In this tutorial, we design a complete GeoAI workflow for extracting building footprints from high-resolution NAIP aerial imagery. We begin by configuring the geospatial deep learning environment, downloading raster imagery and vector labels, and inspecting their spatial properties before generating georeferenced image chips and segmentation masks. We then train a U-Net model with a ResNet-34 encoder, evaluate its learning behavior, and apply sliding-window inference to an unseen scene. Beyond semantic segmentation, we convert predicted masks into cleaned and regularized building polygons, calculate IoU and F1 metrics, explore zero-shot segmentation with Grounding DINO and SAM, and compare the results with a pretrained Mask R-CNN instance segmentation model. We also demonstrate how the same pipeline extends to real-world areas using NAIP imagery from Microsoft Planetary Computer and building labels from Overture Maps.
Copy CodeCopiedUse a different Browserimport os
import subprocess
import sys
import time
import warnings
warnings.filterwarnings(“ignore”)
IN_COLAB = “google.colab” in sys.modules
def pip_install(packages, quiet=True):
“””Install packages with pip from inside the notebook process.”””
cmd = [sys.executable, “-m”, “pip”, “install”, “–upgrade”]
if quiet:
cmd.append(“-q”)
subprocess.run(cmd + list(packages), check=False)
try:
import geoai
except ImportError:
print(“>>> Installing geoai-py and friends (takes ~2-4 minutes on Colab)…”)
pip_install(
[
“geoai-py”,
“segmentation-models-pytorch”,
“buildingregulariser”,
]
)
try:
import geoai
except Exception as e:
raise SystemExit(
f”Import failed after install ({e}).n”
“=> Runtime > Restart session, then re-run this cell. ”
“The install is cached, so it will be fast the second time.”
)
import geopandas as gpd
import matplotlib.pyplot as plt
import numpy as np
import rasterio
import torch
from rasterio.plot import plotting_extent
from IPython.display import display
print(f”geoai : {geoai.__version__}”)
print(f”torch : {torch.__version__}”)
print(f”CUDA available: {torch.cuda.is_available()}”)
if torch.cuda.is_available():
print(f”GPU : {torch.cuda.get_device_name(0)}”)
else:
print(“!! No GPU detected. Training will still run but be much slower.”)
print(” Colab: Runtime > Change runtime type > Hardware accelerator > T4 GPU”)
DEVICE = geoai.get_device()
print(f”geoai device : {DEVICE}”)
CFG = {
“tile_size”: 512,
“stride”: 256,
“buffer_radius”: 0,
“architecture”: “unet”,
“encoder”: “resnet34”,
“encoder_weights”: “imagenet”,
“num_channels”: 3,
“num_classes”: 2,
“batch_size”: 8,
“num_epochs”: 12,
“learning_rate”: 1e-3,
“val_split”: 0.2,
“window_size”: 512,
“overlap”: 256,
“run_zero_shot”: True,
“run_pretrained”: True,
“run_real_aoi”: False,
}
WORK = “/content/geoai_tutorial” if IN_COLAB else os.path.abspath(“geoai_tutorial”)
os.makedirs(WORK, exist_ok=True)
os.chdir(WORK)
print(f”working dir : {WORK}”)
def banner(text):
print(“n” + “=” * 92 + f”n {text}n” + “=” * 92)
def timed(fn, label):
“””Run fn(), report wall time, never let one step kill the notebook.”””
banner(label)
t0 = time.time()
try:
out = fn()
print(f”n[OK] {label} — {time.time() – t0:.1f}s”)
return out
except Exception as exc:
import traceback
print(f”n[SKIPPED] {label}n{type(exc).__name__}: {exc}”)
traceback.print_exc(limit=3)
return None
HF = “https://huggingface.co/datasets/giswqs/geospatial/resolve/main”
train_raster_url = f”{HF}/naip_rgb_train.tif”
train_vector_url = f”{HF}/naip_train_buildings.geojson”
test_raster_url = f”{HF}/naip_test.tif”
def step1():
train_raster = geoai.download_file(train_raster_url)
train_vector = geoai.download_file(train_vector_url)
test_raster = geoai.download_file(test_raster_url)
for p in (train_raster, train_vector, test_raster):
print(f” {os.path.getsize(p) / 1e6:8.2f} MB {p}”)
return train_raster, train_vector, test_raster
paths = timed(step1, “STEP 1 — Downloading sample NAIP imagery and building labels”)
TRAIN_RASTER, TRAIN_VECTOR, TEST_RASTER = paths
We configure the environment, install the required GeoAI and deep learning libraries, and verify GPU availability. We define the central configuration parameters for dataset creation, model training, inference, and optional processing stages. We then create the working directory, define reusable execution utilities, and download the NAIP imagery and building footprint labels.
Copy CodeCopiedUse a different Browserdef step2():
info = geoai.get_raster_info(TRAIN_RASTER)
for k, v in info.items():
print(f” {k:<16}: {v}”)
print(“n— per-band statistics —“)
print(geoai.get_raster_stats(TRAIN_RASTER))
print(“n— vector info —“)
vinfo = geoai.get_vector_info(TRAIN_VECTOR)
for k, v in vinfo.items():
print(f” {k:<16}: {v}”)
gdf = gpd.read_file(TRAIN_VECTOR)
print(f”n {len(gdf)} training buildings | CRS {gdf.crs}”)
print(gdf.head(3))
geoai.view_vector(
gdf,
raster_path=TRAIN_RASTER,
outline_only=True,
edge_color=”yellow”,
outline_linewidth=0.8,
figsize=(11, 11),
title=”NAIP training scene + building footprints”,
)
try:
display(geoai.view_vector_interactive(TRAIN_VECTOR, layer_name=”Buildings”))
except Exception as e:
print(f” (interactive map unavailable here: {e})”)
return gdf
LABELS_GDF = timed(step2, “STEP 2 — Inspecting raster + vector data”)
TILES_DIR = os.path.join(WORK, “tiles”)
def step3():
stats = geoai.export_geotiff_tiles(
in_raster=TRAIN_RASTER,
out_folder=TILES_DIR,
in_class_data=TRAIN_VECTOR,
tile_size=CFG[“tile_size”],
stride=CFG[“stride”],
buffer_radius=CFG[“buffer_radius”],
all_touched=True,
skip_empty_tiles=False,
quiet=False,
)
n_img = len(os.listdir(f”{TILES_DIR}/images”))
n_lbl = len(os.listdir(f”{TILES_DIR}/labels”))
print(f”n chips: {n_img} images / {n_lbl} masks”)
if isinstance(stats, dict):
tot = max(stats.get(“total_tiles”, n_img), 1)
print(f” tiles containing buildings: {stats.get(’tiles_with_features’)} ”
f”({100 * stats.get(’tiles_with_features’, 0) / tot:.1f}%)”)
print(f” foreground pixels: {stats.get(‘feature_pixels’):,}”)
geoai.display_training_tiles(TILES_DIR, num_tiles=6, figsize=(18, 6))
return stats
TILE_STATS = timed(step3, “STEP 3 — Exporting image chips and label masks”)
We inspect the raster and vector datasets to understand their coordinate systems, dimensions, statistics, and feature structures. We visualize the building labels over the aerial imagery and generate an interactive map for spatial exploration. We then divide the source imagery into overlapping georeferenced chips and create matching raster masks for model training.
Copy CodeCopiedUse a different BrowserMODEL_DIR = os.path.join(WORK, “models_unet”)
BEST_MODEL = os.path.join(MODEL_DIR, “best_model.pth”)
def step4():
geoai.train_segmentation_model(
images_dir=f”{TILES_DIR}/images”,
labels_dir=f”{TILES_DIR}/labels”,
output_dir=MODEL_DIR,
architecture=CFG[“architecture”],
encoder_name=CFG[“encoder”],
encoder_weights=CFG[“encoder_weights”],
num_channels=CFG[“num_channels”],
num_classes=CFG[“num_classes”],
batch_size=CFG[“batch_size”],
num_epochs=CFG[“num_epochs”],
learning_rate=CFG[“learning_rate”],
val_split=CFG[“val_split”],
save_best_only=True,
early_stopping_patience=5,
verbose=True,
)
print(f”n best checkpoint: {BEST_MODEL}”)
print(f” size: {os.path.getsize(BEST_MODEL) / 1e6:.1f} MB”)
return BEST_MODEL
timed(step4, f”STEP 4 — Training {CFG[‘architecture’]}/{CFG[‘encoder’]} ”
f”for {CFG[‘num_epochs’]} epochs”)
def step5():
hist_path = os.path.join(MODEL_DIR, “training_history.pth”)
geoai.plot_performance_metrics(
history_path=hist_path,
figsize=(15, 5),
verbose=True,
save_path=os.path.join(WORK, “training_curves.png”),
)
h = torch.load(hist_path, weights_only=False)
best_ep = int(np.argmax(h[“val_iou”])) + 1
print(f”n best val IoU {max(h[‘val_iou’]):.4f} at epoch {best_ep}”)
print(” Reading the curves: val loss rising while train loss falls => overfitting;”)
print(” both flat and high => underfitting (more epochs, bigger encoder, or more chips).”)
timed(step5, “STEP 5 — Training diagnostics”)
We train a U-Net semantic segmentation model with a ResNet-34 encoder using the prepared image and mask tiles. We configure the training process with validation splitting, early stopping, checkpoint saving, and performance monitoring. We then load the training history, plot the learning curves, and identify the epoch that produces the highest validation IoU.
Copy CodeCopiedUse a different BrowserPRED_MASK = os.path.join(WORK, “test_prediction.tif”)
PRED_PROB = os.path.join(WORK, “test_probability.tif”)
def step6():
geoai.semantic_segmentation(
input_path=TEST_RASTER,
output_path=PRED_MASK,
model_path=BEST_MODEL,
architecture=CFG[“architecture”],
encoder_name=CFG[“encoder”],
num_channels=CFG[“num_channels”],
num_classes=CFG[“num_classes”],
window_size=CFG[“window_size”],
overlap=CFG[“overlap”],
batch_size=4,
probability_path=PRED_PROB,
)
geoai.print_raster_info(PRED_MASK, show_preview=False)
geoai.plot_prediction_comparison(
original_image=TEST_RASTER,
prediction_image=PRED_MASK,
titles=[“NAIP test scene”, “Predicted building mask”],
figsize=(16, 8),
prediction_colormap=”viridis”,
save_path=os.path.join(WORK, “prediction_comparison.png”),
)
with rasterio.open(PRED_MASK) as src:
m = src.read(1)
px = float(abs(src.transform.a) * abs(src.transform.e))
print(f” predicted building pixels: {int((m > 0).sum()):,} ”
f”({100 * (m > 0).mean():.2f}% of scene, ~{(m > 0).sum() * px:,.0f} m2)”)
timed(step6, “STEP 6 — Sliding-window inference on the test scene”)
VEC_RAW = os.path.join(WORK, “buildings_raw.geojson”)
VEC_ORTHO = os.path.join(WORK, “buildings_orthogonal.geojson”)
VEC_FINAL = os.path.join(WORK, “buildings_final.geojson”)
def step7():
grouped = geoai.region_groups(
PRED_MASK,
connectivity=2,
min_size=50,
out_image=os.path.join(WORK, “test_prediction_cleaned.tif”),
)
clean_mask = os.path.join(WORK, “test_prediction_cleaned.tif”)
raw = geoai.raster_to_vector(
clean_mask,
output_path=VEC_RAW,
threshold=0,
min_area=15,
simplify_tolerance=0.5,
)
print(f” raw polygons : {len(raw)}”)
ortho = geoai.orthogonalize(
input_path=clean_mask,
output_path=VEC_ORTHO,
epsilon=1.5,
min_area=15,
)
print(f” orthogonalized : {len(ortho)}”)
final = geoai.regularization(ortho, angle_tolerance=12, simplify_tolerance=0.4)
final = geoai.add_geometric_properties(
final, properties=[“area”, “perimeter”, “solidity”, “elongation”, “orientation”]
)
final.to_file(VEC_FINAL, driver=”GeoJSON”)
print(f” final footprints : {len(final)}”)
print(final.head())
if “area” in final.columns:
print(“n footprint area stats (m2):”)
print(final[“area”].describe().round(1).to_string())
fig, axes = plt.subplots(1, 2, figsize=(16, 8))
with rasterio.open(TEST_RASTER) as src:
rgb = src.read([1, 2, 3]).transpose(1, 2, 0)
rgb = np.clip(rgb / np.percentile(rgb, 99), 0, 1)
ext = plotting_extent(src)
for ax, g, t in zip(axes, [raw, final], [“Raw polygonization”, “Orthogonalized + regularized”]):
ax.imshow(rgb, extent=ext)
g.plot(ax=ax, facecolor=”none”, edgecolor=”red”, linewidth=1.1)
ax.set_title(t)
ax.set_axis_off()
plt.tight_layout()
plt.show()
try:
display(geoai.view_vector_interactive(VEC_FINAL, layer_name=”Predicted buildings”))
except Exception:
pass
return final
FINAL_GDF = timed(step7, “STEP 7 — Vectorizing and regularizing the predicted footprints”)
We apply sliding-window inference to an unseen NAIP scene and generate both prediction and probability rasters. We remove small noisy regions, convert the predicted mask into vector polygons, and regularize the footprint geometries to produce cleaner building boundaries. We also calculate geometric properties and compare the raw polygonized results with the orthogonalized and regularized outputs.
Copy CodeCopiedUse a different Browserdef step8():
gt_raster = os.path.join(WORK, “train_gt_mask.tif”)
geoai.vector_to_raster(
vector_path=TRAIN_VECTOR,
output_path=gt_raster,
reference_raster=TRAIN_RASTER,
fill_value=0,
all_touched=True,
dtype=np.uint8,
)
train_pred = os.path.join(WORK, “train_prediction.tif”)
geoai.semantic_segmentation(
input_path=TRAIN_RASTER,
output_path=train_pred,
model_path=BEST_MODEL,
architecture=CFG[“architecture”],
encoder_name=CFG[“encoder”],
num_channels=CFG[“num_channels”],
num_classes=CFG[“num_classes”],
window_size=CFG[“window_size”],
overlap=CFG[“overlap”],
quiet=True,
)
metrics = geoai.calc_segmentation_metrics(
ground_truth=gt_raster,
prediction=train_pred,
num_classes=2,
metrics=[“iou”, “f1”],
)
print(“n — pixel-wise metrics (class 0 = background, class 1 = building) —“)
for k, v in metrics.items():
print(f” {k:<10}: {np.round(v, 4)}”)
print(“n Building-class IoU is the number that matters; background IoU is inflated”)
print(” by the huge negative class and always looks great.”)
geoai.plot_prediction_comparison(
original_image=TRAIN_RASTER,
prediction_image=train_pred,
ground_truth_image=gt_raster,
titles=[“Imagery”, “Prediction”, “Ground truth”],
figsize=(18, 6),
save_path=os.path.join(WORK, “accuracy_comparison.png”),
)
return metrics
METRICS = timed(step8, “STEP 8 — Quantitative accuracy assessment”)
def step9():
if not CFG[“run_zero_shot”]:
print(” disabled in CFG”); return
chip = os.path.join(WORK, “test_chip.tif”)
with rasterio.open(TEST_RASTER) as src:
b = src.bounds
cx, cy = (b.left + b.right) / 2, (b.bottom + b.top) / 2
half = min((b.right – b.left), (b.top – b.bottom)) / 6
bbox = [cx – half, cy – half, cx + half, cy + half]
geoai.clip_raster_by_bbox(TEST_RASTER, chip, bbox=bbox, bbox_type=”geo”)
print(f” clipped chip: {chip}”)
sam = geoai.GroundedSAM(
detector_id=”IDEA-Research/grounding-dino-tiny”,
segmenter_id=”facebook/sam-vit-base”,
tile_size=1024,
overlap=128,
threshold=0.3,
)
out_mask = os.path.join(WORK, “zeroshot_mask.tif”)
gdf = sam.segment_image(
input_path=chip,
output_path=out_mask,
text_prompts=[“building”, “house”, “rooftop”],
polygon_refinement=True,
export_polygons=True,
min_polygon_area=30,
simplify_tolerance=1.5,
)
geoai.plot_prediction_comparison(
original_image=chip,
prediction_image=out_mask,
titles=[“Chip”, “Zero-shot: ‘building / house / rooftop'”],
figsize=(14, 7),
prediction_colormap=”viridis”,
)
print(f” zero-shot objects found: {len(gdf) if gdf is not None else 0}”)
geoai.empty_cache()
timed(step9, “STEP 9 — Zero-shot text-prompted segmentation (Grounding DINO + SAM)”)
We evaluate the segmentation model by comparing its predictions with rasterized ground-truth building labels. We calculate pixel-level IoU and F1 metrics and visualize the imagery, predictions, and reference masks together. We then apply Grounding DINO and SAM to perform zero-shot building segmentation using text prompts without additional model training.
Copy CodeCopiedUse a different Browserdef step10():
if not CFG[“run_pretrained”]:
print(” disabled in CFG”); return
extractor = geoai.BuildingFootprintExtractor(model_path=”building_footprints_usa.pth”)
gdf = extractor.process_raster(
TEST_RASTER,
output_path=os.path.join(WORK, “buildings_maskrcnn.geojson”),
batch_size=4,
confidence_threshold=0.5,
overlap=0.25,
mask_threshold=0.5,
min_object_area=100,
filter_edges=True,
)
if gdf is None or len(gdf) == 0:
print(” no instances returned”); return
print(f” building instances: {len(gdf)}”)
reg = extractor.regularize_buildings(gdf, min_area=20, angle_threshold=15)
reg.to_file(os.path.join(WORK, “buildings_maskrcnn_regularized.geojson”), driver=”GeoJSON”)
extractor.visualize_results(TEST_RASTER, gdf=reg, figsize=(12, 12))
if FINAL_GDF is not None:
print(f”n your U-Net : {len(FINAL_GDF)} polygons”)
print(f” pretrained R-CNN: {len(reg)} polygons”)
print(” Different counts are expected: U-Net merges adjacent roofs, Mask R-CNN splits”)
print(” them into instances. Pick the paradigm that matches your downstream question.”)
geoai.empty_cache()
timed(step10, “STEP 10 — Pretrained Mask R-CNN instance segmentation”)
def step11():
if not CFG[“run_real_aoi”]:
print(” disabled (set CFG[‘run_real_aoi’] = True to run; needs open internet)”)
return
bbox = (-83.9400, 35.9500, -83.9250, 35.9600)
items = geoai.pc_stac_search(
collection=”naip”,
bbox=list(bbox),
time_range=”2021-01-01/2023-12-31″,
max_items=3,
)
print(f” STAC items found: {len(items)}”)
aoi_dir = os.path.join(WORK, “aoi”)
tif = geoai.download_naip(bbox=bbox, output_dir=aoi_dir, max_items=1, preview=False)
print(f” NAIP: {tif}”)
ovt = os.path.join(aoi_dir, “overture_buildings.geojson”)
geoai.download_overture_buildings(bbox=bbox, output=ovt, overture_type=”building”)
print(f” Overture buildings: {ovt}”)
print(geoai.extract_building_stats(ovt))
raster = tif[0] if isinstance(tif, (list, tuple)) else tif
geoai.export_geotiff_tiles(
in_raster=raster,
out_folder=os.path.join(aoi_dir, “tiles”),
in_class_data=ovt,
tile_size=512,
stride=256,
)
print(” AOI dataset ready — feed it to train_segmentation_model() exactly as in STEP 4.”)
timed(step11, “STEP 11 — (optional) Real AOI: Planetary Computer NAIP + Overture Maps labels”)
def step12():
outputs = [f for f in sorted(os.listdir(WORK))
if f.endswith((“.tif”, “.geojson”, “.png”, “.pth”))]
print(” artifacts produced:”)
for f in outputs:
print(f” {os.path.getsize(os.path.join(WORK, f)) / 1e6:8.2f} MB {f}”)
zip_path = os.path.join(WORK, “geoai_results.zip”)
subprocess.run(
[“zip”, “-qr”, zip_path, “.”, “-i”, “*.geojson”, “*.png”, “*.tif”, “-x”, “*tiles*”],
cwd=WORK, check=False,
)
print(f”n bundle: {zip_path}”)
if IN_COLAB:
print(” Download it with: from google.colab import files; ”
“files.download(‘%s’)” % zip_path)
timed(step12, “STEP 12 — Results summary”)
banner(“DONE”)
print(“””
Where to go next
—————-
* Swap the head, keep the code: architecture=”deeplabv3plus”, encoder_name=”efficientnet-b3″
(or any timm encoder) in train_segmentation_model().
* 4-band NAIP (RGB+NIR): num_channels=4 everywhere; the first conv is auto-adapted.
* Multi-class land cover: geoai.train_segmentation_landcover() + geoai.export_landcover_tiles(),
with DiceLoss / FocalLoss / TverskyLoss from geoai.landcover_train for class imbalance.
* Instance segmentation you train yourself: geoai.train_MaskRCNN_model() then
geoai.instance_segmentation(…, vectorize=True).
* Object detection with georeferenced boxes: geoai.train.object_detection() /
geoai.object_detection(text=”cars”) for open-vocabulary Grounding DINO.
* Foundation models: geoai.prithvi_inference (NASA/IBM Prithvi), geoai.universat_inference,
geoai.DINOv3GeoProcessor for embeddings and similarity maps.
* Change detection: geoai.change_detection (torchange backends).
* Deploy: geoai.export_to_onnx() + geoai.onnx_semantic_segmentation(), or the QGIS plugin.
Docs and notebooks: https://opengeoai.org | Book: https://book.opengeoai.org
“””)
We use a pretrained Mask R-CNN model to extract individual building instances and compare them with the U-Net results. We optionally create a real-world dataset by downloading NAIP imagery from Microsoft Planetary Computer and matching building labels from Overture Maps. We finally collect the generated rasters, vectors, plots, and model outputs into a compressed results package for further analysis or download.
In conclusion, we completed an end-to-end geospatial deep learning pipeline that transforms raw aerial imagery into structured and analysis-ready building footprint data. We prepared training samples, trained and evaluated a semantic segmentation model, generated seamless predictions, and refined raster outputs into orthogonalized vector geometries with useful spatial attributes. We also examined alternative extraction approaches through zero-shot foundation models and pretrained instance segmentation, which helps us understand the trade-offs between custom training, prompt-based detection, and ready-to-use models. By packaging the generated masks, probability rasters, evaluation plots, trained weights, and GeoJSON outputs, we created a reusable foundation that we can adapt for land-cover mapping, infrastructure detection, change analysis, and large-scale GeoAI applications.
Check out the Full Codes here. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
The post A Tutorial on GeoAI: Designing Footprint Extraction from NAIP Imagery Using U-Net, Grounding DINO, SAM, and Mask R-CNN appeared first on MarkTechPost.