Land Cover Accuracy Assessment in QGIS (3/n)

hi guys welcome to the third video of the series in this video i will demonstrate how to perform accuracy assessment of a classified land cover data using qgis uh accuracy assessment of land code data is crucial to understand how good the classification of data is or to correct systematic errors in the data especially if you are going to use it for further scientific analysis or even if you let us say want to train a machine learning model using this at a later stage you may want to know the quality of your training data so let's get started okay here we have the land cover data that i demonstrated in the previous video which is this and then the satellite data which we use to prepare this i've changed the band combination to 543 like before okay so without delaying first we go to uh view and then panels scp dock check this and then we go to band set in single bands list i refresh it and then add it to the list so now we have this as a listed product now i will create a new training data set i'll copy the location here i will click on basically this is accuracy assessment data set okay now we want random cells in this area such that each class is equally represented for that i'll go to scp basic tools multiple roi okay so the catch here is that in this bar you have to make this minimum and maximum one here also you make it one how many number of points do you want for each class i'll take 10 for now but i would say take anywhere between 25 to 50 points per class so that you do accuracy assessment on 100 to 200 points now i check this box which says stratified for all values as in like it will take 10 10 values for all classes in the classified raster and what are the classes let's say if you have four classes and you want to take only two then i'll remove these and keep only the first two but i want to do accuracy assessment on all four classes so i'll keep this just like they are okay so then i'll click click on create points okay now that points have been listed i'll keep everything default change nothing and click on run this may take a minute to finish okay the processing completed in roughly two minutes and all the points are listed here now um you can close this for now so it has 10 points for each class but we don't know which point belongs to which class and that's the catch here that we should not see what class that particular point belongs to i am removing this layer and the key point is to classify each point manually based only on the raw multispectral data so that we can compare our observation here with the classified raster and then do the accuracy assessment i would recommend not to start in order because uh points are in order that we created them for instance first and will be urban and vegetation and so on if you when you do it for a lot of points uh then it will be it will make much more sense i will randomly jump to any point uh based on this i will judge what that point belongs to i think this is urban so i will come back here and keep it one i'll unselect it i'll go to some other point zoom to layer that's water body according to me i'll come back here and make it three the other one then i'll keep this in vegetation class so i'll make two here and unselect that then i'll go again to some pixel zoom to it i'll keep this in water class similarly i'll go and collect all the samples all 40 samples i have now collected all the samples so as you change the values here it will automatically create uh subclasses here uh 1 3 2 and 4 and so on so the catch is that you don't only have to look at that pixel you need to zoom in zoom out and then look at the surroundings a bit and then take a distance a lot of pixels can be confusing also because after all this is a medium resolution satellite data 30 meters and there are obviously mixed pixels so that is okay but this accuracy assessment will give you at least a rough idea of if certain classes are systematically mixing up with others class and other classes and so on to check the accuracy we need to add the classified land cover data so that we have the classified data in this both i'll go to scp post processing accuracy and i'll refresh this so that this updates ah in the first one classification i'll choose the land cover layer in the second one the vector raster where classes are i will select this accuracy assessment layer that i have created and in this vector field i'll choose this mcid because that's what we are dealing with again if you in the previous video used subclasses also you can always go with cid which will give you sub class wise classification metric so i'll click on mcid and then hit run it prompts me to save the error raster i'll say accuracy roster the classification results are out now so first so first you see let's say uh accuracy raster in this you you can see a couple of points i'm closing this sap doc now in this you see a couple of points with different numbers so wherever your training points were you will see these numbers and that's what they are the table here will have information on these numbers for instance one represents all a built up classes that were classified as built up and so on and so cells with value 10 represent built up classes that are actually one but in classified raster it's four others class and so on this is a main thing the confusion matrix which will give you an idea of the actual ground truth classes and their comparison with classified data so manually i classified 13 cells as built up out of which nine were built up in a classified raster also one was vegetation one was water and two were in others category so the thing is that you have to check at this diagonal values one for one two to two three two three and four to four they should be as high as possible this one looks okay also because i did it only for 40 points uh to be very sure and confident you can go ahead and collect more samples when you come down here you have producers and users accuracy class wise class one two three four and then you have the overall accuracy of the raster uh and the uh kappa classification coefficient so this table is also available as a csv in the same folder you can open it the values are in this form you need to select this go to data text to columns limited separate by tab if you don't do it then everything is in one column do this and then you have the same table here also so that's how you perform the accuracy statement make sure that your uh overall accuracy kappa coefficient and these diagonal values are as high as possible this is this can be a tedious task but that's the only way around of being confident about your data if you really want to randomize your points so that they are not in order as we collected you can easily do that just right click on this temporary layer open attributes and click on editing click on field calculator so in this update existing field we will update fid and here we will create random numbers from let us say 1 200 okay so we will have random numbers now i'll save it here i'll export this to the directory i'll export this as this is hp file i won't add it to the map okay so now that is done i'll delete all the signatures here and then i will go to scp basic tools import signatures here i'll select a vector accuracy assessment dot shp here mcid can be mcid you just need to match the fields and only in any one let's say ceid field i will replace it with a width fid that we created with random numbers i have unchecked this box which is a calculate signature because we already have signature we are going to create signature in that and then import vector ok so the points are here with these random numbers right now they are in order of 10 10 classes when i click on cid which is random numbers this set of training points they shuffle so now you can easily go ahead and collect the signatures one by one so hope you will be able to do it on your own on your own data thanks for watching this video

Channel: Pratyush Tripathy

Views: 599

Rating: 4.7647057 out of 5

Keywords: remote sensing, qgis, gis, landsat, land cover, supervised classification, confusion matrix, accuracy assessment

Id: Uk4UOdAP8iE

Length: 11min 27sec (687 seconds)

Published: Sat Jun 26 2021