The End Of Knowledge - Vault 4/80 - AI For Good - 2023 - Building a foundation for geospatial AI

graph LR classDef workshop fill:#f9d4d4, font-weight:bold, font-size:14px classDef goals fill:#d4f9d4, font-weight:bold, font-size:14px classDef needs fill:#d4d4f9, font-weight:bold, font-size:14px classDef challenges fill:#f9f9d4, font-weight:bold, font-size:14px classDef education fill:#f9d4f9, font-weight:bold, font-size:14px classDef future fill:#d4f9f9, font-weight:bold, font-size:14px A[Building a foundation
for geospatial AI] --> B[Workshop: geospatial
AI foundation. 1] A --> C[ITU started geospatial
AI in 2022. 2] A --> D[Goal: common language,
understanding. 3] A --> E[Growing data, novel
AI needed. 4] A --> F[Rapid geospatial
AI development. 5] A --> G[Mexico applies AI
for statistics. 6] B --> H[Mexico projects: experimental
statistics. 7] B --> I[Need trained people
in geospatial AI. 8] B --> J[Multidisciplinary approach for
new curricula. 9] B --> K[Concerns: quality, reliability,
verification. 10] B --> L[University survey: geospatial
AI importance. 11] C --> M[Graduate-level respondents:
significant hours. 12] C --> N[Importance: data quality,
learning, trust. 13] C --> O[Geoinformatics programs: computer
science, geospatial. 14] C --> P[Private sector survey:
skills, experience expectations. 15] C --> Q[31% confident, 69%
lack talent. 16] D --> R[62% upskilling
employees. 17] D --> S[Graduates lack required
skills. 18] D --> T[Computer science vs
geomatics background. 19] D --> U[Key skills: programming, geospatial
data, ML, DL. 20] D --> V[Desire for curated
research hub. 21] E --> W[Challenges: spatial autocorrelation,
heterogeneity. 22] E --> X[Multimodal foundation models
next phase. 23] E --> Y[Geospatial AI for all,
like search engines. 24] E --> Z[UN GGIM coordinates
geospatial management. 25] E --> AA[One UN data
hub launched. 26] F --> AB[AI insights from UN
data hub. 27] F --> AC[Smaller companies face
IP challenges. 28] F --> AD[Basic skills needed
in new hires. 29] F --> AE[Library science era
ending. 30] F --> AF[Advance multidisciplinary
work. 31] G --> AG[Encode flexibility in
systems. 32] G --> AH[Common definition
challenges. 33] G --> AI[Bias: more data from
developed regions. 34] G --> AJ[Enhance capacity in
developing regions. 35] G --> AK[Interpretability, data quality,
validation. 36] H --> AL[Build on existing
ontology work. 37] H --> AM[Geospatial AI skills
cross domains. 38] H --> AN[Atlantic Council connects
policymakers, technologists. 39] H --> AO[Key audiences: decision makers,
developers, practitioners. 40] H --> AP[Design balanced
curricula. 41] I --> AQ[Modular curriculum content
for flexibility. 42] I --> AR[Collaboration, sharing existing
content. 43] I --> AS[Working groups for curriculum
development. 44] I --> AT[Real-world challenges
in curricula. 45] I --> AU[Modernize existing geospatial
standards. 46] J --> AV[Stackable curriculum
components. 47] J --> AW[Teach limitations, uncertainties
of models. 48] J --> AX[Foster collaborations for
education. 49] J --> AY[Train geospatial data
scientists. 50] J --> AZ[Pre-trained models for
fine-tuning. 51] K --> BA[Need model hubs with
documentation. 52] K --> BB[Standards for data quality,
performance. 53] K --> BC[Integrate domain knowledge
in AI. 54] K --> BD[Ensure AI is not
a black box. 55] K --> BE[Democratize access with
easy interfaces. 56] L --> BF[Ethical considerations in
education. 57] L --> BG[Enhance, not replace,
traditional expertise. 58] L --> BH[Shift effort to higher-value
applications. 59] L --> BI[Continuous learning for
evolving field. 60] class A goals class B,C,D,E,F,G needs class H,I,J,K,L,M,N,O,P,Q challenges class R,S,T,U,V education class W,X,Y,Z,AA framework class AB,AC,AD,AE,AF future class AG,AH,AI,AJ,AK future class AL,AM,AN,AO,AP future class AQ,AR,AS,AT,AU future class AV,AW,AX,AY,AZ future class BA,BB,BC,BD,BE future class BF,BG,BH,BI future

Resume:

1.- Workshop organized by ITU and stakeholders to discuss building a foundation for geospatial AI.

2.- ITU started geospatial AI activities in 2022, including webinars and challenges.

3.- Goal is to define common language and understanding of geospatial AI concepts, principles and techniques.

4.- Geospatial data repositories are growing, putting pressure on analysis tools. Novel AI methods are needed.

5.- Rapid development of geospatial science and AI has made geospatial AI an important technique for geospatial big data analysis.

6.- Public sector in Mexico is increasingly applying AI and machine learning to generate official statistical and geospatial data.

7.- Projects in Mexico are developing alternative AI sources to generate experimental statistics, like the geospatial data cube.

8.- Growing adoption of geospatial AI in national agencies creates need for people trained in fundamentals and state of the art.

9.- New geospatial AI curricula should take a multidisciplinary approach to understand fundamentals, state of the art, and applications.

10.- Key concerns for AI-produced official data are quality assurance, reliability, and verification/validation processes.

11.- Survey sent to universities about current geospatial AI offerings and importance.

12.- Most respondents were graduate-level, dedicating significant hours, believing geospatial AI is very relevant to introduce.

13.- Strong agreement on importance of data quality, metadata, supervised/unsupervised learning, applications, performance, uncertainty, trustworthiness.

14.- Some universities have dedicated geoinformatics programs combining computer science and geospatial components.

15.- Survey sent to private sector about geospatial AI skills and experience expectations.

16.- 31% of companies are confident in AI capabilities, 69% lack geospatial AI talent and find it hard to hire.

17.- 62% focusing on upskilling employees in geospatial AI rather than hiring.

18.- Two-thirds believe recent graduates don't have required geospatial AI skill sets for jobs.

19.- Opinion split on preferring computer science vs geomatics educational background.

20.- Key skills: programming, dealing with geospatial data/databases/remote sensing, machine learning, deep learning, domain knowledge.

21.- Desire for curated geospatial AI research hub to keep up with advances instead of scattered sources.

22.- Challenges in applying traditional AI/ML to geospatial data due to spatial autocorrelation and heterogeneity.

23.- Next phase of geospatial AI is multimodal foundation models for querying and prediction.

24.- Transition from geospatial experts to geospatial AI enabling anyone to access insights as easily as using a search engine.

25.- UN GGIM supporting coordination of geospatial management across UN system via UN Geospatial Network.

26.- "One UN" data hub launched to compile geospatial assets from 40 UN agencies.

27.- Opportunity to apply AI on top of the UN geospatial data hub for new insights and applications.

28.- Smaller geospatial AI companies face IP challenges when collaborating with academia due to publishing requirements.

29.- Need for basic programming and geospatial skills in new hires to quickly contribute in fast-paced startups.

30.- Library science era of geospatial is ending, making way for AI-powered insights without needing to understand raw data/pixels.

31.- Geospatial AI efforts should advance multi-disciplinary work at intersection of machine learning, geospatial data and domain applications.

32.- Geospatial AI systems need to encode flexibility for different cultural/political/social interpretations, not just technical considerations.

33.- Lack of a common definition for domains like forests or croplands globally poses challenge for training geospatial AI.

34.- Bias in geographic data, with more data from developed regions, is an important factor for geospatial AI development.

35.- Enhancing geospatial AI capacity in developing regions like Africa is crucial for globally representative and impactful applications.

36.- Interpretability and explainability of AI models, data quality and validation are key for geospatial AI.

37.- Ontology work traditionally done in geospatial field is relevant to geospatial AI, shouldn't be reinvented but built upon.

38.- Geospatial AI skills are applicable across many domains beyond just geography/GIS as geospatial data is pervasive.

39.- Atlantic Council program connects policymakers and technologists globally to develop responsible AI practices.

40.- Identified three key audiences for geospatial AI curricula: decision makers, developers/engineers, and practitioners/analysts.

41.- Challenges in designing balanced geospatial AI curricula that incorporate new topics while maintaining necessary traditional geospatial principles.

42.- Some universities developing modular curriculum content at different depths to flexibly meet learner needs.

43.- Collaboration and sharing of existing educational content, not reinventing from scratch, is an efficient way forward.

44.- Working groups with diverse expertise can be formed to collaboratively develop geospatial AI curricula components.

45.- Geospatial AI curricula should expose learners to real-world challenges like noisy/incomplete data, not just neat pre-processed examples.

46.- Consider modernizing existing geospatial standards (e.g. from OGC, ISO) to incorporate geospatial AI considerations.

47.- Build geospatial AI curricula as stackable components from high-level overviews to in-depth technical training.

48.- Important to impart limitations and uncertainties of geospatial AI models, not just the techniques themselves.

49.- Foster collaborations between universities, public sector, private sector and international organizations to advance geospatial AI education.

50.- Train students to be "geospatial data scientists" with combined programming, big data, machine learning and geospatial skills.

51.- Growing availability of pre-trained geospatial AI models that can be fine-tuned, in addition to training new models from scratch.

52.- Need for geospatial AI model hubs akin to software package repositories, with documentation, validation, use cases.

53.- Standards and best practices for assessing and reporting quality of training data, model performance, uncertainties, fit for purpose.

54.- Integrating domain knowledge from application sectors into geospatial AI through collaborations and education.

55.- Geospatial AI should not just be a "black box" to end users but they should have sufficient literacy to use it appropriately.

56.- Democratizing access to geospatial AI through easy-to-use interfaces, not just catering to technical experts.

57.- Ethical considerations, potential for bias, fairness, transparency, accountability crucial to integrate into geospatial AI education.

58.- Geospatial AI will enhance but not replace traditional geospatial and domain expertise; need to train experts at their intersection.

59.- Progress in geospatial data infrastructures and platforms will shift effort from data wrangling to higher-value geospatial AI applications.

60.- Geospatial AI is a rapidly evolving field that will require continuous learning, updates to educational curriculum and workforce upskilling.

Knowledge Vault built byDavid Vivancos 2024