Features extraction

The term Feature Extraction refers to techniques to extract meaningful and informative attributes from images. The features can be local statistical moments, edges,... or even textures !

Image feature extraction creates representations that are more compact and meaningful than the raw pixel data. The extracted features minimize irrelevant and redundant data, often referred to as "noise", which improves the robustness and accuracy of models (for image classification, object detection,...).

By reducing the data's dimensionality, feature extraction also cuts down the computational resources needed for real-time applications.

This page presents various libraries and tools used for feature extraction, categorized by their main functions. These range from traditional computer vision techniques to machine learning algorithms. Note that the list of tools presented here is non-exhaustive.

Practical example

Features extraction can be used for crop segmentation to divide agricultural fields into distinct regions based on crop types or health, enabling precise monitoring and management.

Original Image	Image manipulation	Segmentation	Classification
Bring your own satellite image (Pleiades, WorldView, PNEO, CO3D,...)	Computation of the NDVI (Normalized Difference Vegetation Index) from red and near-infrared bands. This metric is used to distinguish land cover types, and to monitor plant health and growth over time.	All plots are separated into spatial zones with consistent radiometries using a segmentation method (for exemple Watershed)	A crop type has been associated to each plot using machine learning (in the image above we used a random forest classifier)

Image preprocessing

Opening the image

Opening satellite or raster images typically involves reading geospatial data formats and extracting metadata. Please refer to the tools presented in the Data access page.

Image Manipulation (Pixel based)

These tools focus on pixel-wise transformations, band-level operations, and general image processing tasks, such as index calculation and inter-band arithmetic operations.

Function	Tools	Related use cases
Pixel based image manipulation	OTB (Orfeo Toolbox) – otbcli_RadiometricIndices	Radiometric indices extraction
	Georastertools – radioindice	Compute various vegetation and environmental indices on satellite or raster data
	GEOWombat	Efficient computation of indices across large raster datasets

Statistics computation (Pixel neighborhood or region based)

These libraries help compute statistical measures from images or geospatial data over pixel neighborhoods or regions:

Function	Tools	Related use cases
Statistics computation	OTB (Orfeo Toolbox) – otbcli_LocalStatisticExtraction	Computes local statistics over a specified pixel neighborhood
	Georastertools – zonalstats	Compute statistical measures over zones defined by a vector shapefile
	DEMcompare	Compare digital elevation models (DEM) through statistical metrics

Segmentation

Segmentation aims to partition an image into meaningful regions that correspond to objects or areas of interest. Pixels with similar characteristics (such as intensity, texture, or spectral signature) are grouped into coherent segments or regions. In a SAR image, segmentation can be used to delineate flooded versus non-flooded areas by clustering pixels based on their backscatter intensity.

Segmentation methods can range from simple thresholding techniques to more advanced algorithms, such as the watershed method or region-growing approaches.

Function	Tools	Related use cases
2D/3D rasters	OTB (Orfeo Toolbox)	Segmentation algorithms for remote sensing applications
	Magiclip	Grabcut algorithm for remote sensing optical imagery
	Scikit-Image	Segmentation methods (watershed, k-means, random walker, Chan-Vese...)
	OpenCV	Multiple segmentation methods, tutorials are available for thresholding and watershed methods
2D/3D point cloud	Cloud Compare	Generic point cloud processing software
	Scikit-learn	Offers a dedicated clustering module

Segmentation tools often use morphological operations to enhance results. These operations—such as dilation, erosion, opening, and closing—are especially useful for refining binary masks, removing noise, and separating connected objects (for example. buildings, roads,...).

Function	Tools	Related use cases
Morphological operations	OTB (Orfeo Toolbox) - otbcli_BinaryMorphologicalOperation	Application dedicated to Morphological features extraction
	Scikit-Image	Contains a dedicated morphology module (closing, opening, skeletonization,...)
	OpenCV	Widely used library for preprocessing steps such as filling gaps in segmented regions or eliminating small artifacts in satellite images. A short guide is available available

Classification (Supervised / Unsupervised Learning)

Classification involves assigning predefined labels to a point, pixel or region based on its features (e.g. spectral values, texture, shape). It is a decision-making process that uses statistical or machine learning models to categorize image data.

In satellite imagery, classification helps determine what each pixel or region represents (e.g., water, forest, urban). A Random Forest classifier can be used to label segments in an image as "urban", "vegetation" or "water".

Machine learning libraries for classifying features extracted from images:

Function	Tools	Related use cases
Classification	OTB (Orfeo Toolbox)	Provides applications to train a supervised or unsupervised classifier from different set of features and to use the generated classifier for vector data classification. A guide is available
	Scikit-learn	Traditional machine learning models for supervised and unsupervised classification
	iota2	Automated land cover classification from satellite images
	GEOWombat	Scalable tools for satellite image classification using machine learning models across large geospatial datasets
	XGBoost, LightGBM	Gradient boosting for efficient and large-scale data classification