eCognition Alternatives: Object-Based Image Analysis Without the Licence

Trimble's own product page carries this line: "Trimble eCognition is no longer actively sold. This page remains active for existing customers to access software downloads." I read it on 23 August 2026. So the question has changed. It is no longer whether the licence is worth renewing, it is which free tool covers which part of the job, and one part has no free replacement.

I co-founded TerraLab, so AI Segmentation near the end is ours. Same columns, same one-line limit as everything else.

What eCognition actually gives you

eCognition is an object-based image analysis environment rather than a button that finds buildings, and you steer every stage of it by hand.

Segmentation comes first. Multiresolution segmentation starts from single pixels and merges neighbours while a homogeneity threshold holds. You set a scale parameter for object size, split homogeneity between colour and shape, then split shape between smoothness and compactness. Shape cannot exceed 0.9. Those four numbers are the craft, and you retune them per sensor and per landscape.

Then a class hierarchy you write yourself: classes, inherited features, membership functions over spectral means, texture, area, neighbours and relations to parent and child objects. Then you run it, look, change the numbers, run it again.

The output that matters is the rule set, saved as a .dcp file or packaged as a .dax solution, not the polygons. That file is the method. Hand it to a colleague, run it over a thousand tiles on eCognition Server, get the same classification every time, and show an auditor why a polygon was called a building.

Nothing free replaces that today. Every tool below gives you a step. None gives you a portable, versioned rule file a colleague can run unchanged and defend in a review. If your deliverable is the method rather than the map, keep a machine that still opens your .dcp files.

Trimble's rule set documentation and the multiresolution segmentation reference are both still online and free to read.

Route 1: you only want the segmentation step

These carve the image into objects. They do not classify them.

1. Orfeo ToolBox

The closest free thing to eCognition's segmentation stage, and the only one built for large scenes.

• Install OTB, then the OrfeoToolbox Provider plugin, which left QGIS core at 3.36. The CookBook has the wiring. The command line works too.
• Runs on your machine, CPU only, no account.
• In and out any raster. LargeScaleMeanShift chains smoothing, segmentation, small-region merging and vectorization, and writes a vector file whose polygons already carry the per-band mean and standard deviation. That table feeds the classifier next.
• Cost free, Apache 2.0.
• Fails on the defaults: spatial radius 5, range radius 15, minimum size 50 pixels, tiles of 500 by 500. Expect several runs before the object size is right.

orfeo-toolbox.org · LargeScaleMeanShift

2. GRASS i.segment

Already inside , because GRASS ships with it.

• Install nothing. The GRASS provider is in Processing out of the box.
• Runs on your machine, CPU only, no account.
• In and out an imagery group. Out comes a segment raster, which i.segment.stats turns into per-object statistics you can classify.
• Cost free, GPL.
• Fails on patience. Two methods exist, region growing and mean shift, and the region-growing threshold has no physical meaning, so you find it by bisection. The i.segment.uspo addon automates that search and takes hours.

i.segment manual · i.segment.stats · i.segment.uspo

3. SAGA

The quickest way to see a segmentation without reading a manual.

• Install the Processing Saga NextGen Provider plugin, plus SAGA itself. It left the QGIS core Processing set.
• Runs on your machine, CPU only, no account.
• In and out grids in, a segment grid and optional polygons out.
• Cost free. GPL, with the API under LGPL.
• Fails on anything that has to scale or repeat. Its Imagery, Segmentation library holds Seeded Region Growing, Superpixel Segmentation, Watershed Segmentation and three helpers, and it is fast to try. It carries no object-feature framework, so the statistics step is yours to build.

saga-gis.sourceforge.io

4. scikit-image and RSGISLib

For people who would rather write the segmentation than click it.

• Install a Python environment. scikit-image is a pip or conda install, RSGISLib is conda and heavier.
• Runs on your machine, CPU only, no account.
• In and out arrays in, label arrays out. Georeferencing is yours to handle, through rasterio or GDAL.
• Cost free. scikit-image is BSD-3-Clause, RSGISLib is GPL v3.
• Fails on the plumbing. Felzenszwalb, quickshift, SLIC and watershed are four lines each, and none knows what a CRS is. RSGISLib's Shepherd segmentation is the closest published match to eCognition, it installs from conda-forge only, and its site would not answer over https on 23 August 2026, so use the repository.

scikit-image segmentation examples · RSGISLib repository

Route 2: you want supervised classification

You draw training areas, the tool learns them, the whole raster gets labelled. This is the half of eCognition that free software genuinely covers.

5. Semi-Automatic Classification Plugin

The one most eCognition refugees land on, and rightly.

• Install the plugin, then let it fetch its dependencies. It sits on Remotior Sensus and uses scikit-learn and PyTorch.
• Runs on your machine, CPU is fine, no account. An optional tool downloads Landsat and Sentinel-2 scenes, the only time it talks to a server.
• In and out a multi-band raster plus training areas you draw. Out comes a classified raster, an accuracy report, and a vector layer if you ask. Algorithms: maximum likelihood, minimum distance, spectral angle mapping, random forest, support vector machine and a multi-layer perceptron.
• Cost free, GPL v3.
• Fails on objects. It classifies pixels, not segments, so a roof comes back speckled and you clean it up afterwards.

Manual · Plugin listing

6. dzetsaka

Smaller, faster to learn, and it will train a model in about a minute.

• Install the plugin. When you pick an algorithm whose library is missing, it offers to pip-install it and prints progress to the QGIS log, so you rarely open a terminal.
• Runs on your machine, CPU only, no account.
• In and out a raster plus a shapefile of regions of interest. A classified raster and a saved model file. Eleven algorithms, from a Gaussian mixture model that needs nothing beyond QGIS to XGBoost and CatBoost.
• Cost free, GPL v3.
• Fails on text labels. The ROI column must hold class numbers, not names.

Repository · Plugin listing

7. GRASS i.maxlik and r.learn.ml2

Already installed, and the pair that scripts best.

• Install nothing for i.maxlik. r.learn.ml2 is an addon and needs scikit-learn.
• Runs on your machine, CPU only, no account.
• In and out a signature file from i.gensig for the first, a raster stack plus labels for the second. Classified rasters out.
• Cost free, GPL.
• Fails on discoverability. The parameter names assume you already know the GRASS data model. Once you do, they drop straight into a bash or Python pipeline, the closest anything free gets to an eCognition Server batch.

i.maxlik · r.learn.ml2

8. Orfeo TrainImagesClassifier

Pair it with LargeScaleMeanShift and you have the real object-based chain: segment, compute per-object statistics, classify.

• Install OTB, as above.
• Runs on your machine, CPU only, no account.
• In and out image and training vector pairs in. A model file out, plus a confusion matrix if you ask. It writes no image and no vector, so ImageClassifier or VectorClassifier applies the model afterwards.
• Cost free, Apache 2.0.
• Fails on first contact. It is built on LibSVM, OpenCV machine learning and Shark ML, each with its own parameter block. Read the CookBook page before you run it, not after.

TrainImagesClassifier

Route 3: you want the object out, without writing a rule set

Model-based tools. You give up the tuning and the audit trail, and you get a result today.

9. Deepness

Runs any ONNX model you supply over a raster, in tiles.

• Install the plugin, then it pip-installs OpenCV and ONNX Runtime on first start.
• Runs on your machine. CPU works, a GPU needs your own CUDA setup. No account, nothing leaves.
• In and out any raster layer, including an XYZ or WMS basemap, plus your ONNX model. A vector or raster layer out.
• Cost free, Apache 2.0.
• Fails on imagery outside the model's training range. A model trained on 25 cm imagery will not read a 5 cm drone flight, and the plugin cannot warn you.

Repository

10. Geo-SAM

Segment Anything in a QGIS panel, one prompt at a time.

• Install the plugin, in the plugin manager since version 2.0. It downloads its own checkpoints from inside QGIS.
• Runs on your machine, fast enough on a laptop CPU. No account, nothing leaves.
• In and out a raster file or an XYZ, WMS or WMTS layer. One polygon per click or box, saved to a Shapefile.
• Cost free, MIT.
• Fails on area work. It returns one object at a time. Two modes exist since 2.0, and in the pre-encoded one you can only prompt inside the extent you encoded first.

Repository · Plugin listing

11. samgeo

The same models as a Python package, for whole scenes.

• Install a Python environment, then the package. Budget an afternoon the first time.
• Runs on your machine, GPU recommended. No account, but the weights are a real download.
• In and out a GeoTIFF, or basemap tiles it fetches. Mask rasters plus GeoJSON, Shapefile or GeoPackage.
• Cost free, MIT.
• Fails on office laptops, and on CRS mismatches, because it reprojects nothing. This is a library, so you write the code unless you add its companion plugin.

samgeo.gishub.org · Repository

12. GeoAI

A panel per task, plus a trainer for your own models.

• Install the plugin, plus a PyTorch stack underneath. The project says the CUDA dependencies make the install hard, and it is right.
• Runs on your machine, NVIDIA card in practice. No account, nothing leaves.
• In and out a raster layer. Polygons and masks, one panel per task.
• Cost free, MIT.
• Fails on the install, more often than on the imagery. Once it runs it is the widest toolbox here: Segment Anything, tree crowns, water masks, Mask R-CNN.

Repository · Plugin listing

13. Mapflow

Draw an area, pick a class, get vectors back from their servers.

• Install the plugin only.
• Runs on Geoalert's servers. No GPU, an account is required, and the imagery leaves your machine on every job.
• In and out an area you draw. Vector layers by class.
• Cost credits at $0.10 each. Buildings are 10 credits per km², a zoom 18 basemap adds 3, so about $1.30 per km². The free tier is 250 credits, capped at 25 km².
• Fails on dense fabric. Terraced houses merge into one polygon, visible on their own screenshots.

mapflow.ai

14. AI Segmentation, ours

Draw a zone, type the object in plain words, get every match as polygons.

• Install the plugin only. No weights, no CUDA, no Python environment.
• Runs on our servers for the two cloud modes, and on your machine for the small local model in Semi-Auto. No GPU either way. A free account is required.
• In and out orthophotos and basemaps. A styled GeoPackage out, one polygon per object, with a label, a class, a confidence score and an area.
• Cost the free plan covers 3 km² of Automatic detection and 50 Semi-Auto object detections a month, no card. Pro is €39 a month for 300 km² and 2,000 object detections. Semi-Auto on your own machine has no counter.
• Fails on rules that forbid any upload. The two cloud modes send imagery to a server, and only the small local model stays on the machine. We read imagery and nothing else, so a scanned cadastral sheet gives nothing.

Product page · Registry entry

Try AI Segmentation free in QGIS, no card needed

Side by side

How to choose

• Your deliverable is the method, not the map. Nothing here replaces a .dcp. Keep a machine that opens your rule sets, and rebuild the method as a documented Processing model or a Python script before that machine dies.
• You segmented, then classified per object. Orfeo, both halves: LargeScaleMeanShift for the objects, TrainImagesClassifier for the labels. The only free chain that carries object statistics end to end.
• You mostly drew training areas and pressed classify. SCP for an accuracy report, dzetsaka for a model in a minute.
• You want buildings out of imagery this afternoon. Deepness with a ready model, or Mapflow or ours if the imagery may leave the machine.
• The imagery may never leave the machine. That one answer rules out Mapflow and our two cloud modes, and leaves everything else here.

Start with i.segment on one tile tonight. It is already in your Processing toolbox, it costs nothing, and twenty minutes with the threshold will tell you how much of your eCognition work was the segmentation and how much was the rule set.

The eight AI plugins compared goes deeper on the model-based half. The QGIS AI hub lists the rest.