From urban systems and open data to AI-assisted Insight
Course Positioning
Course Positioning
When you meet an urban issue, look past the single event or layer to the system behind it. Then find the data, choose a visualization approach, and explore and verify with an AI Agent.
The core shift “This is not a layer,
it is a system.”
Learning Outcomes
Learning Outcomes
Expand a single phenomenon into its upstream, downstream, environment, people and facilities, and identify the system relationships that must be observed together.
Break a system down into Question → Evidence → Data, and recognize data sources, reliability, and spatial and temporal forms.
Move from observation to finding, insight and action, and understand that correlation is not causation.
Teaching Flow
Week 08 · 360-minute full version
Six Parts map to six stages of thinking. Instructors can skim each summary first, then expand to check the actual chapters, cases and student deliverables.
First understand the upstream, downstream, environmental conditions and affected parties behind a phenomenon, then decide why the data belongs together.
Main case
From rainfall, watersheds, rivers, reservoirs and intakes, through to household water use, flooding and shortage, plus upstream factories, livestock and land use.
“I want to map factories.”
“I want to know what possible pollution sources lie upstream of the water intake.”
Too much: flooding and disastersToo little: shortage and groundwaterUpstream: pollution risk
Unfold a single phenomenon into system relationships that can be discussed and transferred.
Don’t rush to data portals. First rewrite each system element as “what do I want to know, and what evidence do I need?”
Know what you are looking for, rather than just collecting whatever layers are available.
Recognize official, community, remote-sensing, sensor and unstructured data, and check source, time, coverage and reliability.
The data already has fields or a standard format. The focus is reliable ingestion, field normalization, time ordering and spatial conversion; an LLM is not always needed.
Event type, severity, validity period and Polygon are already structured; you still need to handle missing geometry, time fields and expiry rules.
The API does not always return tables; imagery also needs observation time, bbox, dimensions and format before it can be arranged into a playable time series.
Limit: An API returning data does not mean it can go straight onto the screen.
RSS, announcements, PDFs and images are readable but have no consistent fields; after AI-assisted extraction, rules, gazetteers and human review still act as gatekeepers.
Multi-source RSS → URL deduplication → AI extracts place names, category and newsworthiness → validation against the township name list → structured events.
Daily JPG → image text transcription → regex parsing → Chinese/English cross-validation → vectorized activity zones → needs_review.
Limit: AI does not decide official coordinates; results that fail quality checks must be marked as pending verification.
Official / API, community and civic, Remote Sensing, Sensor / Real-time, PDFs / announcements / news / RSS / images
Common checksSource · Update time · Coverage · Gaps · Metadata · Reproducibility · Need for manual review
List 5–10 candidate datasets, marking access method, the role of AI, Source, Spatial, Temporal and limits.
Based on the question, scale and temporal character, choose among map, chart, timeline, flow, animation and 2D / 3D.
Scale matters: A park may be a Polygon at the Taipei City scale but a Point at the Taiwan-wide scale. Not all data needs animation or 3D.
Visualization is not “draw a map whenever there are coordinates.” First decide whether the question is about location, movement, comparison, trend or composition.
Once position, time and altitude are strung into 2D / 3D trajectories, stacking many flight tracks can form a visual “fingerprint” of an airport. Overlaying restricted zones, terrain or weather then reveals detours, holding patterns and differences between conditions.
A trajectory map offers leads worth following up; it does not prove causation by itself. It still needs verification with time, regulations, weather or other data.
Data such as population, fire incidents and water use does not always need coordinates. Depending on the question, it can become KPI cards, ranking bars, time trends, composition charts or lookup tables.
Real data comes first; where data gaps or prototypes remain, label them clearly and never present a design mock-up as live conditions.
Explain why this visual form answers the question better than another.
Take an Observation through locating, comparing and interpreting until it becomes an Insight / Hypothesis, without presenting association as causation.
There are many factories here.
Factories cluster along a particular river stretch.
A cluster of high-pollution-risk facilities upstream of a water intake merits closer checks on discharge, water quality and land use.
“Worth investigating” is an actionable next step, but it does not mean causation has been proven.
Clearly mark findings, associations, possible causes and what still needs verification.
AI Agents widen the exploration space while humans keep responsibility for judgment; a group workshop then carries the Hypothesis into Week 09 verification.
Finding data · Reading APIs · OCR · Cleaning · Spatial Join · Coding · Debugging · Monitoring
Defining the question · Judging plausibility · Interpreting relationships · Avoiding far-fetched causation · Deciding the final Insight
raw data · processed data · script · prompt / rule · metadata · source · timestamp
The faster the AI, the more validation matters.Students pick one topic from urban heat, flooding, elderly mobility, food supply, public transport or urban green space, and complete three deliverables with AI.
Map the interacting system elements behind the problem.
List 5–10 datasets with their sources and spatial and temporal forms.
If I show A, B and C together, I expect to observe X.
Acquire → Transform → Visualize → Interact → Validate → Interpret
What students take away is more than a map.