Two types of feature classification were explored in this exercise using a sample Quickbird image of Salzburg and the eCogntion software. Pixel-based and object-based classification techniques were integrated to classify the satellite image into different categories including water, agricultural land, artificial surface, etc.
The 4-band Quickbird image was loaded into eCognition with the addition of the fifth layer (thematic) pertaining to the (sample) air quality of Salzburg.
First, a chessboard segmentation was done which segmented the image into a grid with a size of 10 units.
At this point, we cannot see that many information about the objects. In the Image Object Information window, several information was added including:
Mean value of the blue band
Max. pixel value of the blue band
Area of the object
Shape Index of the object
Value of the thematic layer
NDVI Value
To add the NDVI value, a customized object feature was created with the use of the “Create New Arithmetic” function.
NDVI=(NIR-RED)/(NIR+RED)
By selecting two different objects from the image, we can see the differences in the Blue band, the Air quality layer, and their NDVI values.
By examining the image object information window, we can infer that the values for Shape Index and Area will not change since the objects are segmented equally like a chessboard. Also, when it comes to the NDVI value, the water features should have a value < 0.
For this particular segmentation method, the Area and Shape Index do not really matter since the objects are segmented into equal sizes.
In this example, a different segmentation procedure called multiresolution segmentation was used. Compared to the previous one, multiresolution segmentation considers the scale in the process of segmenting the image into meaningful objects.
Selecting the two objects again, the boat and the water, we can observer that the shape of the objects dramatically changed as compared to the previous segmentation. This time, the boundaries are clearly defined which makes it more usable in creating “meaningful objects”.
Now that the image has been segmented in a way we wanted it to be, we can proceed to the classification. To distinguish the vegetation from the non-vegetation features, we can use a threshold value for the NDVI.
By using the Update Range, the NDVI values for the new objects are displayed.
Initially, two classes, Water and Vegetation were created, and the objects were classified based on their corresponding NDVI values. For Vegetation, NDVI >= 0.25, and for Water, NDVI < -0.15.
The result of the initial classification is displayed below.
Another classification method was used to identify the boat in the image. This time, a Class-related feature (Relative Border To) was used to identify the objects that are border by water. After creating the boat class, another process was ran to select classify the objects that are 100% surrounded by water.
Next, the air quality layer was integrated to the analysis. The goal is to identify the vegetation areas with low and high air quality. To do this, two separate classes (low air quality and high air quality) were created and nested under the Vegetation class.
To identify which areas that has a high and low air quality, another Assign Class process was added in the process tree, taking into account the mean value of Layer 5 which corresponds to the air quality layer.
High Quality: Mean Layer 5 >= 50
Low Quality: Mean Layer 5 < 50
Collapsing the layer will also automatically display the higher class which is Vegetation, instead of the two classes below it.
Lastly, a new level was created below the initial multiresolution segmentation with scale=200. The scale used this time was 50, while shape and compactness values are both 0.5. With this, smaller objects are created from the previous ones.
Bigger objects from the super-class like Vegetation can be used to directly address the objects in the smaller class, in this case, air quality. For instance, a new “Relations to super objects” was added in the image object information window to determine if a specific object in the smaller class lies within the bigger class of vegetation. Relations to Superobjects features describe an image object by its relations to other image objects of a given class, on a higher image object level in the image object hierarchy.
