geomaster – skill Dữ liệu & Khoa học cho agent AI

GeoMaster

Geospatial analysis across vector/raster GIS, remote sensing, spatial ML, terrain,

networks, and scientific applications. Start with the relevant workflow, inspect

input provenance, and load only the reference needed for the task.

Tested scope and installation

The local recipe suite targets GeoPandas 1.2.0, Rasterio 1.5.2, Shapely 2.1.2,

PyProj 3.8.0, Rioxarray 0.23.0, Xarray 2026.9.0, OSMnx 2.1.1 and

PySTAC Client 0.9.0. See [review and source ledger](references/review.md) for

execution limits. Examples requiring actual input files are templates; authenticated

Earth Engine/CDS/commercial services, desktop GIS and GPU training remain illustrative.

# In a dedicated environment; install only the workflow's optional dependencies.
uv venv --python 3.13
uv pip install geopandas==1.2.0 rasterio==1.5.2 shapely==2.1.2 pyproj==3.8.0
uv pip install rioxarray==0.23.0 xarray==2026.9.0 'dask[array]' scikit-learn==1.9.1
uv pip install pystac-client==0.9.0 planetary-computer==1.0.0 odc-stac
# For osgeo/native CLI or PDAL, use a separate conda-forge environment:
# conda create -n geo-native -c conda-forge python=3.13 gdal pdal python-pdal

Install TorchGeo/PyTorch, RSGISLib, Py6S/6S, ArcPy, QGIS or other specialist

runtimes separately when required. A Rasterio wheel includes its own GDAL library;

it does not install osgeo or the GDAL command-line programs.

Workflow

1. Record product/collection ID, acquisition time, processing baseline, license,

band names, scale/offset, QA meaning, horizontal/vertical CRS and units.

2. Verify CRS from authoritative metadata. set_crs labels coordinates;

to_crs transforms them. Never infer an unknown CRS from plausible bounds.

3. Align extent, affine transform, dimensions, pixel registration and resolution.

Reproject categorical masks with nearest-neighbor resampling; choose an appropriate

resampler for continuous data. Matching array shape alone does not prove alignment.

4. Preserve nodata, clouds, shadows and saturation masks. Convert unsigned integers

to floating point before differences; apply the provider's radiometric transform once.

5. Run the analysis at a defensible support/resolution. Keep training labels and

validation blocks separate. Terrain elevation and horizontal units must agree.

6. Export CRS, transform, valid-data mask, units, model settings and provenance.

Check numeric expectations on a small known fixture before scaling up.

Local raster recipes

Import the bundled [raster helper](scripts/raster_workflows.py) from its directory

(add that directory to PYTHONPATH or run from it). It implements small in-memory

recipes; use windows/Dask for larger data. Writers require a new output path. It does not infer band identities or masks.

NDVI

from raster_workflows import write_ndvi

# This example assumes a VERIFIED four-band B02/B03/B04/B08 stack whose mask
# already excludes clouds/shadows, and harmonized DN reflectance = DN * 0.0001.
write_ndvi('s2_masked_stack.tif', 'ndvi.tif', red_band=3, nir_band=4,
           scale=0.0001, offset=0.0)

Do not use those indices or calibration for an arbitrary sentinel2.tif.

SAFE products and STAC assets are often separate single-band rasters. Additive

radiometric offsets affect even NDVI. EVI/SAVI require physical reflectance.

normalized_difference keeps undefined ratios and invalid cells as NaN, not zero.

Terrain

import rasterio
from raster_workflows import terrain_metrics

with rasterio.open('dem_metres.tif') as src:
    slope_deg, aspect_deg, shade = terrain_metrics(
        src.read(1, masked=True), src.transform, src.crs)

The helper requires a north-up projected metric grid and elevations in metres.

Aspect points downslope clockwise from north; it is undefined for flat cells.

Slope is resolution-aware; nodata contaminating a derivative stencil stays invalid.

Hillshade is an illumination visualization, not hydrological flow or exposure risk.

Classification

import geopandas as gpd
from raster_workflows import classify_imagery

training = gpd.read_file('training.gpkg')  # polygons with class_id in 1..65534
model = classify_imagery('masked_features.tif', training, 'classified.tif')

The helper checks CRS, geometry, labels, valid training pixels and overlapping

labels, then preserves nodata in a uint16 output. It fits a small demonstration

model; it does not measure accuracy. Use spatial/temporal holdouts at the field,

scene or regional level before reporting predictive performance. See

[machine learning](references/machine-learning.md).

Vector analysis

import geopandas as gpd

zones = gpd.read_file('zones.geojson')
points = gpd.read_file('points.geojson')
if zones.crs is None or points.crs is None:
    raise ValueError('Resolve missing CRS before analysis')
points = points.to_crs(zones.crs)
joined = gpd.sjoin(points, zones, how='inner', predicate='within')
stats = joined.groupby('zone_id')['value'].agg(['count', 'mean', 'std'])

# Local data only: verify the estimated CRS area of use before accepting it.
metric = zones.to_crs(zones.estimate_utm_crs())
metric['area_m2'] = metric.area
buffers = metric.geometry.buffer(1000).to_crs(zones.crs)

within excludes boundary points; overlapping zones can duplicate observations.

Choose and document a boundary/overlap policy. A projected CRS need not use metres

or preserve area. UTM suits local regions, not every national, polar or global task.

Cloud catalogs and Earth Engine

[Data sources](references/data-sources.md) documents current STAC, CDSE, CDS,

Overpass and geocoding contracts. Discovery is separate from downloading pixels.

from pystac_client import Client
import planetary_computer

catalog = Client.open('https://planetarycomputer.microsoft.com/api/stac/v1',
                      modifier=planetary_computer.sign_inplace)
search = catalog.search(collections=['sentinel-2-l2a'],
                        bbox=[-122.5, 37.7, -122.3, 37.9],
                        datetime='2023-06-01/2023-06-30',
                        query={'eo:cloud_cover': {'lt': 20}}, max_items=5)
items = list(search.items())
if not items:
    raise ValueError('No matching scenes')
# Save item IDs/properties; inspect asset keys, scale/offset and QA before load.

limit controls page size; max_items bounds total traversal. Planetary Computer

signing returns expiring SAS asset URLs; use unsigned IDs/metadata for durable

provenance and sign near the time of access. Earth Engine requires prior

ee.Authenticate() and ee.Initialize(project='your-registered-project').

Cloud filters at scene level do not replace per-pixel masks. The

[remote-sensing reference](references/remote-sensing.md) includes SCL masking and

region summaries with an explicit exclusive date end.

Networks

import osmnx as ox
import networkx as nx
G = ox.graph_from_place('Portland, Maine, USA', network_type='drive')
G = ox.routing.add_edge_speeds(G)
G = ox.routing.add_edge_travel_times(G)
origin = ox.distance.nearest_nodes(G, -70.26, 43.66)
destination = ox.distance.nearest_nodes(G, -70.27, 43.67)
route = nx.shortest_path(G, origin, destination, weight='travel_time')

This makes public Nominatim/Overpass requests. Handle no-route results and inspect

imputed speeds; these estimate free-flow time, not observed traffic. Use graph CRS

for nearest-node coordinates and distinguish travel-time seconds from metres.

Efficient storage

import rioxarray
from rasterio.shutil import copy as rio_copy
from rio_cogeo.cogeo import cog_validate

cube = rioxarray.open_rasterio('large.tif', masked=True,
                             chunks={'band': 1, 'x': 1024, 'y': 1024})
# Compute bounded windows or reductions; do not force the entire cube into memory.
rio_copy('input.tif', 'output_cog.tif', driver='COG', compress='DEFLATE')
valid, errors, warnings = cog_validate('output_cog.tif')
if not valid:
    raise ValueError(errors)

Tiling/compression alone does not establish COG layout. Preserve or regenerate

appropriate overviews at COG creation; do not mutate the result in place.

References

• [Coordinate systems](references/coordinate-systems.md): authority, axis order, projections
• [Core libraries](references/core-libraries.md): GDAL, Rasterio, Fiona, Shapely, PyProj, GeoPandas
• [Remote sensing](references/remote-sensing.md): optical, SAR, hyperspectral, QA and calibration
• [Machine learning](references/machine-learning.md): spatial holdouts, RF/SVM, neural models, XAI
• [GIS software](references/gis-software.md): QGIS, ArcGIS, GRASS, SAGA
• [Scientific domains](references/scientific-domains.md): marine, atmosphere, hydrology, agriculture, forestry
• [Advanced GIS](references/advanced-gis.md): 3D, trajectories, topology, networks
• [Big data](references/big-data.md): Dask, cloud access, GPU, COG, GeoParquet, Zarr
• [Industry applications](references/industry-applications.md): urban, hazards,

… (nội dung đầy đủ có trong gói tải về)