Unsupervised Classification using the Semi-Automatic Classification Plugin version 7

[Music] hi i'm luca congedo and you're watching from js remote sensing in this tutorial we're going to perform a supervised classification of a sentinel 2 image using the same automatic classification plugin particularly we are going to perform a clustering using the exodata algorithm of course the same method could be applied to any multispectral image please remember to always update the plugin to the latest version available so here in the qgis interface we go in the sap menu the preprocessing menu and select the sentinel to tab here we have the sentinel 2 conversion 2 and we select with this button here the directory containing the sentinel 2 bands here in this case the sentinel-2 sample which is previously downloaded and we select it and as you can see here the table is filled with the name of the bands of the sentinel 2 bands you should check this option here create bandset and use bandset tools and then we click run to start the conversion process we create a new directory for converted bands here click ok and so the conversion process will start as you can see here the progress bar and so when the conversion is completed here we can see here the sentinel two bands loaded in the qgis map here you can see the converted bands and also if we come back to the scp bandsaw tool here we can see that the sentence two bands converted to reflectance are already loaded in this bandsaw two the center wavelength already defined according to the sentinel's two bands here so we can remove this bandsawed one which is empty click here close tab and yes so we have this bandset one which is the input for clustering in the next step [Music] so now that we have the benefit one defined we can go here to band processing clustering here the clustering tool this is the interface of the clustering tool here we select the input bandset which is the advanced one we select here the isodata method for clustering and here we set the distance threshold which is used by isodata for merging signatures set the value of 0.01 and we set here a number of classes of 10 which will be in the classes identified by the ease of data we set here the maximum number of iterations here we can set here 10. of course these parameters can be defined according to your test or your study area and find the optimal values for these parameters we set here the iso data maximum standard deviation as value 0.2 the iso data minimum class size in pixels here in 10 and here we can select the criteria about the creation of seed signatures which are the starting signatures for the clustering process here in this case we'll use the use random seed signatures so random seeds are selected in the image as distance algorithm we can choose between minimum distance spectral angle mapping so we select the minimum distance and we could also save the resulting signal tools to signature release but we can leave this unchecked in this tutorial and then we click run to start the clustering process here we set the name of the output classification for this distance clustering and click save and so the process the clustering process will start here as you can see the progress bar after all the iterations here we have the result of clustering here are the 10 classes you can see here the result of the classification and of course we also have here in the output interface we have the mean spectral signal tour for each link cover class which is of course the result of clustering so now that we have the classification here we should define and set the corresponding length cover class so now we should perform a foot interpretation we can use here this bench one to create a color composite rgb color composite in order to identify each line cover class and the related iso data class so here this is a network color composite you can change here the rgb tool the color composite we can set here for instance this false color composite 732 which corresponds to the new infrared the red and the green bands and if we click with the identify tool of qgis here or the classification you can see here the class 5 for instance of this pixel which belongs to vegetation class you can see here in red in this color composite this other class here with class 1 as you can see here is more related to their soil and so we go on clicking and identifying every class of the iso data classification identify the corresponding land cover class so when we have completed the identification process of every land cover class we go here to post-processing reclassification here we click refresh and select the clustering classification here you can click this button here to calculate unique values as you can see the table is filled with a old value which is the classification and we can simply change the new value which is our land cover classification system so we can set here for instance 3 for the class 1 2 for the class 2 the iso data classification and so on of course in your results the the numbers could vary because of the results of course being a supervised classification the class values could be different in your case then we click run and we set here the output name for instance reclassified here and so here you can see they reclassified the raster we can of course set the symbology here properties you can set here unique values and we of course need to classify all the values and of course change the colors according to our needs so 0 which is unclassified you can set for instance this color blue for the class 1 water you can set another color for the class built up here for instance red and a different color for instance green for vegetation here class three and yellow for bear soil class 4. now we click ok as you can see this is the result of the unsupervised classification reclassified to our cover classes and here you can see there are several classification errors for instance these red areas which should be urban area but actually they are soil you can see vegetation is classified pretty good of course this is a unsupervised classification so it depends on the parameters and our needs we should probably repeat the classification and maybe change the parameters to get better results and we can see here urban area classified correctly so now that we have a land cover classification we can see there are several errors and maybe we could enhance the classification removing isolated pixels as you can see here there are several isolated pixels for instance here could be classification errors so we can use post-processing tools to enhance the classification for instance removing these isolated pixels and create a cleaner classification so here in the cpu menu the classification cf tool here we click here the refresh to refresh the list of the single band rasters here we select reclassified we set here the size threshold so we set here 2 for removing only isolated pixels and we set 8 as pixel connections in order to remove isolate pixels also considering diagonal pixels then we click run and save the output file instance sieve and we can see here the result you can see we have removed isolated pixels of course we could also use other post-processing tools for refining the classification so we have performed the unsupervised classification of a sentinel-2 image that can be a good starting point for assessing and analyzing the spectral signatures of the cover for any comments or questions please join the facebook group of the sema automatic classification plugin thank you for watching you

Channel: Luca Congedo

Views: 5,520

Rating: 4.9694657 out of 5

Keywords: QGIS, Supervised Classification, Land Cover, unsupervised classification, Sentinel-2

Id: kRPNjNvrLPU

Length: 12min 25sec (745 seconds)

Published: Wed Jan 27 2021