Object-based image analysis (OBIA)

Human visual perception almost always outperforms computer image processing algorithms, For example, your brain knows a river when it sees one. But a computer can’t distinguish rivers from lakes, roads or sewage treatment plants.

Especially with the spatially extreme and spectrally minimal resolution UAV image data, a change in thinking must take place. It makes more sense to think in terms of objects or entities to be identified rather than classifying in terms of individual pixels. The basic principle of object-based image analysis (OBIA) is to segment first and then classify.

The segmentation process is algorithm-dependent but looks iteratively for similarities in space, structure and channel dimensions for grouping neighboring and similar pixels into objects. This segments are classified in a next step using supervised training data.

General Workflow

The example of OBIA classification is typical of the manual performance of such operations using a software package. It consists of the following steps:

data acquisition (orthophoto, training data).
generation of spatial segments on the basis of the
extraction of suitable description parameters
model training
classification of the input data set.

In software-based processing, additional steps often have to be performed for technical reasons. Even more often, the processing of the individual steps is not necessarily linear, since intermediate results are used repeatedly in different steps. The following figure shows below step by step described process as a graphical model, which you can integrate in QGIS as a tool in the Processing Box.

OBIA classification Workflow for Orthoimages

OBIA classification Workflow for Orthoimages

For reference you may Download the basic data. In Addition you may download the upper OBIA-workflow as an QGIS-Model. You can add this to your QGIS project with pushing the first icon “Models” on the processing sidebar and choose Add Model to Toolbox. Please note that is running with fixed default values. For modifying it you need to right-click on the model and choose Edit Model.

Step by step tutorial

In the following step by step guide an OBIA approach with QGIS and the OTB Toolbox is carried out as an template example. There are many segmentation algorithms and integrated classification methods. The Mean-shift method used here and the subsequent training with Support Vector Machine is a robust and common method. Especially the extension of the feature space (here called Range Radius) and the search space (Spatial Radius) as well as the size of the segmented objects (Minimum Region size) is crucial for a satisfying result. The principle is transferable to the different forms of the OBIA and despite abundant literature and some good instructions it is a free empirical game.

Step-1 Create Training Sample Points by manual digitizing

If you need to learn how to digitize with QGIS you may follow this tutorial. However we will only digitize Points and not polygons.

Activate under Main Menu->Settings->Digitize and check Reuse last entered attribute values. this will makes in much more comfortable to digitize training points of one class in series.

Create a point vector file and digitize the following classes:

class	CLASS_ID
water	1
meadows	2
meadows-rich	3
bare-soil-dry	4
crop	5
green-trees-shrubs	6
dead-wood	7
other	8

Provide at minimum 10 widely spread sampling points.

Save this file naming it sample.gpgk.

< Keep in mind >

Here, the training data is digitized on the screen (as is often the case) using the god’s eye method as an example. For a rough classification of the selected classes, this is of course useful and sufficiently accurate.

However, please keep in mind that these training samples are usually collected in the field. This means that utilizing a GPS or exact map work, the positions, and their affiliations are collected and entered into a corresponding table. Often this is accompanied by a vegetation survey, soil survey, or limnological or other surveys.

Step-2 Segmentation

In the search field of the Processing Toolbox type segmentation and double click Segmentation.

select the input image: example-5.tif
select meanshift from the drop-down list Segmentation algorithm
The Spatial Radius value can be set to 25. This is determining the spatial range of the segementation and is also experimental. Try to identify the scale of your major classes in pixel
The Range Radius value can be set to 25. We are dealing with RGB images that have a range 0-255. The optimal value depends on the datatype, the dynamic range of the input image and requires experimental trials for the specific classification objectives
Set Minimum Region size in pixels to 25. The minimum size of a region (in pixel unit) in segmentation. Smaller clusters will merged to the neighboring cluster with the closest radiometry
keep Processing mode as Vector
Tick 8-neighborhood connectivity
Set Minimum object size in pixels to 200
Name the Output vector file as segments-meanshift.shp
Push Run

Check the results: Load the output vector file segments-meanshift.shp into your QGIS session and place it on top of the image example-5.tif
Colorize the vector layer in the QGIS Layers window. Right click Properties->Symbology->Simple Fill, Fill Style: No Brush and Stroke color: white.
Check the features with a visual inspection. Is the result goal keeping? If not, start over… .

Step-3 Feature extraction

Type zonalstats in the search field of the Processing Toolbox and open the tool ZonalStatistics. You find it under the image manipulation section of OTB.

Select as input image: example-5.tif.
Select input vector data the vector file with segments from above segmentation segments-meanshift.shp
Name for the output vector: segments-meanshift-zonal.shp.
Push Run.

Step-4 Joining training data with segements

Type in the search field of the Processing Toolbox join and double click Join Attributes by Location.

select the Base Layer: segments-meanshift-zonal.shp
select the Join Layer: sample.shp
Join Type: choose Take Attributes of the first matching…
Tick Discard records which could not be joined
Provide an output filename

Sometimes the geometry is broken. Type fix in the search field of the Processing Toolbox and open Fix Geometries which will in most cases do the job.

Step-5 Training

Type train in the search field of the Processing Toolbox and open TrainVectorClassifier

In the field Vector Data List select the correct vector file by clicking … and browse directly to the file containing the training area polygons segments-meanshift-zonal.shp
Provide the Output model filename as lahn-gi-spann-obia.model
In the field Field names for training features copy and paste "mean_0 stdev_0 mean_1 stdev_1 mean_3 stdev_3 mean_2 stdev_2"
The name of Field containing the class id for supervision is CLASS_ID
Classifier to use for training: libsvm Usually the straighforward Support Vector Machine is doing a good job
SVM Kernel Type: linear
SVM Model Type: csvc
Set Parameters optimization to ON
Push Run

Step-6 Classification

Type class in the search field of the Processing Toolbox and open VectorClassifier

In the field Vector Data select manually the correct vector file clicking … and browse directly to the file containing the training area polygons containing segments and features for the whole image lahn-gi-spann-segments-meanshift-zonal.shp
The name of the upper input model file is lahn-gi-spann-obia.model
Output field containing the class is CLASS_ID
Copy and paste into the field Field names to be calculated same attributes as above: "mean_0 stdev_0 mean_1 stdev_1 mean_3 stdev_3 mean_2 stdev_2"
Name the output vector data lahn-gi-spann-classified_obia.shp.
Push Run
Finally load the output vector file into QGIS and apply the same QGIS style used for the training data. Layer->Layer properties->Symbology->Style->Load style....

Result

You will see partly predominantly excellent classification. However, there are also significant misclassifications.

What could be the reason for that?
What is a weakness of this approach?
Any suggestion to improve?