top of page

Mapping with context: How object-based classification enhances geospatial insights

Automated machine learning for map classification offers a time and labour-efficient solution to extract insights from high-resolution satellite data of large areas. Here at Adatos, we are incorporating object-based classification into our projects to enhance classification accuracy and provide a wider range of classification approaches for our clients.

As opposed to pixel-based classification, object-based classification segments images into distinct features based on their spatial context before performing class predictions. This method can help to reduce speckle in the output and create a cleaner, intuitive map. Source: Uddinkabir

Object-based classification, an alternative to pixel classification

Pixel-based classification, a common method of classification, focuses on individual pixels without considering spatial relationships or contextual information. Pixel-based classification is suitable for areas that are more homogenous as it enables straightforward differentiation of land cover based on pixel information. On the other hand, object-based classification leverages on contextual details to classify land cover. For some use cases, object-based classification achieves greater accuracy because it ensures objects with complex shapes are not misclassified as noise.

The watershed algorithm is an example of an object-based classification technique. The algorithm treats the image as a topographic surface, where each pixel corresponds to a specific elevation value. Initially, markers are defined in the image, which represent the starting points for the watershed segmentation process. These markers are typically derived from image attributes or user inputs. Then, the algorithm propagates the watershed basins from the markers, following the gradient paths until they meet at the boundaries of objects.

Watershed segmentation provides several benefits. By using the watershed algorithm, it is possible to detect and separate connected objects accurately, even when they exhibit complex shapes or are closely adjacent. Additionally, this technique can preserve the detailed boundaries of objects, leading to improved spatial precision in the resulting segmented image.

In carbon accounting, this segmentation method can be used to detect smaller areas of deforestation activities which might otherwise be misclassified as noise by pixel based algorithms.

But even with the popularity of these algorithms, there are challenges that require fine tuning of these machine learning models.


There are two common challenges in object based classification - under segmentation and over segmentation.

Under-segmentation: The segment carved out by the algorithm contains more than one class of objects. For instance, a bush and a tree are lumped in a single class. This leads to insufficient detail and inaccurate representation of actual land cover.

Over-segmentation: The algorithm wrongly cuts between what should have been one class of objects. For instance, an algorithm could excessively divide a forest of trees into multiple classes due to the variation in individual tree crowns.

Here at Adatos, we work closely with clients to understand their challenges and tailor solutions to meet their unique classification needs and deliver accurate results for informed decision-making and analysis.

We can help

In an era of pressing challenges like environmental degradation, and food security, the need for advanced image analysis techniques is crucial to save time and get the latest insights. Object based classification has the potential to make maps more interpretable and give decision-makers the insights needed to address their challenges head-on. Here at Adatos, we can help you see the bigger picture.


bottom of page