Industry applications

Workflows below are illustrative study designs; they do not establish operational hazard, traffic, agronomic or infrastructure decisions. Their geometry/counting recipes can be tested independently of the domain model.

Urban land use and population

Land cover and land use are different labels: optical spectra alone do not establish residential versus commercial use. Combine appropriately licensed imagery, building attributes and independent labels. Validate raster grids, named bands and QA before feature extraction. Use a spatial holdout and per-class confusion matrix as in machine-learning.md.

skimage.feature.graycomatrix summarizes the supplied quantized patch. Calling it once on the full image produces global scalars, not a per-pixel texture map. For local textures, state window size, quantization and boundary/nodata policy and check chunk halos. Do not apply remove_small_objects directly to a categorical raster expecting each disconnected class region to be handled independently.

Dasymetric redistribution must conserve each source-zone total. This tested function allocates a scalar count over nonnegative cell weights for one source zone:

import numpy as np

def redistribute_population(total, weights):
    weights = np.asarray(weights, dtype=float)
    if not np.isfinite(total) or total < 0 or not np.isfinite(weights).all() or (weights < 0).any():
        raise ValueError('Population and weights must be finite and nonnegative')
    if weights.sum() == 0:
        if total == 0:
            return np.zeros_like(weights)
        raise ValueError('Positive population has no eligible destination')
    return total * weights / weights.sum()

Apply independently within each census zone and assert sums after masking/cropping. Do not multiply this allocation by the original population raster again. Counts and densities require different resampling; record rounding and reconciliation policy.

Disaster management

Flood risk requires a calibrated discharge/scenario, hydraulic model, connected inundation, exposure and vulnerability. A DEM threshold using invented water levels for 10/100/500-year events is not a return-period model. Keep hazard probability, water depth, exposed people/assets and damage separate, with common units and time. Use hydrology for the preprocessing sequence and validate against observed events; clipping settlements yields exposure, not damage.

Wildfire assessment requires justified fuels, moisture, topography, weather and a calibrated spread/risk model. An arbitrary product of scaled wind/slope/vegetation is an illustrative score, not probability. Reclassify NumPy arrays using explicit lookup or boolean masks; arrays do not have map_classes. Keep wind speed/direction and fuel-model provenance, uncertainty and out-of-domain conditions with any output.

Utilities and infrastructure

For vegetation corridors, project linework to a locally appropriate metric CRS, buffer a justified corridor distance, reproject that geometry to the raster CRS, and pass the open Rasterio dataset to rasterio.mask.mask, not a NumPy array. The returned affine transform locates the cropped pixels. Masked height values and wire height/sag/clearance uncertainty need explicit treatment.

from rasterio.mask import mask
import rasterio

metric_lines = power_lines.to_crs(power_lines.estimate_utm_crs())
corridors = metric_lines.geometry.buffer(buffer_metres)
with rasterio.open('vegetation_height.tif') as src:
    if src.crs is None:
        raise ValueError('Raster CRS missing')
    heights, cropped_transform = mask(src, corridors.to_crs(src.crs), crop=True, filled=False)
# row,column component centres must be transformed through cropped_transform.

A chosen canopy threshold marks screening candidates, not regulatory clearance. Compute area from valid projected-cell support, not a hardcoded 1 m² per pixel. Export georeferenced findings for review; output generation is not authorization to create work orders.

For pipeline least-cost routing, encode truly prohibited cells as inaccessible edges, not merely high finite cost (which can still be selected). Raster costs must have units, resolution and directional assumptions. Account for diagonal step length and corner cutting, preserve barriers, check endpoint accessibility and handle no path. Verify the reconstructed route independently against constraints and cumulative cost.

Transportation and transit

OSMnx constructs topological street networks from OSM. graph_from_gdf(roads) does not exist: graph_from_gdfs(nodes, edges) requires correctly indexed node/edge tables and CRS, or download a bounded walk/drive graph. Lines that cross on a map may be bridges/tunnels and must not be automatically connected.

For a walk-time accessibility template on an existing graph:

import networkx as nx

def reachable_walk_nodes(graph, origin, minutes=15, speed_kph=5):
    if minutes < 0 or speed_kph <= 0:
        raise ValueError('Nonnegative time and positive walking speed required')
    graph = graph.copy()
    for _, _, data in graph.edges(data=True):
        length = float(data['length'])
        if not np.isfinite(length) or length < 0:
            raise ValueError('Edge length must be finite nonnegative metres')
        data['walk_seconds'] = length / (speed_kph / 3.6)
    return nx.single_source_dijkstra_path_length(
        graph, origin, cutoff=minutes * 60, weight='walk_seconds')

This is fixed-speed walking with the graph's direction/access restrictions. It excludes transit waiting, timetables, transfers, grade and partial terminal edges. A convex hull of reachable nodes can include inaccessible water/barriers; label it a visualization approximation, not a service-area boundary. Edge count is not a travel-time measure.

AADT is vehicles/day, not a peak-hour flow. Do not divide it by vehicles/hour capacity without a calibrated temporal conversion. Nearby sensor interpolation alone does not model connectivity, turns or demand; validate predictions on withheld roads/times.

Sources: Rasterio masking, OSMnx contracts, NetworkX Dijkstra, Scikit-image GLCM.