Lab 7: Object-based Classification

Background
The goal of this lab is to learn and utilize eCognition, a state of the art object-based processing tool. First  homogeneous spatial and spectral clusters are produced over the image or known as objects. a random forest and support vector machine classifiers will be used which require training samples. Finally an output will be produced showing the newly classified image.
Methods
First a project was created in eCognition, then objects need to be created on the image, this is done through the process tree utilizing multiresolution segmentation, with a scale parameter of 9. This produced the objects shown in figure 1 and 2. Finally training samples are taken to be put in the process tree. this trains the classifier.
Figure 1: This shows a mess of blue over the image, but thwese are all polygons that contain similar pixels.

Figure 2: This image shows a zoomed in section of figure 1, this allows the actual polygons to be seen.
Next the same process was performed but for an SVM rather than random forest. This was similar to the random forest, yet there were different parameters.

Then finally a UAS image was classified, this was done using an RF classifier, except the scale parameter was 120, this allowed for the best grouping of pixels over the UAS imagery (Figure 3).
Figure 3: This shows the polygons, or objects over the UAS image.

Results
The results of performing a random forest object based classification provide a fairly accurate classification (Figure 4). There was some error in some places where urban areas was classified as bare soil, but this was fixed by manually editing polygons. The SVM classifier produced a similar output (Figure 5), with similar errors of miss categorized pixels.

Figure 4: This shows the random forest object based classification over Eau Claire and Chippewa Counties.
Figure 5: This shows the support vector machines object based classification over Eau Claire and Chippewa Counties.
Finally a Random Forest classification was performed on UAS imagery, it worked well, but error is seen, where a vehicle is classified as a building, and additionally the building on the eastern side apparently has a car on its roof. 





Sources

Landsat satellite image is from Earth Resources Observation and Science Center, United States Geological Survey. 
UAS image is from UWEC Geography & Anthropology UAS Center. 





Comments

Post a Comment

Popular posts from this blog

Lab 6: Digital Change Detection

Image
Background The goal of this lab is to develop an image that shows change detection of land surface features over time. First a means of performing a qualitative change detection will be performed, then we will quantify post-classification change detection, and a model will be developed that will map detailed change of land use land cover images over time. Methods Part 1: Change detection using Write Function Memory Insertion The first part of the lab, a Write Function Memory Insertion is utilized. This processes utilizes near infrared bands, when the two images are brought into the Write Function, an output image is produced, and the image has areas highlighted in pinkish red, these are areas that have experienced change. Specifically in this lab, ec_envs1991.img and ec_envs_2011.img were used to perform this change detection. Part 2: Post-classification comparison change detection   In this part of the lab a From-to change detection using classified images of the Milw...
Read more

Lab 5: classification Accuracy Assessment

Image
Goal The goal of this lab is to understand how to evaluate the accuracy of classification results. Accuracy assessments are mandatory after performing an image classification, this is a portion of the post-processing stage of remotely sensed data. Methods An accuracy assessment is performed on a supervised and an unsupervised classification from labs 3 and 4. This is done by utilizing the accuracy assessment tool in Erdas. A high resolution reference image of the same study area is used to perform the assessment. This is done by generating  random points on the reference image, 125 random points in this lab,  and classifying those points based on the high resolution image utilizing the same classification codes for water, forest, agriculture, etc. (Table 1). By identifying these points, a comparison report is then produced (Figure 1). This allows for an error matrix to be created from the results of comparing the random points to the classified images. Table 1: ...
Read more

Lab 10: Radar Remote Sensing

Image
Goal The goal of this lab is to understand how to perform basic preprocessing and processing of radar images. Specifically noise reduction through speckle filtering, spectral and spatial enhancement, multi-sensor fusion, texture analysis, polarimetric processing, and slant-range to ground-range conversion. When this lab is completed, the basic knowledge of processing on any type of synthetic aperture radar image will be gained. Methods Speckle Supression In this lab radar imagery is analyzed and processed through Erdas Imagine, and ENVI. Noise reduction methods for preprocessing of radar imagery was performed through speckle filtering, edge enhancement and image enhancement. To perform speckle filtering, Radar speckle suppression tool is used. this removes common salt and pepper effect. this is done by utilizing a calculated coefficient of variation (Figure 1). Figure 1: coefficient of variation is 0.274552. This variable is then used to run the Speckle Suppression Functio...
Read more