 Welcome to today's lecture. So, let us begin with a new module today. And before I show you the contents, till now, if you quickly think about the lectures attended so far, you can understand that we were concerned more about the measurement as in what exactly is measured by a radar system, what are the fundamental properties of a synthetic aperture radar image, you know we were more concerned about the measurement, about the image fundamental properties. Now let me ask you that why are you possibly measuring SAR images? Your obvious answer would be of course for recognition of objects in a scene, for recognition of targets in a scene, isn't it? Now say you visit a shop say it is a stationary shop and when you look around, you do see a certain order in which each item has been arranged, isn't it? Say all the pencils and books and pens all of them are neatly arranged in order such that the items which are similar in nature are grouped together, okay? Let me re-itrate similar in nature are grouped together. Now even in digital image processing using microwave images, we can perform a similar operation wherein we can represent pixels or classify pixels which represent a particular feature. Now here by feature I mean urban area or sand, fallow land, vegetation, water body, note here that I have used the term classify but then you can argue that I know how a synthetic aperture radar image looks like, you know it is monochrome, very difficult to visually identify what is what, very difficult to demarcate features as it is. Then if I cannot see the features, how am I going to classify? Very valid point. So, this is what we shall be covering as part of module 3 on image classification. We will learn about classification of synthetic aperture radar images. We will cover supervised, unsupervised and fuzzy classification methods and their accuracy assessment. So, let us begin. We are in lecture 1 of module 3. See before I discuss about any specific methods to classify synthetic aperture radar data, I shall take the liberty to first introduce you to image classification in general, the concept of image classification and this module was created thinking about two audiences in mind. First is for those who want a non-mathematical introduction to the underlying concepts, non-mathematical. Second is for a slightly advanced audience who prefer mathematical explanations. So, proceeding forward, what you see in front of you is the land use land cover map of India 2015-2016 from NRSC, National Remote Sensing Center ISRO. You can see that this map gives you different kinds of features. Built-up area is designed a specific color, cariff crop or plantations, evergreen forest, water bodies all are designed a particular color. What you see here is a map wherein the pixel identified as built-up or cariff crop, it is assigned one particular label, one particular numerical label, it can be 1 or 2 or 3 and so on. Which means this is not an image but it is a map. In image classification, the output what we are expecting will look something like this. So, let us try to define it little bit more in detail. The process of image classification shall consist of two stages. And the first stage, we need to recognize or identify real world objects, okay, recognition. And by real world objects, I mean water body, vegetation, buildings, built-up area and so on. And second is labeling of pixels to be classified, okay. Let me re-itrate. The process of image classification shall consist of two stages. Number one is recognizing or identifying the real world objects. And number two is labeling the pixels to be classified. Again, what you see here is a map wherein each pixel, each pixel of an image has been identified according to some feature, some real world object on the earth surface. And it is assigned a particular numerical label, okay, fine. So, what is the need of image classification? Why are we learning this topic? As I mentioned in the example before we began the class that is, you demarcate similar objects in shell according to their nature in a shop. When you enter a shop, there is a certain kind of order that you see when you look around. In image classification, it involves automatically categorizing all the pixels in an image into land cover classes, which means properties of a pixel are used to label that pixel. They currently form the backbone of most of the multi-spectral classification processes. Again, by image classification, I mean identifying the pattern associated with each pixel position in an image in terms of the characteristics of the objects that are present at the corresponding point on the earth surface. Now think about synthetic aperture radar image, image classification, identifying the patterns associated with each pixel position in an image in terms of the characteristics of the objects that are present at the corresponding point on the earth surface. And it involves an automatic process automatically categorizing all the pixels in an image into land cover classes. And we shall start with spectral pattern recognition which utilizes the values of a pixel, okay, let it be backscatter values or digital numbers in various wave bands, okay. And to summarize, we shall be covering two approaches to pixel labeling. One is supervised, second is unsupervised, two approaches to pixel labeling. We will also get an introduction about fuzzy classification but that shall come later for now understand that there are broadly two processes, two approaches to pixel labeling, one is supervised, another is unsupervised. Now let me give you another example, you know, say you are given the task of creating say a facial recognition system in your department or in your office, you may have seen such systems in movies, right. So how do you possibly begin to create such a system? Think about it. The aim is that only the students or the faculties or the staff of your department should be granted access to enter that building based on your developed facial recognition system. Now in case you are a working professional, let me re-itrate. The aim is you need to create a facial recognition system such that only the employees of your office are granted access to the office. If you think about it, the technology should essentially have a database, isn't it? A database, database of what? Database of faces. And every time someone wants access, their own face would need to be matched to a database for similarity, isn't it? So I am using the word similarity, which means if similarity exists, then access is granted and if there is no similarity, access is not granted. Essentially, when this comparison is being made, the system is checking for patterns, checking for similarity. Okay? A similar approach is followed in the case of image classification. So we will understand this further when we try to gain knowledge about the geometrical model of classification. Okay? So just to understand this a little bit more further, let us try to understand the geometrical basis of classification. Imagine that as a hydrologist, you have two sets of values with you. Okay? Say they are the river discharge and the catchment area. And say for different catchment areas, you have the corresponding river discharge for a hypothetical set of basins. They are a pair of values. Catchment 1, river discharge 1, catchment 2, river discharge 2, catchment 3, river discharge 3 and so on. Okay? You have a pair of values of river discharge and catchment area. Now if we plot these values along the x-axis and along the y-axis, we are presented with a diagram like this. Isn't it? Let us scatter plot. On the x-axis, we have catchment area and on the y-axis, we have river discharge values. When we look at this plot, we can quickly infer two things. Isn't it? Number one, axis of Euclidean space can be used to represent both catchment area and river discharge. Axis of Euclidean space. The axes are the x and y-axis of a Cartesian coordinate system. Axis of Euclidean space. Again our eyes and brain together tell us that visually there appears to be two groups or let us just say two clusters of points when you look at this diagram. Two clusters of points visually you get a feeling. We can get to estimate the compactness of the two clusters which means that say I am asking you which cluster do these remaining points belong. You see one group of clusters I will call it as one which are tightly knit close together and you see another group of points I will call it as group two which are close together. Now my question is you have two other points which neither belong to one neither belong to two. So which group do they possibly belong to? How will you respond? So you can make an intuitive guess that say if this value if it is closer to the center of cluster one it can be assigned to cluster one. Again if this value is closer to the center of cluster two it can be assigned to cluster two. This means that we are using a two dimensional Euclidean space to estimate distance between points. Say I am closing my eyes and one of my students is standing across the room you know. I am asking her to direct me to where she is standing. I am closing my eyes and then I am asking her to direct me to her where she is standing. Probably she might say that I have to take a certain number of steps along the x direction and then further a certain number of steps along the y direction to direct me to where she is standing to her location. So even here we are using the Cartesian coordinate system, isn't it? Now if we wish we can even draw a line to demarcate both these clusters group one and group two. We can even demarcate and say that both the clusters are separate. For example as the example cited is in 2D space with just two axes x and y. I can draw a line in the space between the two groups to demarcate them and such a line is called as decision boundary. Let me write it down. Decision boundary. In the same example say I am going to add a third dimension, a third dimension which is elevation above mean C level. Same example which means assume that for the set of points that you already have you have one more set of information which is namely elevation above MSL mean C level. So let us try to plot the values in three dimensional space wherein I have on the x axis catchment area say on the other axis I have river discharge and elevation above mean C level. The same set of points I am using it to be represented using three axes. So what to do now? Again the eyes and the brain they tell me that there are two distinct set of clusters. I am using the word cluster, two distinct set of clusters or groups of points which are tightly knit together. In this case the decision boundary is going to be a plane instead of a line, is not it? Because we are talking about a three dimensional space and not a two dimensional space. All right? So while you hold this thought let us try to understand about how to mathematically represent this distance. Remember we are talking about the distance in space between two points using either two axes of the Cartesian coordinate system x axis and y axis or three axes of a Cartesian coordinate system that is x, y and z axis. So how can I mathematically represent this distance? Using the distance formula, is not it? We can estimate the degree of separation of two clusters or two groups of points by looking at the distance between their centers and the scatter of points around those centers. Here distance is used as a metric or measure of similarity or dissimilarity and the Euclidean distance between two points can be estimated using the following relationship, is not it? So say in this diagram in front of you, assume there are two distinct features, say this represent a part of a satellite image. One is, one feature is represented by blue color and another feature is represented by pink color. So once again the aim of image classification is to recognize or identify real world objects and by real world objects I mean water body, vegetation, buildings, built up area and so on. So in this case in this small example there are two features and the second part of image classification or the second step in image classification is I want to label the pixels to be classified so that at the end of classification the output is going to look something like this. Output is going to look something like this wherein a numerical label is given to each pixel. Say all the pixels that correspond to vegetation which are assigned V here are given a numerical label 1 and all the pixels that correspond to water body which is represented by W here are given a numerical label of 2. But then if you look closely in the diagram somewhere I have indicated question mark, is not it? Somewhere question mark, question mark, question mark. So these pixels have a combination of both the features, some part of vegetation and some part of water, is not it? Now this is where classification algorithm finds its importance. We will discuss about mixels when we start our introduction on fuzzy classification that is what to do when a pixel has a part of multiple features. Say for a pixel 30 percentage represents vegetation and say 70 percentage represents built up area or let me complicate the example a bit further. For one single pixel 30 percent represents vegetation, 30 percent represents water body, 30 percent represents built up area and 10 percent represents fallow land. So ultimately what label will you give such a pixel? The answer you will get once we discuss about fuzzy classification, but for now as part of this lecture I want you to understand the concept the geometrical basis of image classification. Remember we have not begun to understand image classification and synthetic aperture radar images. We are just trying to understand what image classification means and then towards the upcoming lectures we will try to understand how this can be exactly applied to synthetic aperture radar images. So just to re-itrate image classification it consists of two stages. Number one is to recognize or identify real world objects and number two is to label the pixels to be classified. So let me try to show you a glimpse of Thuvan portal which is an Indian web based utility prepared by the Indian space research organization Thuvan portal. You can access the thematic services present and the portal opens up. For your work you can download the land use land cover map from this portal. For example say I am going to click on LULC 50k 2015-16 and say I can click on any state I am going to go with say Kerala okay and then once I click on view I get a classified map wherein there are different land use land cover features represented in different colors. You know we have forest, we have built up area, I have agricultural lands, wetlands, water bodies and so on. The statistics are given for each of the land use land cover class. Metadata is present and the web service details as well as the overlay is present. For example say I want to highlight the transportation network say the national highway gets highlighted. Say I want to highlight the water bodies, reverse reservoirs they get highlighted in a different color. So you can see that if you are interested to know more about the thematic based products that are available in Bhuvan portal I would urge you to visit this website and to download the products of course you need to create a login ID password okay. Different products are possible for example what you see here is the urban land use product okay. Again with different legends I can browse the metadata, the details are given as well as I can go to overlay option and see the transportation network highlighted here as well as the water bodies. The administrative layers can be clicked on a very useful portal and I thought this is the right way to begin our module 3 on image classification because now you have an idea about what are the end results products of image classification and as part of the next lecture we shall be learning how exactly to perform image classification. So let me end today's lecture. I hope you understood the concepts of image classification and please do visit the utility prepared by ISRO and I will see you in the next class. Thank you.