Exercise 3: Multitemporal Change Detection and Monitoring - continued
Classify only lava
Create the training field file and replace all classes except lava by NOT LAVA (here named: no). Run the classification again and you will obtain a result showing only pixels classified as lava.
You can also use the prepared file for training fields: TF_2001_lava-no.sav and TF_2003-01-31_lava-no.sav.
Classification results for 2001 and 2003
Post-classification filter
Some pixels are wrongly classified. They are often isolated or in small groups. Filtering is a technique used to remove them. We can also assume that single pixels will not be lava. Therefore we apply a filter that removes isolated pixels, or where only 2 pixels are neighboured (as done in the LEOWorks tutorial on p.57).
For both classified images (with lava, non lava) apply the following filter:
Enhance > Filter > Median 5 x 5
You can see that many pixels have been removed.
Band Arithmetics
Filtered classification results for Landsat images of 2001 and 2003
Band arithmetics (see LEOWorks tutorial p. 23) are used to subtract the filtered classification result of 2001 from the subsequent one (2003) that contains the new lava flows.
Image > Arithmetic > Subtract (filtered image 2003 – filtered image 2001)
Assuming an ideal classification result, only the changes between the two images should show up in white after the band subtraction. There are still many wrongly classified pixels, especially in the south next to Lake Kivo, where a different amount of pixels has been classified as lava / vegetation. You can clearly see one of the two major changes that occurred during the Nyiragongo 2001 and Nyamuragira 2002 eruptions.
1. What can you see after doing the subtraction?
2. Looking at the subtraction result, why do you think one part of the new lava flow is clearly visible but on the other one it is not?
| 
