 And welcome to part 2 of the tutorial 6, Synthetic Aperture Radar Image Classification. So, in the first part of the tutorial, we learnt how to perform SAR image preprocessing using snap toolbox. We learnt how to open the image and how to calibrate the data, whereby just to re-itrate radiometric calibration, the aim is to convert backscatter intensity as received by sensor to normalized radar cross section or sigma naught imagery. And then we performed multi-looking after which the image shall have approximately square pixels because we are converting from slant range to ground range. And then we tried to subset the image and perform speckle filtering again by speckle and referring to the random constructive and destructive interference that results in salt and pepper noise throughout the image that is grainy type appearance throughout the image. So, to reduce this effect, we applied speckle filters to be more specific, we applied leaf filter as an example and then we also saw how to perform de-squeuing and terrain correction. Here terrain correction is performed to correct for geometric distortions. And then what we did was we created a third band by performing band resuing using snap. We created an RGB image and then ending with unsupervised classification of that image. So, we assigned certain clusters and then performed unsupervised classification to see the classification map. So, this was done as part of part 1. In this section of the tutorial 6, we shall be learning how to perform supervised classification again on the same LO's Pulsar SAR imagery. So as before, the data set that we will be using is the LO's Pulsar data downloaded for the Mumbai region which has been used in part 1 of the tutorial 6, a little bit about supervised classification what it is and how to perform this before we begin the exercise. So, for unsupervised classification, you may have noticed that you are entering the number of classes which means you do not have a prior a priori information that is we do not know beforehand how many land cover types are present in the imagery or how many should be specify as classes. The computer will automatically group these pixels with similar characteristics into unique clusters that follow some statistically determined criteria. Now, coming on to supervised classification, so here the identity as well as location of land cover types, let it be urban area or agriculture area or water body, so everything are known a priori known beforehand. This is by either field data collection or your own personal experience or by conducting surveys and so on. So, in the image, the analyst will typically try to locate these areas and a supervised classification requires a training data sites whose spectral characteristics are known. So, in specific for this particular tutorial, we shall have a look at the maximum likelihood method of supervised classification. Now, assume you have a cloud of points, you know, cloud of points which may represent image pixels of a certain class say urban region, okay. So, the geometrical shape of this point cloud, I am going to call it as point cloud. So, the geometrical shape of this point cloud can be described by an ellipsoid, okay, ellipsoid. Now, say the image you are working with has k number of classes in which urban area is just one class, okay. Say the image has fallow land, it has water bodies, it has vegetations and so on. So, in maximum likelihood classification, we tend to determine the equi-probability contours for all k classes which means we are trying to find out the probability of each pixel belonging to each of the k classes. And that pixel can be finally assigned to the class for which the probability of membership is highest, okay. For example, I take the first pixel of the image and then I try to find the probability of that pixel to belong to each of these classes. Probability of that pixel to belong to the urban area class, to belong to the water body class, vegetation class and so on. And then finally, at the end of maximum likelihood classification, I am going to assign that pixel to the class to which it has highest probability, okay. So, what you see here, the density function of a normal random variable with mean, mu and variance sigma square, why I am showing you this is? Because in maximum likelihood classification algorithm, the probability that a pixel is a member of a class is given by the multivariate normal density. So, to reiterate in this particular algorithm, the first step is you pick up training pixels, training data pertaining to each class. For example, if there is a water body class in your image, you zoom in and then pick out those pixels that are representative of water bodies. And then similarly, you pick up training pixels for each class and then the maximum likelihood classifier will evaluate the variance and covariance while classifying. And then finally, the probability density function is calculated for each class, which means probability of each pixel belonging to each class is calculated. So, let us try to understand this further through the exercise. So, if you remember, in part 1, what we did was, we stopped at subsetting the image which has been subjected to calibration, radiometric calibration, which has been subjected to multi-looking that is underscore ML, underscore SPKs, it has been subjected to speckle filtering and then de-squeuing that is DSK and then TC that is terrain correction. So, this is what you see in front of you is an RGB image that has been subjected to radiometric calibration and geometric correction. It has been subjected to speckle filtering, multi-looking, de-squeuing and then what you see here is the RGB image, which means it is ready for classification. So, what I do first is I try to click on the vector and create the vector data container. I am going to name it class 1 to keep it generic. If you are absolutely sure, you can also give it specific names like water bodies or urban areas and so on, I am going to keep it more generic. So, what we do is, you can highlight the image, create a new vector data container, name it and then use the select polygon tool which will give you something like a rubber band like tool. At the click of a mouse, it is going to give you a tool which will allow you to select your area of interest or AOI. So, as this is just for representation purpose, I am going to show you how to select the area of interest and in between you know if you want to pan or say zoom to a particular area, you can again click and then zoom in, repeat, you can create on the polygon tool just as I have done and then select which container your selected pixels are going to go into and in between I am zooming in. So, what I am doing is, I am trying to pick out pixels of the same class from different areas of the image so that it is not clustered in one particular region. Because training sites within image need to be very carefully selected and they should be representative pixels for that particular land use class. So, what I am doing is, I am viewing the image, zooming in to specific sections of the image and I am selecting areas of interest using the rubber band type tool that is polygon shape file. Now, conversely there is also another method to select an area of interest so that instead of using this particular tool, the analysis may need a specific seed location and you get to select that seed location using a cursor. So, you click on any part of the image that particular seed location say X comma Y, you are indicating it as the location and then it begins evaluating the neighboring pixel values and slowly starts expanding like some sort of an amoeba you know as long as it finds a pixel nearby with similar spectral characteristic, this process is going to be continued. And it is one of the most effective means of collecting homogeneous training data but here what you see in the screen is I am using the inbuilt tool that is available in SNAP that allows one to create a vector class which we can name as any land cover class say urban area, water body and so on. So, here I have kept it generic as class 1, class 2, class 3, class 4, etc. The final aim or goal is you should collect well distributed training samples that are non-autocorrelated. I am using the term non-autocorrelated because as training data collected from autocorrelated data is going to have variance and we do not want that. So if you have noticed I have come up with four classes again this is for representation purpose and every time I am zooming into a particular class I am using the tool selecting an area and when I click on the image it is going to come up with a pop-up asking me which class should I assign the pixels to, okay, alright. So there you see the training classes have been selected in different parts of the image. You know just to mention that the SNAP gives you options which helps you create the scatter plots or for example correlation plots. I can just specify the x-axis and y-axis say for example I have specified h h, sigma naught h h image as 1 axis and as y-axis have specified sigma hv image. I also get to create scatter plots like this, okay, with the limits specified, okay. Similarly, I can have several plots that I can create say histograms or correlative plot and so on, okay. Just to show you an example I can see the histogram that has been created for each class, for class 1, class 3, class 2 and class 4. To perform supervised classification we have to go to the option raster classification, supervised classification and maximum likelihood classifier wherein the pop-up will ask me to specify the image and the feature bands. So in this case I am going to select everything and then it takes a few seconds to complete the classification process. You can see towards the left side the classification result has been displayed. So I am going to click to band and then click to label the classes, okay. Now yes, so you see the output which is displayed here, okay. As before I can even play around with the colors say I am not happy with the color assigned to one particular LandCover class. Lot of options are there in SNAP. So I am going to go to view tool windows and then color manipulation which will allow me to specify change the colors that has been assigned to different class. Now say I am not happy with the classification map. So let me try to reclassify with more number of samples. Let me go to classification, supervised classification and number of algorithms are built in. I am going to increase the number of training samples and try classifying again because the accuracy of supervised classification depends upon how good your training data set is. So in this particular exercise I quickly showed you how to collect the data, training data of each class from the image. A great deal of care and effort should go into picking up these training data sites because they are representative training pixels for each class, okay. So I have completed the classification again and let me see if the image is any better. See here for the training data site I have to refer to the sample size and representativeness of the sample. Now sample size does not mean bigger the sample better your results, okay. How you select it and how representative these training pixels are for each class that is more important. Here I am just trying to show you the results when we try to increase the number of samples assuming the samples that we have collected are representative training pixels, okay. Now the number of variables whose statistical properties are to be estimated and the amount of variability in each class all these are important when it comes to supervised classification. So you see this is the output I am getting when I am using 2000 number of samples for each class by samples I mean training data set, okay. As before let us try to display the same you may feel it is slightly better, is not it. And if you use the different colors sometimes it helps to bring out the contrast and you are able to judge whether your classified map is better or not. So that is the whole purpose of color manipulation, yes, okay. So there are several inbuilt algorithms in SNAP that help you to perform supervised classification and in this part of the tutorial we just showed you how to work with one particular algorithm that is maximum likelihood algorithm. Now so overall in this tutorial we have learnt how to perform pre-processing of SAR imagery and how to carry out unsupervised classification and supervised classification. So let me hope that you found this section of the tutorial useful I will see you in the next lecture. Thank you.