Lite Classify

by LeoBar · a lightweight desktop app for object-based land-cover mapping of drone orthomosaics

User manual

Turn a drone orthomosaic into a checked land-cover map

Lite Classify splits an aerial image into objects, describes each one by its colour, texture, shape, height and appearance, helps you find the classes the scene really contains, learns those classes from examples you click, and checks the finished map against an independent sample. This manual walks one real -hectare survey through every step: objects, nine classes, training objects and independently labelled check points.

• Bare
• Crop 1
• Crop 2
• Grass
• Water
• Tree
• Road
• Building
• Mud

Part A

Getting started

Install, meet the worked example, learn the window.

Part B

How it works

Segments, the 24 measurements, DINOv2 appearance, class discovery, Random Forest, accuracy.

Part C

Workflow

Every control in the four tabs, step by step.

Part D

Reference

Files, parameters, troubleshooting, glossary, licence and citation.

1Welcome

Lite Classify is a desktop tool for geographic object-based image analysis (GEOBIA) of very-high-resolution drone imagery. It was built to map land cover in tropical landscapes, from canopy gaps and restoration patches to farms and campuses, and it works for any cover types you can recognise by eye in an RGB orthomosaic: trees, crops, grass, bare ground, roads, buildings, water.

Lite Classify picks up where Lite Flight leaves off. Plan and fly the survey with Lite Flight, process the photos into an orthomosaic and surface model in your photogrammetry software, then bring both into Lite Classify to map them. It is one self-contained program that runs offline on a single workstation, with no Python, GIS software, account or internet connection needed.

You work through four tabs, left to right. Each tab hands its result to the next.

Computing is the quick part. On the worked example, segmentation and feature measurement for objects ran in minutes on a desktop GPU. Most of your time goes into reviewing groups, labelling examples and checking the result.

2Installing and launching

Lite Classify ships as a self-contained Windows program; you do not need Python or any GIS software installed.

1. Download an installer from the release page, or from the Google Drive mirror. Setup CPU (about 320 MB) runs on any Windows 10/11 PC. Setup GPU (about 1.8 GB) adds NVIDIA CUDA support: much faster segmentation and appearance features on an NVIDIA graphics card. Both work offline once installed.
2. Run the installer. It installs for your user account, with no administrator rights needed, into %LOCALAPPDATA%\Programs\Lite Classify, and adds a Start-menu shortcut (a desktop shortcut is optional). Installing one edition over the other replaces it. Uninstall from Windows Settings › Apps.
3. Start Lite Classify from the Start menu. A splash screen shows while the libraries load; the first launch can be slow, later launches are quicker.
4. Check the console at the bottom of the window. It should say Load an image to begin. Every action the app takes is logged there, including errors.
5. Check the Segmentation Engine box on Tab 1. If GPU is selectable, an NVIDIA CUDA graphics card was found; segmentation and the DINOv2 appearance features will run on it. If it is greyed out, everything runs on the CPU, just slower.

3The worked example

Every figure and number in this manual comes from one drone survey: the ha CAFS experimental agricultural site of the University of the Philippines Los Baños, used for field experiments in agriculture-related courses. It lies along UPLB Pili Drive in Los Baños, Laguna, northeast of the Mount Makiling Forest Reserve and outside its boundaries. In one flight it holds what an agricultural map has to separate: annual crops in rows and blocks, thin grass strips between bare beds, dying grass that is nearly the colour of soil, mud, a pond and ditches, a road lined with trees, and buildings.

Orthomosaic

RGB, 8-bit, cm/px

Native size

× px

Surface model

DSM, cm/px

Working grid

5 cm · × px

Mapped area

ha

Coordinates

WGS 84 / UTM 51N

DSM heights

– m

Objects

The orthomosaic was captured at about 2.2 cm per pixel. The worked example segments it on a 5 cm working grid into small objects, about 2 m² each, so that crop rows and beds stay separate, and adds the DINOv3 appearance features (Part B, 5.4). Grass strips narrower than an object are recovered after classification by the greenness refinement (Part B, 5.9). The DSM comes from the same photogrammetry run on a coarser grid; Lite Classify places it by its coordinates, so the two need not share a pixel size.

The sample data (orthomosaic, DSM and the saved workspace) can be downloaded from the sample-data folder on Google Drive, so you can follow every step on the same data.

4A tour of the interface

The window has a menu bar, four tabs, a control column on the left of each tab, a large map viewer, and a console strip along the bottom that logs every step. All four viewers share one copy of the image, so switching tabs is instant even on large mosaics, and they share the zoom and pan: switching tabs keeps the same view. The bar above each map shows the image's native and working resolution, size, coordinate system and bit depth; below that, the map coordinates under the cursor.

Menus

Closing the window, New and Load Workspace all ask Save / Discard / Cancel when there is unsaved work.

The four tabs

Each tab is described control by control in Part C.

Mouse and keys