What is Image Classification in Remote Sensing?
Image classification is the process of assigning land cover classes to pixels. For example, these 9 global land cover data sets classify images into forest, urban, agriculture and other classes.
In general, these are three main image classification techniques in remote sensing:
- Unsupervised image classification
- Supervised image classification
- Object-based image analysis
Unsupervised and supervised image classification techniques are the two most common approaches. However, object-based classification has been used more lately because it’s useful for high-resolution data.
What are some of the differences between supervised and unsupervised classification? Find out more by reading.
READ MORE: 15 Free Satellite Imagery Data Sources
In unsupervised classification, it first groups pixels into “clusters” based on their properties. In order to create “clusters”, analysts use image clustering algorithms such as K-means and ISODATA. For the most part, they can use this list of free remote sensing software to create land cover maps.
After picking a clustering algorithm, you identify the number of groups you want to generate. For example, you can create 8, 20 or 42 clusters. To be clear, these are unclassified clusters because in the next step, you manually identify each cluster with land cover classes. For example, if you want to classify vegetation and non-vegetation, you’ll have to merge clusters into only 2 clusters.
Overall, unsupervised classification is the most basic technique. Because you don’t need samples for unsupervised classification, it’s an easy way to segment and understand an image.
Unsupervised Classification Steps:
- Generate clusters
- Assign classes
In supervised classification, you select representative samples for each land cover class. The software then uses these “training sites” and applies them to the entire image.
Supervised classification uses the spectral signature defined in the training set. For example, it determines each class on what it resembles most in the training set. The common supervised classification algorithms are maximum likelihood and minimum-distance classification.
Supervised Classification Steps:
- Select training areas
- Generate signature file
Object-Based (or Object-Oriented) Image Analysis Classification
Supervised and unsupervised classification is pixel-based. In other words, it creates square pixels and each pixel has a class. But object-based image classification groups pixels into representative shapes and sizes. This process is multi-resolution segmentation or segment mean shift.
Multiresolution segmentation produces homogenous image objects by grouping pixels. It generates objects with different scales in an image simultaneously. These objects are more meaningful because they represent features in the image.
But most importantly, you can classify objects based on texture, context and geometry.
In OBIA, you can use multiple bands to create objects and then classify them. For example, OBIA can take infrared, elevation or a shapefile to classify each object. Also, layers can have context with each other. For example, objects have proximity and distance relationships between neighbors.
Nearest neighbor (NN) classification is similar to supervised classification. After multi-resolution segmentation, the user identifies sample sites for each land cover class. Next, they define statistics to classify image objects. Finally, nearest neighbor classifies objects based on their resemblance to the training sites and the statistics defined.
Object-Based Nearest Neighbor Classification Steps:
- Perform multiresolution segmentation
- Select training areas
- Define statistics
Remote Sensing Data Trends
In 1972, Landsat-1 was the first satellite to collect Earth reflectance at 60-meter resolution. At this time, unsupervised and supervised classification were the two image classification techniques available. For this spatial resolution, this was sufficient.
However, OBIA has grown significantly as a digital image processing technique.
Over the years, there has been a growing demand for remotely sensed data. There are hundreds of remote sensing applications include food security, environmental concerns and public safety. To meet demand, satellite imagery is aiming at higher spatial resolution at a wider range of frequencies.
Remote Sensing Data Trends:
- More ubiquitous
- Higher spatial resolution
- Wider range of frequencies
But higher resolution images does not guarantee better land cover. The image classification techniques used are a very important factor for better accuracy.
Selection of Image Classification Techniques
Let’s say you want to classify water in a high spatial resolution image.
You decide to choose all pixels with low NDVI in that image. But this could also misclassify other pixels in the image that aren’t water. For this reason, pixel-based classification like unsupervised and supervised classification gives a salt and pepper look.
Humans naturally aggregate spatial information into groups. Multiresolution segmentation does this task by grouping homogenous pixels into objects. Water features are easily recognizable after multiresolution segmentation. This is how humans visualize spatial features.
- When should you use pixel-based (unsupervised and supervised classification)?
- When should you use object-based classification?
As illustrated in this article, spatial resolution is an important factor when selecting image classification techniques.
When you have low spatial resolution, both traditional pixel-based and object-based image classification techniques perform well.
But when you have high spatial resolution, OBIA is superior to traditional pixel based classification.
Unsupervised vs Supervised vs Object-Based Classification
A case study from the University of Arkansas compared object-based vs pixel-based classification. The goal was to compare high and medium spatial resolution imagery.
Overall, object-based classification outperformed both unsupervised and supervised pixel-based classification methods. Because OBIA used both spectral and contextual information, it had higher accuracy. This study is a good example of some of the limitations of pixel-based image classification techniques.
Growth of Object-Based Classification
Pixels are the smallest unit represented in an image. Image classification uses the reflectance statistics for individual pixels.
There has been much growth in the advancements in technology and the availability of high spatial resolution imagery. But image classification techniques should be taken into consideration as well. The spotlight is shining on the object-based image analysis to deliver quality products.
According to Google Scholar’s search results, all image classification techniques have shown steady growth in the number of publications. Recently, object based classification has shown much growth.
This graph displays Google Scholar’s yearly search results using the “AllinTitle:” search phrase.
If you enjoyed this guide to image classification techniques, I recommend that you download the remote sensing image classification infographic.
1. Blaschke T, 2010. Object based image analysis for remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing 65 (2010) 2–16
2. Object-Based Classification vs Pixel-Based Classification: Comparitive Importance of Multi-Resolution Imagery (Robert C. Weih, Jr. and Norman D. Riggan, Jr.)
3. Multiresolution Segmentation: an optimization approach for high quality multi-scale image segmentation (Martin Baatz & Arno Schape)
4. Trimble eCognition Developer: http://www.ecognition.com