mod07lec 31

[Music] [Music] [Applause] [Music] [Applause] [Music] hello everyone and welcome to horse and this one we are going to discuss about how multi spectral analysis especially involving several bands that means multivariate analysis can be done in digital image processing so in this way we will be going through the scatter plot two-dimensional plot and band ratio analysis principal component analysis and D correlation stretch so these are the topics for this first what is basically multivariate image statistics because we know that the remotes most of the remote sensing data is acquired in multi spectral form that means lot of bands data is coming and therefore we need to analyze that data so it is often concerned with the measurement of how much radiant flux is reflected or emitted in case of thermal infrared from an object in more than one band because this is the main thing is not a single band analysis where a stir gram is more than sufficient and stretching options are also simple but when we go for multi bands or multivariate analysis things becomes little complicated so it is useful to compute multivariate statistical measures such as covariance and correlation among several went to determine how the measurements co-vary this is required because if they are all say seven or eight bands in a once scanner our sensor and then we many of these bands might be highly correlated to each other and so that whenever we are going for some color combinations or composites we need to have those images which are different than each other and therefore the bands which are correlated to each other may be dropped in color composite so this kind of analysis will allow to understand and select the relevant bands it includes this multivariate analysis as I have mentioned earlier when the sewing techniques between discuss several limitations and advantages with Bendall show principal component analysis very popular one PCA feature selection classification and accuracy assessment so mainly image transformation and that is multivariate that the image transformation typically involves the manipulation of multiple bands of data whether from a single multispectral image or from two or more images of the same area acquired at different times that is multi to important data image data so that can also be analyzed in this way and either way the image transformations generate new images from two or more sources which highlight particular features or properties of interest better than the original input image in case of single band images most of the things which we have been doing through the hello T's and maybe in some cases new images may not be generated but in a multi variate analysis these things are there so first very simple multivariate analysis involving more than one band may be the image division or bender band division so the most common transforms applied to me is data that is image division on pixel by pixel basis carried out the following operations for example band 1 / band - so a new band is generated and resulting data then rescale to fill the range of display because if you divide the number of the range of pixel values will might reduce and therefore you want to rescale to the display range that is between 0 to 255 in most of the cases so this is very much required in most of these multivariate analysis this rescaling is the last sister which has to be done every time and this division is applicable commonly also called a band ratio and double very popular technique which we will see so mathematically basically if you look that V V IJ are equal to V V IJ K divided by V V IJ l and in this one we we I J K that is the brightness value at the location line I pixel L J and K band of imagery that means the bend and the bend one in this case k equal to one we can consider so bend one and each pixel basically this division will be made pixel by pixel so that is why locations of pixels are given in all and these two bands K and L and output band is the ratio image which is given denoted here as R so this L is the brightness value of the same location in Bendele the corresponding pixel will be divided this one has to remember and the ratio value at the same location that is the in in case of output image and in denominator is 0 if be end of then demand denominator baby is made 1 because there might be a situation when and this and the bend L that means may be bent to when we create the ratio the value becomes say in finite so we rescaled to 1 and what are the applications of ratio major one is to remove the shadow effects induce by the rugged topography and basically our ultimate aim in all these image processing techniques to improve the image interpretation so our images becomes much better for interpretation so this is what the undesirable effects on recorded radiances that is the variable illumination caused by the variations in the topography in daytime if you are if you are working in a hilly region then one once of the valley might be eliminated the other one will be in the shadow and this kind of effects and this differential solar and illuminations and can cause difficulties while making image interpretations so if we imply burn ratio techniques then these effects can be minimized that is the effects of topography due to the difference here and solar illuminations sometimes differences in giving that is a Bank values from identical surface materials are caused by topographic slope and aspect shadows or seasonal changes so all these the orientation of slope and the reference to the position of the Sun will play major role here and these conditions hamper the ability of interpreter as I have already mentioned to correctly identify surface materials or land use in a remote sensing and ultimately it will affect our interpretation in ratio transformations can be used to reduce the effect of such environmental conditions which are present in a hilly regions especially here is the example is given here that one side of the slope is illuminated on this left side whereas in the right side of the slope is in dark and therefore like in digital numbers of different features which are given here life or the the slope which is and when they might be having a pixel value of 48 in case of been V 50 but when they create a ratio it becomes 0.9 a 6 of course the pixel value has to be in ratio image has to be integer value so it is 3 scale later on to make a risk to integer and same like in case of shadow and Bend of an event a orb and when it is just 18 there as in bender B it is 19 so again is getting quite close as in case of sunlight and shadow and therefore the effect of shadow is getting reduced another example this is the and this was the from deciduous trees forests or vegetation then you are having examples of coniferous trees that here the in case of Sun dated slope it is been one this 31 pixel value in case of been to orb and B it is 45 so they show comes 0.69 and in case of shadow again 11 and 16 ratio comes 0.69 so when we will make the integers values then these values are coming very close whereas in Odin elements these values where at very far and this this is what the advantage we are removing the effects of shadow which is has been which has been caused due to the presence of rugged topography so you will be are having same cover type no problem and if you are having different cover types then different depending on which features are there or which land cover are there that and that will reduce the effects of topography radiance that is only 50 percent of radiance of sunlit as in many cases we are seeing here generally and the ratio nearly identical whereas the end result becomes identical and hence removal of topographic effects now applications of studies may be that radial issues may provide unique information not available in any single point because two bands information are being used and they issue a simple mathematical technique and ratio discriminate and subtly spectral variations and it contains aspect of shape and spectral reflectance curve of different earth surface covers can be brought out of by ratio ratios are independent of absolute pixel values and ratios can be used to generate false color composites by beining three monochromatic rishu datasets that means if you are having many many bends you can create ratios of two different bands having three outputs o3 ratio outputs these three the Shu outputs can be combined and can be created a color composite so I say multi bands ratios instead of involving just three wins in this world and this ratio color composites you would be involving six bands and less effects of topography and therefore the image interpretation may become very reliable so they show clearly portray the variations of slopes of spectral reflectance curves between two bands and volts here is the example here like in case of Landsat TM you know that they they are covering first in the visible part then when food is in the near-infrared part and then five is the middle infrared four and then then seven is you know in far basically far infrared now then the here band six is not mentioned because bent six is thermal so that has not been considered in this particular slide now different objects which are present on the earth their spectral curves would be different in different bands this is what you are seeing so if I create a ratio between say band 4 and band 3 then I am I am not only removing the topographic effects but I am involving two bands and the same time because here the vegetation is making major change here so in this spectral curve and therefore these bands and this kind of been racing will help us similarly wherever the major changes are there in case of vegetation in case of dry or wet soil or water bodies turbid water bodies and so on clear lake water bodies or behave differently and the ratios can exploit that one and you can have a better output image for reliable interpretations image interpretations for example which brands to ratio example this is what in the previous figure I was mentioning that the healthy vegetation reflects strongly in near-infrared because of high chlorophyll content and the spectrum by absorbing strongly in the visible red as you can see here that healthy vegetation and will have a high reflectance in near-infrared in case of Landsat team band 4 whereas in the red band of land said a team that is been three you don't have much reflection and this is this is what that the healthy vegetation reflects strongly in near-infrared portions of a spectrum while absorbing strongly in the visible red and because of this sharp curve we can exploit so other surfaces types such as soil and water so numerical reflectance in both near infrared and red portions it is not necessary that all objects or all kind of features which are present on the earth will behave different vegetation behaves differently water body dry soil website with the all are having different spectral curve so we need to understand the spectral cousin accordingly we should choose the bend and then create the Bend ratio thus the ratio wave of near-infrared and / red would result in ratio much greater than one for vegetation and issues around one for soil and water and this is that is why I was mentioning that where the curve has is a there is a change in the curve these two adjacent bands will give you much better results does the discussion of vegetation from other surface covers types is significantly enhanced in such ratio outputs and also we may be better able to identify areas of on stress vegetation vegetation suffering from and lack of water or maybe disease or something these by implying this ratio you can clearly identify and the veget health of the vegetation we show mania rainfall reflectance there is a shift then if healthy widgets and less chlorophyll content it will the curve will shift towards the red and the ratio would be lower than the healthy green vegetation so implying such a band ratio technique even in case of vegetated areas the health of vegetation can also be assessed easily and this is an example of a real example that this is infrared channel this is these are from IRS history and this is the red Channel and as you can see that there is the vegetated area which is coming dark in red channel Morrison and further definitely it will have a better limited part or a high brightness compared to red part but when we created this ratio then all the features which are present become much more clear second thing is you will observe that the effect of topography especially in this part which is the co Alex in this here they will the that has been reduced and you see this image which is more or less and quite representing a flat terrain but we know that it is not flat but for interpretation it might be better for certain applications so that is a big advantage that the discarnation power of the output image becomes much higher much reliable compared to individual bands if I would have been using so what's the advantage of binary swing then the been ratio of two beams as have been mentioned removes much of the effect of illumination in analysis of spectral differences and this is possible because the composition between two geologic units for example the limestone which is gray white and in red bed which is the red determine the ratio of any two bands whereas the illumination determination magnitude of the Dean received by the satellites and here is the example this is the Sun and the energy is coming it is getting reflected so if you are having this is the illuminated part so and this is that word band this is the limestone and they are they are doing different different reflection as you can see 46 23 23 and likewise and however ideas which are having is or atmospheric distortions or having this because of relief scattering and this these numbers from these rocks or reflection from these rocks is reduced significantly roughly half say inverse significant however in the areas with significant days the performance of ways you is much improved by removing the hedge produced by release scattering in that Mesilla and therefore bandwidth swinging when you find that there is a haze then we can employ again been dressing to improve our images so the product image the ratio image will help the less effects of scattering because that is causing and the haze so release scattering effects can also be minimized and this is the example here and that this is the band urban blue radiation so strong topography effects where the ratio band 3 y went to removes much of the effect of illumination and he'll differences of rock and here you can not identify different rocks very clearly and the effect of topography is very much dominating but here the effect of topography has reduced whereas the discrimination between two rocks or many rocks have increased so that's the advantage of veneration another example about the color composites so do you - vendor issues five by two and seven by four involving Landsat TM data and here that all these three are in RGB that is the color composite oven ratios so three well - which goes in red Channel five by two goes in green Channel seven by four goes in blue Channel and when we put we get a output image now we'll compare this output of it either with five by two single or seven by four which are involved here but you don't get that kind of discrimination power of different features which are present then in this color composite of band ratios so it is a very good tool an image processing to exploit the maximum variations which are presenting different bands and can include a new product here been to bank three then five invent a four and seven so two three four five and seven five bands and characteristics variations have been incorporated in this color composite in normally we can only incorporate about three in a standard false-color composite so that is why it is giving much better results and now little advanced vendors showing techniques and that we end up just no joy by dividing but making more different combinations or bringing little more arithmetic in it and that is very popular one is called normalized difference vegetation index or NDVI and it is for using coarse relatively coarse resolution data like no IV HRR or modest data Google at global scale and DVI maps are being used so we can assess and that what is happening to the vegetation cover globally and in this one the more infrared band is subtracted by red band and this is / near-infrared + red band and that we would mdvi will do it will bring and highlight the health of vegetation so we are improving LED show stress vegetation and V NIR that is very infrared band saws chlorophyll and that is why these two bands are involved in this normalized difference vegetation index here is a example like this is the healthy vegetation on right side you are seeing the vegetation which is not having much chlorophyll so here in infrared the reflection is coming about 50 percent and invisible just 8% but when we go for the dry leaves or vegetation which is suffering from stress levy lack of water or some disease then when neuron further deflection has reduced here by roughly 20 percent and whereas visible reflection has increased very significantly from 8 to 30 and if we see the ratios then say this is a healthy vegetation ratio is 0.72 whereas in case of a not the healthy vegetation the vegetation under the stress is just 1 4 so that's the advantage that you can assess in a vegetated area the vegetation which is suffering from stress and they will that this collision between healthy and a vegetation suffering from is class will be enhanced as we can see here and this is you have the examples of at continental scale or country scale and you can Northern Ireland and some part of Europe is here so this is the NDVI of June 2003 this is the NDVI of October 2003 you can see that how this in DUI has encountered decrease so in October because now this is the onset of winter and you are having very less NDVI value but in case of June which is you are having full of grimly and NDVI is touching even up to one so that's the you can assess that how vegetation cover at if we because generally we imply course resolution data so at country scale or continental scale and this vegetation conditions can be assessed now in further to this there is a further analysis which is called PCA or principal component analysis a where we use the Scott a scatterplot two-dimensional histogram we have already discussed when we discussed storuman scatter plots so we will just directly go for this and I will explain in this graphics that IV plot bind1 and band to like this now if they are correlated like this this is this see that they are very much correlated so then then then this origin is shifted in the center and this rotation is there so this becomes so the religion instead of here is shifted here and they the principal component one this is aligned with the maximum variations which are present between these two MA tube ends and the next one is the pc2 which is perpendicular to PC one and now I think that if you are having a multiband data or multispectral data so you can graphically I can represent mainly three components pc-1 pc2 pc 3 but in with using a digital image processing software we can we can go for many income and not only 3 but maybe 4 5 6 depending what you are trying to achieve so here only I am showing two so because of bend correlation which you are seeing where that one season Bend one is not so much different from what we see on the bend to and as per linear algebra we are making eigen vectors where the I News contain the contribution of each man to the principal component and this is what we see here that wherever the six principal components of banned ones 5 1 2007 of Landsat TM 6 is missing because 6 is thermal channel so that is not involved in the analysis and this is what you are seeing that principal component 1 will always have the maximum variations and the principal component which is lost in this case is 6 which will have maximum noise and as you can see here so all noise which is having the not much relation with any bangs will go in principle component 6 in this particular example and maximum variations which are present in the image in these images of 6 channels 6 bands there all has been and accumulated basically if I can use this word with principal component 1 so principal component the component 1 image will have maximum variations the last one will have maximum noise nothing else and now one more step is also possible because now I will move that this band is and there are other other principal components are highly correlated they are not correlated so I can also make principal compressive so principal components so I can use principal component 1 2 3 and assign say red green blue this thing then I can create color composite of principal component and the such output images not only within 5 or 6 bands in this case in this example 6 bands but when we create a color composite they they are creating much more better images for of course for reliable interpretations this is how the generally statistics which you will get that band 1 here band to band 3 and and then basically here in this as soon has been 6 but then seven basically and this is correlation matrix and this is correlation eigen vectors are there so then when one then one definitely DISA and that percent correlated so one is there and similarly you can see that how they are correlated then one tube and two is point nine so it's a very high correlation and therefore principal component analysis is would be much useful in such cases now the next step in principal component analysis is D correlation stretch which is a given based on the principal component analysis like example is shown here that Bend one has been to scatter plot is showing a highly correlated data and what we can do we can shift the reason as in case of principal component analysis and then stretch along these axis to occupy the full range which is available for display and why this we are depleting each component within each component and at the end in a in F if we see this scatter plot we are occupying a very large area which was not possible and with the normal thing so this say bang bang bang two is a coordinated highly coordinated is there whereas Bank for band seven are less correlated so this is this is the situation and if we plot and see this in queue here this is the original RGB this is how the pixel values are distributed then we'll go for the correlation stretch Decius in RGB color space they are occupying a very large area or space between that between that color cube which we have been discussing also and this is another color combination where if I grow them I am we are occupying much more area which is available within and hsy that a hue/saturation and yellow color space then we are occupying little s so this D correlation stretch and in our java combinations and many times will give you much better results with full of colored images an example here is that this is simple RGB image but given the principal component and then we did the D correlation of different prints components principal controller pieces then and then made a color composite this is what the result is and see that the quality of discrimination here I can discriminate different rocks different gun features comparatively very easily then on simple RGB image this is the advantage of going for multivariate analysis in digital image processing and this brings to the end of this lecture thank you very much [Music] [Applause] [Music] [Applause] [Music] [Music]