Lite Classify
Object-based land-cover mapping of drone orthomosaics, offline on one Windows workstation.
Lite Classify splits an RGB orthomosaic into objects, describes each one by its color, texture,
shape, height and appearance, helps you find the classes the scene actually contains, learns
those classes from examples you click, and checks the finished map against an independent,
design-based sample.
How it works
You work through four tabs, left to right:
- Load, View & Segment: load the orthomosaic and an optional DSM, color-correct, resample,
and segment into objects with SLIC (CPU or GPU, tiled for large scenes). Each object gets 24
features: color, vegetation indices, GLCM texture, shape, elevation, slope, height above
ground, and contrast with its neighbors.
- Discover Classes: each object's surroundings are described by a frozen DINOv2 ViT-S/14
(or, optionally, a DINOv3 ViT-S+/16 weights file you obtain from Meta) at 5 m and 10 m. The objects are over-clustered with k-means, and you review, merge, split
and name the groups to build the class list and the first training samples.
- Classification: label more samples by clicking, dragging or growing a selection, with
help from "Find Similar Segments" and "What to Label Next". Check class separability
(Jeffries–Matusita), train a Random Forest, retrain until the map stops improving, and export
a GeoTIFF with a QGIS style. An optional pixel-level greenness refinement recovers grass
strips narrower than the objects. A large image can be processed one extent at a time.
- Accuracy Assessment: draw a stratified or simple random sample, label the points by eye
or import field data, and get the error matrix plus area-weighted accuracy and area estimates
with confidence intervals (Olofsson et al., 2014; Stehman, 2014).
Running from source
Python 3.13 on Windows 11 is the tested setup.
python -m venv .venv
.venv\Scripts\pip install -r requirements.txt # PyTorch with CUDA 12.4
.venv\Scripts\python liteclassify.py
On a machine without an NVIDIA GPU, use requirements-cpu.txt instead. The DINOv2 weights
(about 85 MB) are downloaded once on first use, or read from models/ if you place them
there.
To build the executables and installers, see installer/README.md.
Repository layout
| Path | Contents |
|---|
liteclassify.py | Launcher: splash screen, then the main window; --selftest for headless checks |
obia_app.py | The Qt application: the four tabs and their workflow |
obia_engine.py | Raster I/O, segmentation, features, classification support, accuracy assessment (no Qt) |
obia_discover.py | DINOv2 context embeddings and the Discover grouping |
obia_ui.py | Reusable widgets: map viewer, toggle switch, collapsible sections, separability figure |
app_info.py | Name, version and branding |
build_exe.py, installer/ | PyInstaller build and the Inno Setup installers |
docs/manual/ | The user manual (HTML) and the scripts that generate its figures |
experiments/cnn/ | The offline experiments behind the DINOv2 appearance features |
third_party/dinov2/ | Vendored DINOv2 backbone code (Apache 2.0) |
License
Lite Classify is free software under the GNU General Public License v3, with the
additional attribution terms in NOTICE.md. Modified versions must keep the
credit line below, be marked as modified, and use a different name:
Based on Lite Classify by L. D. Barua (LeoBar) — https://github.com/leobarua/liteclassify
For other licensing terms, see COMMERCIAL-LICENSING.md.
To contribute, see CONTRIBUTING.md; contributions require agreeing to the CLA.
Citing
If you publish work made with Lite Classify, please cite the version you used:
Barua, L. D. (2026). Lite Classify (Version 1.2) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.23877519
To cite all versions, use the concept DOI
10.5281/zenodo.23462352. Other formats (BibTeX, APA)
are under "Cite this repository" on GitHub, generated from CITATION.cff.
