將地圖選取範圍轉化為 3D 城市模型的實用流程

I build site scale models often enough that I have settled on a repeatable sequence. Nothing here requires a GIS team. It does require knowing what you are modeling for before you start, because the requirements change the workflow more than the tool does.
Step 0: Decide what good enough means
Before I open anything, I write down the tolerance. For a shade question, a half meter height error is fine. For a street wall study, the footprint matters more than the roof. Writing the tolerance down keeps me from chasing detail that will never change the answer, which is the most common way these afternoons turn into weeks.

Step 1: Define the question, then the extent

A rectangle is a scope decision. If the question is shade on one sidewalk, four blocks is plenty. If it is how a new tower changes a corridor, you need the corridor plus a buffer. I write the question down first. It stops the extent from creeping every time I see something interesting at the edge.

Step 2: Collect the reference layer

Before any geometry, I grab what is authoritative: parcel boundaries, building footprints, street centerlines, transit alignments. In Phoenix the light rail line is a hard constraint, and so is the one way grid. Having those as vectors means I can check the generated model against something real instead of trusting how it looks. If a layer is missing, I note it rather than invent it, because an absent vector is easier to explain later than a guessed one.

Step 3: Generate massing from the selection

This is where a tool like GenAnywhere fits. You draw a rectangle on the map, and the system generates a 3D model for that region using AI. The output is massing, not architecture. Roofs are flat, facades are simple, and heights are estimated from imagery. That is fine for site analysis. It is not fine for a rendering you would show a client without a disclaimer.

Step 4: Snap, scale, and sanity check

The first check is always scale. I measure a known block face against the parcel data. If the model is off by more than a few percent, something in the projection is wrong. The second check is height against known buildings. Downtown Phoenix has enough landmarks that a systematic error shows up in about ten minutes.
A quick word on polygon budget
Generated massing is heavier than it looks. Four blocks of low rise with parking structures can still run into millions of triangles once roofs and canopies are included. I merge coplanar faces and drop interior geometry I will never see. If the scene starts stuttering while I orbit, the model is telling me something about how I will use it later.

Step 5: Add the layers you actually need

A model with no data is a sculpture. I add three layers in this order: ground cover, because asphalt and shade drive the heat story; the transit alignment, because it explains traffic; and land use, because it explains why the blocks look the way they do. Air quality goes on top once I have a source I trust.

Step 6: Test one scenario, not ten

The temptation is to run a dozen scenarios. I run one: the change I am actually evaluating. In Phoenix that is usually a height increase or a shade structure. One scenario, compared against the baseline, from the same camera position. Anything more and I stop being able to read the result. The fixed camera is the whole trick, since a new viewpoint can make a change look bigger than it is.

Step 7: Export with the metadata attached

When the scene is right, I export images from fixed camera positions and write down the extent, the imagery date, and the tool version. This sounds like busywork until you come back in six months and cannot remember whether the trees were modeled or invented. The note takes two minutes.
Where this workflow breaks down
Detail. Interior courtyards, facade articulation, and real tree species are all beyond it. It is also bad at change over time, since every model is one snapshot. If those matter, you need survey grade data and a much longer schedule than an afternoon.

What you get for the effort

For block level questions, the workflow gets you to a defensible answer fast. For everything else, it tells you which questions are worth paying a survey team to answer. That is usually the more valuable output.
One more habit: keep the raw selection. When a colleague asks why the model looks a certain way, the rectangle and the generation date are the whole audit trail. Without them you are defending a picture instead of a process.