Drone survey and inspection software for India: open source

Read this first

Limits that apply to every pack

No model has been measured on Indian sites. README.md:116 The shared semantic head was scored on 49 Gorakhpur tiles (satellite/aircraft imagery at 0.448 m/px, one city), which the README says means 'the engine works on Gorakhpur', not drone imagery of Indian sites.
The India packs are geospatial engines. Most of them consume a georeferenced semantic class raster or detections from a validated model, and refuse when the evidence is missing. The only shipped semantic head (shared_semantic_landcover) has two usable classes, vegetation and building; road is weak (IoU 0.366 on the Gorakhpur holdout), water (0.003) and bare_land (0.000) do not work. A pack whose input class is not supplied by a validated model refuses rather than guessing.
Status is the computed status in docs/FEATURES.md on main (generated by tools/feature_status.py with 10 trained models installed locally; CI without weights reaches a lower count). 'verified' means the named tests pass; several India tests use constructed fixtures, noted per row where the registry says so.

Agriculture

Capability	What it measures	What it refuses	Status
Crop canopy segmentation	Crop, soil, unwanted vegetation and water masks with canopy cover and bare/missing regions, from a real semantic raster.	Refuses when the semantic schema has no crop/canopy/vegetation class, or no crop/canopy or soil pixels are present.	verified FEATURES.md:129
Multispectral crop indices	NDVI, NDRE and GNDVI from calibrated bands, reported as spectral indices rather than AI.	A missing band is marked unavailable rather than substituted.	verified FEATURES.md:130
Crop stress and anomaly map	Georeferenced stress/anomaly regions with the sensor, crop and validation scope stated.	Refuses when the source does not identify NDVI, NDRE or GNDVI, or when no valid canopy pixels remain.	verified FEATURES.md:131
Plant and tree counting	Geolocated plant/tree counts from connected instances; missing and health categories only where validated.	Refuses without a stated validation scope; does not invent missing or health categories.	verified FEATURES.md:132

Construction

Capability	What it measures	What it refuses	Status
Construction survey comparison	Two survey DSMs produce mapped rise/fall regions, changed area, added/removed volume and a client-readable report.	Does not invent a semantic cause for a change.	verified FEATURES.md:138
Progress against approved design	Observed coverage inside each element of an explicit approved design model.	No contractual percentage complete: schedule, quantities, hidden work and sign-off are not visible from the air. Refuses a design with no polygon elements or one that does not overlap the survey grid.	verified FEATURES.md:139
Construction site segmentation	Versioned schema and task-trained runtime gate exist.	No production construction head is trained.	in progress FEATURES.md:140

Land and property

Capability	What it measures	What it refuses	Status
Land-survey GIS extraction	Georeferenced polygons of semantic classes extracted from an orthomosaic's semantic raster.	Metric land analysis refuses an unreferenced raster; refuses a class raster with no valid survey pixels. Building and vegetation are the only classes the shipped semantic head supports usably.	verified FEATURES.md:146
Property and encroachment change	Change between surveys against an imported cadastral boundary and an aligned previous survey.	Refuses a cadastral boundary with no polygon features; rasters must share CRS, dimensions and transform, otherwise explicit registration is required.	verified FEATURES.md:147

Mining and quarry

Capability	What it measures	What it refuses	Status
Stockpile selection and measurement	Volume of a pile the operator outlines, with the base-surface method stated, as a mapped client result.	Does not propose stockpile outlines automatically (see below).	verified FEATURES.md:153
Mine and stockpile change	Per-pile and per-pit area/volume change between dated surveys.	—	verified FEATURES.md:154
Automatic stockpile segmentation		Out of v1 scope by decision, 2026-08-17: no corpus exists at drone resolution.	not started FEATURES.md:155
Mine and quarry segmentation		Out of v1 scope by decision, 2026-08-17. MineNetCD change detection reached IoU 0.2955 on about 15 held-out mines (60 per cent of flags wrong, 47 per cent of real change missed) and was rejected.	not started FEATURES.md:156

Power, rail, roads and solar assets

Capability	What it measures	What it refuses	Status
Power-line asset inspection	Mapped real vectors with capture geometry enforced for close-range components.	Refuses without a detection validation scope or a corridor centreline; a validated-model finding needs a confidence in [0, 1].	verified FEATURES.md:162
Railway corridor inspection	Mapped corridor findings with capture geometry appropriate to the component.	Same refusals as power-line inspection. The rail obstacle detector misses about one obstacle in four, so an empty result does not mean the corridor is clear.	verified FEATURES.md:163
Mapped road condition	Road-condition package with metric distance along an explicit surveyed centreline.	Refuses when the semantic schema has no road-surface class, when no road pixels are present, or without a measured surveyed centreline. The shipped semantic head's road class is weak (IoU 0.366).	verified FEATURES.md:169
Solar array and module inventory	Module counts from an instance raster; only layout gaps are called missing.	Refuses without a validation scope or when the instance raster contains no module instances.	verified FEATURES.md:175
Solar RGB and thermal inspection	RGB/radiometric registration from at least six operator-supplied tie points, and module temperatures where registered RGB module polygons and projected inventory polygons agree, with a named severity convention.	Tests use constructed RGB/radiometric fixtures, not a field capture, so this is not field validation; all outputs remain review candidates. Checks tie-point self-consistency, not image feature matching.	verified FEATURES.md:176

Which coordinate system do outputs use in India?

OpenDroneKit picks the UTM zone from the mean position of the dataset's geotags: core/geo.py auto_utm_epsg computes zone = floor((lon + 180) / 6) mod 60 + 1 and returns EPSG 32600 + zone (north) or 32700 + zone (south); auto_utm_epsg_for_fixes averages the geotags first. For mainland India (about 68 E to 97.4 E) that yields EPSG 32642 to 32647 (UTM 42N to 47N); most of the country falls in 43N to 46N (72 E to 96 E). The operator can instead pass an EPSG code (--epsg). geo.py:317

Flying in India

OpenDroneKit has no integration with DGCA’s Digital Sky platform and no regulatory approval of any kind. A search of the public code for DGCA, Digital Sky and NPNT finds nothing. Permissions, registration and airspace checks are the operator’s responsibility, through the official channels:

DGCA Digital Sky (registration and the airspace map)
The Drone Rules, 2021 (Gazette of India, G.S.R. 589(E), 25 August 2021)
Directorate General of Civil Aviation

Related: working offline · measurements · model cards · corridor template for rail, road and power lines.