dzetsaka 
You can download samples to test the plugin on your own.
As this tool was developped during my work in the Guiana Amazonian Park to classify different kind of vegetation, I gave an Teko name (a native-american language from a nation which lives in french Guiana) which represent the objects we use to see the world through, such as satellites, microscope, camera...
dzetsaka : Classification tool runs with scipy library. You can download package like Spider by Anaconda for a very easy setup.
Then, as this plugin is very simple, you will just need two things for making a good classification :
- A raster
- A shapefile which contains your ROI (Region Of Interest)
The shapefile must have a column which contains your classification numbers (1,3,4...). Otherwise if you use text or anything else it certainly won't work.
dzetsaka now supports 11 powerful machine learning algorithms:
- Gaussian Mixture Model (GMM) - Fast baseline classifier
- Random Forest (RF) - Robust ensemble method
- Support Vector Machine (SVM) - High-accuracy classifier
- K-Nearest Neighbors (KNN) - Simple distance-based classifier
- XGBoost (XGB) - State-of-the-art gradient boosting
- LightGBM (LGB) - Fast gradient boosting framework
- Extra Trees (ET) - Extremely randomized trees
- Gradient Boosting Classifier (GBC) - Scikit-learn gradient boosting
- Logistic Regression (LR) - Linear probabilistic classifier
- Naive Bayes (NB) - Fast probabilistic classifier
- Multi-layer Perceptron (MLP) - Neural network classifier
NEW FEATURE: dzetsaka can now automatically install missing dependencies!
When you select an algorithm that requires additional packages (XGBoost, LightGBM), dzetsaka will:
- Detect missing dependencies automatically
- Offer to install them with one click
- Handle the installation process in the background
- Provide real-time progress in the QGIS log
Supported auto-installation:
- ✅ scikit-learn (for RF, SVM, KNN, ET, GBC, LR, NB, MLP)
- ✅ XGBoost (for XGB classifier)
- ✅ LightGBM (for LGB classifier)
No more manual pip commands! Just select your algorithm and let dzetsaka handle the rest.
Simply open terminal and type:
python3 -m pip install scikit-learn -U --user
Method 1 - Using QGIS Python console (Recommended):
- Open QGIS
- Go to Plugins → Python Console
- Type:
import subprocess; subprocess.check_call(["/Applications/QGIS.app/Contents/MacOS/bin/pip3", "install", "scikit-learn", "--user"])
Method 2 - Using Terminal:
If you have Python 3 installed globally:
python3 -m pip install scikit-learn -U --user
Note: On some macOS systems, you may also need to install joblib separately:
python3 -m pip install joblib -U --user
After installation, restart QGIS to ensure the libraries are properly loaded.
For QGIS 3.20 and higher: Open OsGeo shell, then :
o4w_env
python3 -m pip install scikit-learn -U --user
For Qgis 3.18 and lower: Open OsGeo shell, then :
py3_env.bat
python3 -m pip install scikit-learn -U --user
Thanks to Alexander Bruy for the tip.
For Qgis 2:
In the OsGeo setup, search for PIP and install it. Then you have few more steps to do. In the explorer, search for OsGeo4W Shell, right click to open it as an administrator. Now use pip in OsGeo Shell like on Linux. Just type :
pip install scikit-learn
If you do not have pip installed, open osgeo4w-setup-x86_64.exe, select Advanced install and install pip.
You can now use all 11 machine learning algorithms including XGBoost and LightGBM!
dzetsaka automatically optimizes algorithm parameters using cross-validation grid search:
Core Algorithms:
- Random Forest (RF): 5-fold CV, optimizes n_estimators and max_features
- SVM: 3-fold CV, optimizes gamma (0.25-4.0) and C (0.1-100)
- KNN: 3-fold CV, optimizes n_neighbors (1-17)
- GMM: No tuning (fastest baseline)
Advanced Algorithms ⭐:
- XGBoost (XGB): 3-fold CV, optimizes n_estimators (50-200), max_depth (3-9), learning_rate (0.01-0.2)
- LightGBM (LGB): 3-fold CV, optimizes n_estimators (50-200), num_leaves (31-100), learning_rate (0.01-0.2)
- Extra Trees (ET): 3-fold CV, optimizes n_estimators and max_features
- Gradient Boosting (GBC): 3-fold CV, optimizes n_estimators and max_depth
- Logistic Regression (LR): 3-fold CV, optimizes C and penalty
- Naive Bayes (NB): Uses optimal default parameters
- MLP: 3-fold CV, optimizes hidden_layer_sizes and learning_rate
dzetsaka automatically handles sparse class labels (e.g., classes 0, 1, 3 - missing class 2):
- Core algorithms: Work natively with sparse labels
- XGBoost/LightGBM: Automatic label encoding/decoding for compatibility
- Seamless workflow: No manual preprocessing required
Advanced users can provide custom parameters through the processing interface using the parameter grid functionality.
- If your raster is spot6scene.tif, you can create your mask under the name spot6scene_mask.tif and the script will detect it automatically.
- If you want to keep your spectral ROI model from an image, you can save your model to use it on another image.
Online dev documentation is available throught the doxygen branch.
If you use dzetsaka in your research and find it useful, please cite Dzetsaka using the following bibtex reference:
@misc{karasiak2016dzetsaka,
title={Dzetsaka Qgis Classification plugin},
author={Karasiak, Nicolas},
url={https://github.com/nkarasiak/dzetsaka},
year={2016},
doi={10.5281/zenodo.2552284}
}
I would like to thank the Guiana Amazonian Park for their trust in my work, and the Master 2 Geomatics Sigma for their excellent lessons in geomatics.
