promising directions were discussed.

57.- These include: causal forecasting, continual learning, physics-guided machine learning, uncertainty quantification, human-in-the-loop, and tracking distribution shift.

58.- Engaging diverse stakeholders, considering societal impact and accountability, creating incentives, and adapting to changing data are key for AI applications.

59.- Stakeholder involvement throughout the AI development and deployment process is critical to build useful, trusted and ethical systems in high impact domains.

60.- The combination of AI, physics, human knowledge and learning systems is a major challenge and opportunity ahead for sustainability and climate action.

Knowledge Vault built byDavid Vivancos 2024