For decades, we have tended to think about geographic information in terms of maps. We collect data, download files, copy them onto computers, process them with specialized software, and eventually produce something people can see: a map of broadband availability, watersheds, roads, parcels, population, utilities, or community assets. The map is useful, sometimes extraordinarily so, but the map is not the infrastructure. The infrastructure is everything beneath it: the data and its provenance, the standards that describe it, the systems that store it, the networks that move it, the computing resources that analyze it, and the interfaces through which people ask questions.

That distinction resonates with me because much of my life and career has involved finding useful signals within much larger systems of information. I grew up on a farm near Pickford in Michigan’s Eastern Upper Peninsula, where geography is not an abstraction. Distance shapes how people reach school, work, healthcare, markets, and one another. Later, working in signals intelligence in the U.S. Air Force taught me another lesson: collecting information is not the same as understanding it. My subsequent work in information technology, networking, data science, education, and community technology has reinforced the same idea. The challenge is rarely just obtaining more data. It is connecting the right information, understanding its context, and extracting something useful from it.

Cloud computing gives us an opportunity to rethink geospatial infrastructure from that perspective. For rural places like the Eastern Upper Peninsula, that shift could be far more consequential than simply producing faster or more attractive maps. Governments, universities, federal agencies, Tribal Nations, utilities, nonprofits, researchers, schools, libraries, and community organizations collectively hold or provide access to enormous amounts of information about this place. At the national level, the U.S. Geological Survey makes data about elevation, lidar, hydrography, watersheds, orthoimagery, land cover, structures, transportation, governmental boundaries, and geographic names available through The National Map and related services.1 Increasingly, our problem is not whether data exists. The harder problem is connecting that information efficiently enough that communities can use it.

Stop Moving So Much Data

The traditional geospatial workflow remains surprisingly file-centric. We find a dataset, download it, make a local copy, transform it, combine it with another dataset, generate another file, and perhaps publish the result. Another organization downloads its own copy. A researcher creates another. Someone saves one to a network drive. Six months later, someone repeats the process with a newer version. The result is not simply geographic information. It is an expanding collection of geographic copies whose age, provenance, and relationship to authoritative sources can become increasingly difficult to understand.

Having spent much of my career around networks, this workflow has always struck me as an infrastructure problem as much as a GIS problem. Networks teach us to pay attention not only to what information exists, but also to where it resides, how it moves, how often it moves, and whether that movement is necessary. My graduate work in applied data science added another dimension: computation is also a resource, and sound analytical design requires us to be deliberate about which data is read, transformed, joined, and retained. Cloud-native geospatial computing brings those perspectives together.

I encountered these questions firsthand while contributing to a National Science Foundation-supported Internet Measurement Research project involving the University of Michigan and Merit Network. My role centered primarily on the research data infrastructure behind the analysis. I contributed to the software and data pipeline used to integrate the FCC Broadband Serviceable Location Fabric with Broadband Data Collection information, helping transform large, fragmented broadband datasets into research-ready infrastructure. The resulting analytical environment contained approximately one billion records. Working at that scale pushed me deeper into DuckDB, Parquet, Python, SQL, and the practical realities of data engineering, where architectural decisions begin to matter as much as individual algorithms.2

That infrastructure supported research by Amy Stuyvesant, Stilian Stoev, and George Michailidis, reported in Towards Unbiased Inference of Internet Broadband Availability Based on Observational Speed Test Data. Their paper acknowledges my technical contribution to part of the software pipeline joining the FCC Fabric and Broadband Data Collection datasets. The distinction is important: the statistical methodology and research findings belong to the paper’s authors. My role was helping build the data infrastructure that made large-scale integration and analysis possible.2

The experience changed how I think about data infrastructure. At a billion-record scale, questions that may appear to be implementation details become architectural decisions. How should information be organized? Which attributes need to be read? Where should joins occur? Which intermediate products should be retained? How do we maintain the connection between a derived result and the source observations behind it? How can the infrastructure support another analysis without requiring someone to rebuild everything from the beginning? The lesson I took from IMR was not simply that we can process very large datasets. It was that the organization of the data helps determine which questions become practical to ask.

The traditional desktop model made considerable sense when geographic information systems revolved primarily around local computers and file systems. It becomes harder to justify as the size and variety of geospatial information expands. Satellite and aerial imagery, elevation models, lidar, parcel information, transportation networks, environmental observations, broadband data, weather information, and streams generated by sensors and connected infrastructure can become enormous. Repeatedly transferring all that information to ask a relatively small question can be inefficient in ways that affect storage, bandwidth, processing time, cost, and, ultimately, who possesses the technical capacity to use the information.

Cloud-native geospatial architecture suggests a different approach: move the question toward the data rather than continually moving the data toward the question. If I want to examine broadband conditions in one township in the Eastern Upper Peninsula, I may not need a statewide dataset sitting on my computer. I may not need every attribute in the dataset, and I may not need its highest available resolution. I need the geography relevant to the question, the attributes relevant to the analysis, and enough detail to support the decision at hand.

This principle is not merely aspirational. The Open Geospatial Consortium’s Cloud Optimized GeoTIFF standard uses tiled imagery, reduced-resolution representations, geographic referencing, and HTTP range requests so compatible software can retrieve only the portions of an online raster needed for visualization or processing. Instead of downloading an entire image before work can begin, an application can request the relevant internal portions of the file.3

A similar transition is occurring for vector information. Apache Parquet is a column-oriented format designed for efficient storage and retrieval, while GeoParquet extends that model with interoperable geographic types such as points, lines, and polygons. GDAL and a growing ecosystem of geospatial and analytical tools support GeoParquet.4 With column-oriented data, analytical systems can focus on the attributes required by a particular question rather than treating every field as equally necessary.

Cataloging is another part of the architecture. The SpatioTemporal Asset Catalog, commonly called STAC, standardizes metadata structures for geographic assets and how those assets can be discovered and queried. Although STAC began with satellite imagery, it now encompasses a broader range of resources, including aerial observations, synthetic aperture radar, lidar, point clouds, elevation models, vector information, machine-learning labels, and derived products.5

The important idea is larger than any one technology: do not move, store, or process information that the question does not require. Cloud optimization should not simply mean placing conventional files in somebody else’s data center. It should mean organizing information so systems can retrieve the relevant geographic area, time period, resolution, and attributes without transferring everything else.

This is one of the clearest connections between my IMR experience and cloud-native geospatial computing. Processing data at enormous scale taught me that efficiency is not merely about making code run faster. It is also about minimizing unnecessary data movement, selecting appropriate representations, bringing computation closer to the relevant information, and building reusable analytical infrastructure so expensive operations do not have to be repeated every time someone asks another question.

From Files to Questions

There is a deeper transition hidden inside this architecture. We have historically organized much of GIS around files. Cloud-native geospatial infrastructure allows us to organize it increasingly around questions. Instead of beginning with, “Where can I download the data?” we can begin with, “What are we trying to understand?” We might want to know which homes are far from existing broadband infrastructure, which community facilities fall within a service area, where transportation barriers overlap with healthcare access, which environmental observations belong to a particular watershed, or whether infrastructure investments correspond with changing community conditions. Each question may require information maintained by entirely different organizations.

This is particularly relevant to the community-technology work I have been doing in the Eastern Upper Peninsula. When we talk about broadband or digital opportunity, the question rarely remains a telecommunications question for long. Connectivity intersects with education, healthcare, economic development, transportation, public safety, workforce development, and access to community institutions. Looking at those issues separately can obscure relationships that become more apparent when information is connected around people and place.

Modern analytical tools make this query-oriented model increasingly accessible. DuckDB can query Parquet files directly over HTTPS and from object storage without requiring users to import the complete dataset into a traditional database first.6 Apache Arrow provides a language-independent, column-oriented memory format designed for analytical operations and efficient interchange among systems. Its standardized representation supports data locality and vectorized processing and can reduce repeated conversion as information moves among analytical tools and programming environments.7

These technologies do not remove the need for thoughtful data engineering. They change what is possible. A regional analyst can potentially query selected columns from a remote vector dataset, filter records by geographic extent, join those records to local information, and create a visualization without first constructing a permanent copy of every source. The difficult problem increasingly is not drawing the map. It is discovering, connecting, interpreting, and validating the evidence.

Geography provides a powerful mechanism for doing this because nearly everything important to a community happens somewhere. A water sample is collected somewhere. A fiber route passes somewhere. A student learns somewhere. A road deteriorates somewhere. A household experiences connectivity somewhere. Hospitals, libraries, schools, towers, substations, farms, wetlands, businesses, shorelines, trails, and emergency calls all exist within geographic context. Having grown up in a rural place, I think that matters. Communities like ours experience systems spatially. Distance, density, terrain, water, weather, roads, and settlement patterns affect how infrastructure works here, often in ways that disappear when a community becomes merely another row in a spreadsheet.

The answer, however, should not be to place everything into one enormous regional database. Centralizing every dataset could recreate problems of duplication, governance, maintenance, and authority at a larger scale. A better objective is interoperability: allow information to remain with the institutions responsible for maintaining it while establishing standards and mechanisms through which it can be discovered, queried, connected, and appropriately reused. The Open Geospatial Consortium’s modern APIs are designed as building blocks for accessing, integrating, and analyzing geographic information alongside other information on the web.8 That begins to look less like building a giant map and more like building a knowledge commons about place.

Why This Matters for Rural Communities

This distinction is particularly important for rural communities because access to data is not the same as capacity to use it. Large organizations may employ GIS analysts, database administrators, developers, cloud architects, and data scientists. Small municipalities, townships, schools, libraries, nonprofits, and regional organizations generally do not have teams of specialists available for every question. Yet these organizations face increasingly sophisticated questions about broadband, transportation, housing, water, healthcare, emergency response, workforce development, land use, education, and infrastructure investment.

My experience teaching technology has reinforced this point from another direction. Giving someone access to a sophisticated tool is not the same as giving them the capacity to use it meaningfully. We should design information infrastructure so the expertise embedded in our systems can be shared while leaving room for people to contribute the knowledge that only they possess about their communities.

These are not really separate mapping problems. They are interconnected community-information problems. If some of the underlying geospatial capacity can be shared, every organization does not have to recreate the same technical infrastructure before it can participate. A township should not need to build a miniature data center to understand itself. A library should not need its own GIS department to help residents explore community information. A student should not need an expensive workstation to examine environmental observations surrounding their community. A local organization should not have to maintain terabytes of imagery simply to investigate one local question.

Shared infrastructure does not diminish the value of local expertise. It may make that expertise considerably more valuable. Less time can be spent finding, downloading, copying, cleaning, and converting files, while more time can be spent understanding what the information means in the context of a particular place. A regional platform cannot know everything a township supervisor, Tribal natural-resource specialist, librarian, teacher, utility worker, farmer, or resident knows about their community. Good infrastructure should amplify local knowledge rather than attempt to replace it.

The Map Becomes a View Into Knowledge

Once the underlying infrastructure changes, visualization changes with it. The map stops being the final repository of knowledge and becomes one view into a much larger information environment. A resident might explore a simple interactive map while a planner works in a GIS application. A researcher might issue a spatial query, a student might use a notebook, a regional organization might build a dashboard, and an application might request only the geographic features necessary for a particular task. These different interfaces do not require completely separate universes of data. They can increasingly operate against the same underlying evidence.

Artificial intelligence adds another possibility. Someone may eventually ask a regional information system, in ordinary language, where broadband gaps overlap with transportation challenges, which communities are experiencing particular infrastructure pressures, or what is known about environmental conditions within a watershed. But my experience across signals intelligence, networking, data science, and large-scale research infrastructure makes me cautious about confusing the ability to find patterns with the ability to understand them. AI does not eliminate the need for sound data infrastructure. It makes that foundation more important. An eloquent answer is not necessarily trustworthy if we cannot determine where the underlying information came from, when it was collected, what geography it describes, how it was transformed, and which organization is responsible for it.

That is why provenance must be treated as part of the infrastructure rather than as documentation added after the fact. A trustworthy system should help us understand who produced a dataset, when it was created or updated, what method was used to collect it, the geographic scale at which it is meaningful, what transformations have been applied, what restrictions accompany its use, and which authoritative source someone should consult before making an important decision. STAC’s standardized metadata for assets associated with particular places and times provides one useful building block for discovery, but provenance needs to persist through the transformations that eventually create an analysis or visualization.5 It is relatively easy to produce a convincing map. It is much harder, and more valuable, to build a system capable of showing why the map should be believed.

From Spatial Data to Community Knowledge

Imagine this approach operating across the Eastern Upper Peninsula. Environmental observations could be connected with watersheds and Great Lakes conditions without requiring every organization to maintain identical copies of every dataset. Broadband infrastructure could be examined alongside homes, schools, libraries, businesses, and other community institutions. Transportation information could be considered in relation to healthcare, workforce, education, and emergency response. Infrastructure investment could be examined alongside community outcomes over time. The purpose would not be to create one enormous map containing every conceivable layer. It would be to create infrastructure capable of assembling the appropriate evidence when someone asks a meaningful question.

For me, this is where the different strands of my experience increasingly come together. Farm life taught me that place matters. Signals intelligence taught me to look for meaningful information within noise. Networking taught me that connections and architecture matter as much as individual endpoints. Data science taught me to interrogate information rather than simply accumulate it. The IMR work taught me what those principles mean at scale: once an analytical environment reaches hundreds of millions or a billion records, data architecture is no longer something happening behind the research. It becomes part of what makes the research possible. Teaching has reminded me that knowledge becomes more valuable when people can explore and use it themselves. Working with communities across the Eastern Upper Peninsula has demonstrated repeatedly that people closest to a place possess knowledge that will never be fully captured by a remote dataset.

We would still make maps, of course. Maps remain among our most powerful tools for understanding place. But the maps would no longer be the principal asset. The asset would be the underlying capability that allows communities to ask new questions without reconstructing the entire analytical process every time. A map can help answer the question we knew to ask yesterday. Geospatial data infrastructure gives us greater capacity to investigate the question we discover tomorrow.

We sometimes describe cloud computing as though its defining characteristic were simply that someone else’s computers are storing our files. That understanding misses the more consequential opportunity. Standards and implementations such as Cloud Optimized GeoTIFF, GeoParquet, STAC, Apache Arrow, and modern geospatial APIs point toward a different relationship among data, computation, networks, institutions, and people. They can allow computation to occur closer to large datasets, permit users to retrieve only what their questions require, reduce unnecessary duplication, preserve connections to authoritative sources, and make sophisticated geographic capabilities available through many interfaces.

For rural regions such as the Eastern Upper Peninsula, the opportunity is therefore bigger than GIS modernization. Geospatial data can become part of a shared regional knowledge infrastructure: an environment in which communities can discover evidence, understand its provenance, connect information across institutional boundaries, and transform data into locally meaningful knowledge. Geography supplies an especially powerful connective layer because it gives otherwise disconnected observations something in common: place.

We should continue making good maps. But we should invest just as seriously in what lies beneath them. The map is an output. The real infrastructure is our capacity to connect evidence, computation, geography, and local knowledge so communities can better understand the places they call home.

Sources

  1. U.S. Geological Survey, The National Map. ↩

  2. Amy Stuyvesant, Stilian Stoev, and George Michailidis, Towards Unbiased Inference of Internet Broadband Availability Based on Observational Speed Test Data (2025), SSRN Working Paper No. 5372801. The paper acknowledges Jason Kronemeyer’s technical contribution to a portion of the software pipeline joining FCC Fabric and Broadband Data Collection data. ↩ ↩2

  3. Open Geospatial Consortium, Cloud Optimized GeoTIFF Standard, Version 1.0. ↩

  4. GeoParquet, GeoParquet Specification; GDAL, GeoParquet Documentation. ↩

  5. Open Geospatial Consortium, SpatioTemporal Asset Catalog Community Standard; STAC Specification, STAC Overview. ↩ ↩2

  6. DuckDB, Reading and Writing Parquet Files; DuckDB, HTTP(S) Support. ↩

  7. Apache Arrow, Columnar Format Specification. ↩

  8. Open Geospatial Consortium, OGC API Standards. ↩