 This is the lesson six lecture and this has a focus on the idea of change detection using remote sensing. Therefore, the focus is on digital change detection because most, if not all, contemporary remotely sensed imagery is digital in nature. So biophysical materials and human-made features, anthropogenic features on the surface of the earth are inventoried using remote sensing and field techniques. Some of the data are fairly static. They do not change over time like landforms, mountains, rivers and so forth, typically remain just about the same. Conversely, some biophysical materials and human-made features are dynamic and changing rapidly like our cities are growing, our agricultural fields change due to human activity every year and so forth and it is important that such changes be inventoried accurately so that the physical and human processes at work can be more fully understood. So therefore, if you have a imagery at two times acquired through remotely sensed methods, it has a story of a change detection, of a change that took place and what might be the human and the social and the economic processes that are driving this change can be inferred to an extent. It is believed that land use and land cover change is a major component of global change with an impact, perhaps greater than that of climate change and therefore it is not surprising therefore that significant effort has gone into the development of change detection methods using remotely sensed data. This lecture will review how change information is extracted from digital remotely sensed data. It first summarizes the remote sensor system and environmental parameters that must be considered when change detection takes place. Several change detection considerations and methods will be introduced as well. This is a generalized workflow for change detection using remotely sensed imagery and it begins with stating the nature of the change detection problem. What is it that you are looking for? And you begin by specifying the change detection geographic region of interest. The change detection time period is a daily, seasonal, yearly. What's the temporal cycles needed for this change detection? Define the classes of interest in a classification system. Then select either hard or fuzzy change detection logic. We will talk about hard and fuzzy classification briefly at a later time and then select per pixel or object based change detection techniques. Then there are other considerations of significance when performing change detection. The remote sensing system considerations that is knowledge of the sensor and how rapidly does it take the imagery and what kind of resolutions it has. There are environmental considerations as well. What were the atmospheric conditions during the time of data acquisition during both times? Soil moisture conditions, phenological cycle characteristics, the tidal stages, should it be close to the ocean? Then the remote sensing data has to be processed to extract the information. So it would be good to have in situ and collateral data, some ground truthing data to improve your classification. You need to know when your remotely sensed data was collected and what is the repeat time for the collection. Then the pre-processing has to happen where you have to do the geometric correction and the radiometric correction. If it has not been done, you have to select the change detection algorithm, do the appropriate classification, made be supervised, unsupervised, and so on and so forth, perform change detection using GIS algorithms, and then compute the statistics, quantify that change. With a change detection matrix, we will talk about this later. After that, you perform the accuracy assessment. You compare the land cover map with either an image that you have of this region, that is of a higher spatial resolution that you can use for ground truthing, or having access to field data for this accuracy assessment. We will focus in on accuracy assessment in greater detail in lesson eight. And then, once you've done the accuracy assessment, if your product meets the standards that you need for an effective change detection that is good, otherwise, you go back and you repeat parts of this process again iteratively till you have a product that works for you and gives you the ability to answer the question that you were trying to probe. So the dimensions of the change detection region of interest, the ROI, must be carefully identified and held constant throughout a change detection project. So this geographic region of interest, made to be a county, state, or a watershed, is especially important in change detection studies because it must be completely covered by the dates of imagery available, the end dates of imagery available, end could be one, two, three, four, however many repeat dates. Failure to ensure that each of the multiple date images covers the geographic area of interest results in a change detection and change detection maps with data voids that are problematic when computing change statistics and trying to quantify the change. The change detection time period is very important as well. There are certain problems that need very rapid time cycles for the data to be converted into information, as we have seen with several firefighting posts in previous discussions, that the data needs to be changed into information very rapidly, and that is true for most disaster response applications. Sometimes the time period selected over which the change is to be monitored is too short or too long to capture the information of interest. So the analyst must be very careful to identify the optimal change detection time periods. So it will be different for agriculture versus for fighting a fire, for example, or a disaster response. If there is any flooding or a hurricane or a tornado and so forth, then the time cycles for change detection are much quicker and the delta T's are much smaller. This selection is dictated by the nature of the problem. Traffic transportation studies might require a change detection period of just a few seconds or minutes, and images obtained monthly or seasonally might be sufficient to monitor the greening of a continent. Careful selection of the change detection time period can ensure that resource analysis funds are not wasted. It is very important to select an appropriate land cover, land use classification scheme when you are doing change detection, such that your before and after images are codified within the same land use classification scheme. So it is wise to use an established standardized land cover use classification system for change detection, such as the following. So the American Planning Association has its own land based classification standard. The USGS has the Anderson system, the land use and land cover classification system for use with remote sensor data. The US National Vegetation Classification System can also be used. The US Fish and Wildlife Service has a classification of wetlands and deep water habitats of the United States and the International Geosphere Biosphere Program also has its own land cover classification system and use of standardized classification systems allows change information to be compared with other studies. In this course, we have focused mainly on hard classification. Hard land cover classification means that you have assigned the objects or the pixels to a particular land cover class, made to be forest, pasture, agriculture, whatever the case might be. But sometimes you can have a fuzzy classification as well that this pixel or this object could be this class or it could be that other class as well and that's known as fuzzy classification. So most change detection studies have been based on the comparison of multiple date hard land cover classifications of remotely sensed data. The result is the creation of a hard change detection map consisting of information about the change in discrete categories. For example, change in forest class or change in agricultural class or changes in built up urban classes from one class to another, one specific class to another and this is still very important and practical in many instances but we now recognize that it is ideal to capture both discrete and fuzzy changes in the landscape where possible. You can also have change detection based on a per pixel basis or in per object basis using the object oriented change detection. Most digital image change detection is based on processing a date in some previous date and then a date in plus one some subsequent date classification maps pixel by pixel. Okay, but this is changing with the onset of obi a where we are seeing a lot of studies in which the change is done object by object. This is so a pixel based change is commonly referred to as per pixel change detection. Object oriented change detections involves the comparison of two or more scenes consisting of many relatively homogeneous image objects which are patches or segments same thing and the smaller number of relatively homogeneous image objects in the two scenes are then subjected to various change detection techniques. There are remote sensing system considerations to be taken when you are doing remotely sensed change detection. Successful remote sensing change detection requires careful attention to remote sensor system considerations and environmental characteristics in the study area. Failure to understand the impact of the various parameters on change detection process can lead to inaccurate results. Ideally the remotely sensed data used to perform change detection is acquired by remote sensor system that holds the following resolutions constant temporal spatial that includes the look angle spectral and radiometric resolutions and we will discuss these in greater detail in subsequent slides. The remote sensing system parameter temporal resolution is very important for change detection studies. Two temporal resolutions should be held constant during the change detection if possible. First use a sensor system that acquires data at approximately the same time of the day. For example Landsat thematic data thematic mapper data TM data are acquired before 9 45 a.m. for most of the coterminous United States. This eliminates the diurnal sun angle effects that can cause anomalous differences in the reflectance properties of the remote sensor data. What that means is that we want to try to keep the illumination conditions to be about the same in making a comparison and second acquire remote sensor data on anniversary dates for example February 1st 2005 and February 1st 2007 anniversary date imagery minimizes the influence of seasonal sun angle and plant phenological differences that can negatively impact a change detection project. Another critical remote sensing system parameter is spatial resolution involved in the change detection study. So accurate spatial registration of at least two images is essential for digital change detection. Ideally the remotely sense data are acquired by a sensor system that collects data with the same instantaneous field of view on each date meaning to say the same pixel size on each date. For example Landsat thematic mapper data collected at a 30 by 30 meter spatial resolution on two dates are relatively easy to register to one another. What that means is if you pull it up on a geographic information system if you pull up both of these images they will be registered pretty exactly the same where you have the same pixel on the ground in direct conflation on in the two images such that they are lined up with each other as closely as possible. It is possible to perform change detection using data collected from two different sensor systems with different instantaneous field of views or different ground pixels. So for example the Landsat thematic mapper data has a ground pixel of 30 meters by 30 meters and let's say we collected that for date one the initial date and then for the subsequent date date two we have spot HRV high resolution visible XS data which is 20 meters by 20 meters on the ground. In such cases it is usually necessary to decide on a representative minimum mapping unit. So let's say you could decide 20 meters by 20 meters or you could choose 30 meters by 30 meters depending on that particular problem and the issues involved and then you would have to resample both the data sets to this uniform pixel size whichever minimum mapping unit that you decided on. This does not present a significant problem as long as the image analyst remembers that the information content of the resampled data can never be greater than the instantaneous field of view or the pixel of the original sensor system. What that basically means is that if you start resampling an image there is a little bit of information loss that is somewhat inevitable and the analyst needs to bear this in mind. All the remotely sensed data used for change detection should be geometrically rectified to be within plus minus half a pixel of its correct planometric position. This is very important for all of the different data layers the GIS data layers and the image retaking at different times to line up exactly or as close as possible within half a pixel in a geographic information system such that you can conduct meaningful change detection analysis. To further elaborate this point the geometric rectification algorithms are used to register the images to a standard map projection and let's say UTM for most U.S. projects and rectification should result in the two images having a root mean square error or RMSC of less than or equal to half a pixel and misregistration of the two images may result in the identification of spurious areas of change between the data sets. For example, just one pixel misregistration may cause a stable road on the two dates to show up as a new road in the change image and this is very important and important consideration and when you download publicly available data very often you will find that it has not been properly rectified and you get two image layers from two different times and they will not be within half a pixel of each other in many instances not always so therefore it is very important for an analyst to check how closely the two images are registered to each other in the case of a change detection study and if they are not registered very closely to each other then the analyst should do the process of rectification to ensure that the images are lined up to within half a pixel of each other before proceeding. For many satellite systems and aerial systems the look angle gets to be an important variable that has to be accounted for in change detection studies so remote sensing systems like the spot satellite which is a french satellite and the quick bird which is a commercial satellite can collect data at off nadir look angles as much as plus minus 20 degrees from an oblique vantage point so a nadir collection would be when you are looking straight down onto the study area whereas spot and quick bird can take oblique imagery as well which in essence increases the revisit time for these satellites if they are able to look off to the side and two images with significantly different look angles can cause problems when used for change detection for example a spot image of a maple forest acquired at zero degrees off nadir will look directly down on the top of the canopy conversely a spot image acquired at 20 degrees off nadir will off vertical will record reflectance from the side of the canopy differences in reflectance from these two data sets may cause spurious change detection results therefore the data used in remote sensing digital change detection should be acquired with approximately the same look angle if possible the spectral resolution of the sensor being used for the before and after imagery is a very important consideration as well ideally the same sensor is used to acquire imagery on multiple dates when this is not possible the analysts should select bands that approximate each other for example the landsat multispectral scanner bands four five and seven and spot bands one two and three can be successfully used with landsat etm plus bands two three and four and many change detection algorithms do not function well when the bands in one image do not match those of the other image for example utilizing the landsat thematic mapper band one blue with either spot or landsat multispectral data may not be wise because these bands do not correspond well to each other here is a further elaboration on where there may be differences in radiometric resolution between the before and the after imagery in a change detection study so a analog to digital an a to d conversion of the remote sensor data usually results in eight bit brightness values ranging from zero to 255 as is the case in landsat imagery up to landsat seven landsat eight gets to be a 16 bit sensor and then most nape imagery is of eight bit resolution rate bit radiometric resolution such that it has values ranging from zero to 255 so ideally the sensor systems collecting the data should be at the same radiometric precision or resolution on both dates when the radiometric resolution of data acquired by one system for example landsat multispectral scanner one which has six bit data is compared with data acquired by a higher radiometric resolution instrument for example landsat thematic mapper which was on landsat four and five where which acquired eight bit data the lower resolution data in six bits should be decompressed to eight bits for change detection purposes and image processing software can do these types of conversions however ideally the brightness values associated with both dates of imagery are converted to an apparent surface reflection which eliminates this problem of the differences in radiometric resolution sometimes there is the issue of radiometric correction of multiple dates of imagery to perform change detection so let's say if you have the top of atmosphere reflectance then you would have to do the absolute radiometric correction for each image so do the atmospheric corrections and so forth such that you can recover the surface reflectance for each image and then you can do the processing to see what kind of changes occurred over time another option is a relative radiometric correction that you just take the first image as a baseline and what you do is you take the multiple date images and you normalize the image histograms relative to this first baseline image such that you mimic similar lighting conditions before you proceed with your change detection exercise understanding the remote sensing environmental conditions of the study area is of great importance when conducting change detection so therefore this is where domain knowledge comes into play and the failure to understand the impact of various environmental characteristics on the remote sensing change detection process can lead to inaccurate results and when performing change detection it is desirable to hold as many environmental variables constant as possible one example of such an environmental parameter is soil moisture so ideally soil moisture conditions should be identical for the successive for the end dates however one two three four dates of imagery used in a change detection project extremely wet or dry conditions on one of the dates can cause change detection problems it is important to review precipitation records to determine how much rain or snow fell in the days and weeks prior to the remote sensing collection when soil moisture differences between the dates are significant for only certain parts of the study area perhaps due to a local thunderstorm it may be necessary to stratify or to cut out those affected areas and perform a separate change detection analysis which can be added back in the final stages of the project another very important environmental parameter is the phenological cycle characteristics natural ecosystems go through repeatable predictable cycles of development humans also modify the landscape in predictable stages these cycles of predictable development are referred to as phenological cycles analysts use this information to identify when remotely sense data should be collected therefore analysts must be familiar with the biophysical characteristics of the vegetation soils and water constituents of ecosystems and their phenological cycles likewise they must understand human made development phenological cycles such as those associated with residential expansion at the urban rural fringe let's take a focus on the notion of vegetation phenology vegetation grows according to relatively predictable diurnal seasonal and annual phenological cycles obtaining near anniversary images greatly minimizes the effects of seasonal phenological differences that may cause spurious change to be detected in imagery when attempting to identify change in agricultural crops the analyst must be aware of when the crops were planted ideally monoculture crops for example corn wheat are planted at approximately the same time of the year on the two dates of the imagery a month lag date between fields of the same crop can cause serious change detection error here is a graphic that shows the example of vegetation phenological cycles along with landslide multispectral scanner images of one field during a during a growing season at the San Joaquin and Imperial Valley in California so this picture somehow speaks for itself and you can see the different stages for sugar beets cotton and alfalfa and it's got the corresponding multispectral scanner imagery along with it some of it is in as you can see in the color infrared false color combination and some in real color and also bear in mind that the multispectral scanner has an 80 meter pixel pretty large pixel and therefore this must be looking at pretty large fields of these crops furthermore for vegetation phenology and particularly for agriculture changes in row spacing and direction can have an impact on the change detection study these observations suggest that the analyst must know the crops biophysical characteristics as well as well as the cultural land tenure practices in the study area so that the most appropriate remotely sensed data can be selected for change detection so here's a dichotomous key used to identify progressive stages of residential development pointing at the urban suburban phenological cycles and the entire process of how this urban suburban interface is going to change is mapped out like a flow chart please take a moment to look at this process and knowledge of these types of processes is what a domain expert brings to change detection in remote sensing so somebody who has experience and knowledge of the phenomena that is being examined is critical to have a successful remotely sensed data based detection study a change detection study so here's just a visual example of how you have a residential development near denver colorado and how the land cover is changing between two dates in 1976 and then later on in 1978 and this process is now automated in geographic information systems was done more so by hand at that time but you get the basic idea the same kind of mapping and change detection is now done with digital technology the selection of a change detection algorithm is an important consideration as well in this course we will not focus on the different types of change detection algorithms you just need to be aware of the general types of change detection algorithms that exist and this is a topic for yet another course in remote sensing to dig into the details of these algorithms and so the selection of an appropriate change detection algorithm is very important and firstly it will have a direct impact on the type of image classification to be performed if any secondly the selection of the change detection algorithm will dictate whether important from to before and after change information can be extracted from the imagery and many change detection projects require that from to information be readily available in the form of maps and tabular summaries here is a listing of a few of the change detection algorithms that are used for change detection studies and these include multi-date composite image methods image algebra methods which is band differencing and band ratioing post classification comparison binary mask applied to date two ancillary data source used as date one spectral change vector analysis chi-square transformation cross correlation visual on-screen digitization and other knowledge-based vision systems so here is one possible way where you can visualize changes that have occurred in a scene so here we have two black and white images so the first one is taken from the nap program the national aerial photography program it's a black and white image with a two and a half by two and a half meter pixel taken in 1994 and then we have another air photo taken from the cams instrument which is also in roughly the same spectral region again it's a panchromatic black and white image also with the same spatial resolution and if you look at both of these images in one display that is you put the the cams image the image taken at a later time to the red gun and you take the image taken from the before time to the green gun and you don't have any color gun assign assignment for the blue gun and if you look at that image display you can immediately tell the red features are pertaining to the image taken in 1996 and the greenest features are the images taken in 1994 and you can visually begin to discern the changes that have taken place because they are showing up in red from the later image so then there are image algebra change detection methods as well it is possible to identify the amount of change between two rectified images by band ray showing or image differencing image differencing involves subtracting the imagery of one date from that of another so if the two images have almost identical radiometric characteristics that is the data have been normalized or atmospherically corrected in a similar manner and the subtraction results in positive and negative values in areas of radiance change and zero values in areas of no change the results are then stored in a new change image when eight bit data are analyzed in this manner the potential range of difference values can be found in the range of negative 255 to positive 255 and the results can be transformed into positive values by adding an appropriate constant the constant will never be greater than 255 but it just depends on the particular case that what constant you would choose to make sure that all the numbers in the change image get to be positive here is an example of image differencing for change detection so we have a spot panchromatic image over here on par pond in south carolina taken in 89 in april 89 and another one taken in october 89 and by taking the difference of these two images we can begin to see and map the water lily growth that took place in this time period by looking once again at the difference image another method to quantify the changes that have taken place over a period of time using remotely sensed imagery is the post classification comparison change detection post classification comparison change detection is a heavily used quantitative change detection method it requires the rectification and classification of each remotely sensed image the two maps are then compared on a pixel by pixel basis using a change detection matrix unfortunately every error in the individual date classification map will also be present in the final change detection map therefore it is imperative that the individual classification maps used in the post classification change detection method be as accurate as possible here is a graphic that shows the post classification comparison in a change detection study so we have images taken at two dates let's say you have two multi-spectral images at date one and date two both of them have undergone similar pre-processing and then both of them are distilled into a land cover map for date one and a land cover map for date two and from these two land cover maps you can come up with a change map and a change detection matrix as we will see in the subsequent slides so here are two case studies for post classification comparison and change detection so one is for kittridge south carolina between 1982 and 1988 and the other is for fort moltree also in south carolina for 1982 and 1988 and here are the two landsat images that were used in this study in the next step we can see that both the before and after images for both of those these two study areas have been developed into a classified land cover map as shown by getting a difference image between the classified maps we can identify areas of change in both of these two study areas as you can see from this graphic so by looking at the two land cover maps the before and after land cover maps for each one of the study areas you can construct a change detection matrix by which you can quantify the land cover change that has occurred so you can see that in the in the columns in the vertical direction we have the before image 1982 image land cover characteristics listed over here and then in the rows or in the horizontal direction we have the after image 1988 and the land cover class is developed from the the after imagery and then we can see that in the diagonal elements is where there is no change that where you had developed land in 1982 and you still have developed land in 1982 but you can also begin to quantify the changes where you had a particular class in 1982 and we can see that that class has changed into other classes in 1988 so please take a moment to look at this change detection matrix such that it gives you an idea on how land cover change is quantified using this method here we have a simplification of a change matrix to look at the post classification comparison change in a change detection study so you can see that the initial state is the year 2000 and the final state is the year 2007 the diagonal elements represent no change what was initially water is still water along the diagonal elements but the off diagonal elements show you the change that happened from one class to another class and please take a look at this matrix to try to understand the logic of it sometimes for change detection studies using an ancillary data source as date one ends up being very convenient and useful and effective so sometimes there's a land cover data source that may be used in place of a traditional remote sensing image in the change detection process so for example the us fish and wildlife service conducted a national wetland inventory nwi of the united states at a one is to 24 000 scale some of these data have been digitized and in fact most of these have been digitized at this date and are available on the internet so instead of using a remotely sensed image as date one in a coastal change detection project it is possible to substitute a digital national wetland inventory map of the region as the initial state in this case the nwi map is recoded to be compatible with the classification scheme being used and next date two of the analysis is classified and then compared on a pixel by pixel basis with date one information using post classification comparison methods the traditional from to information then can be derived in a change matrix so here's a graphic that shows change detection using an ancillary data source as date one it could be the national wetlands inventory map but it could also be the nlcd the national land cover uh database or could be some other remote sensing product that is standardized in a particular community and that can very well be used for date one such that for date two you can have your remotely sensed imagery from which you derive a classification map and then you can compare the before and after effectively there's another advanced method to look at changes in a study area using remotely sensed imagery and this is known as the spectral change vector analysis so when a land undergoes change or disturbance between two states its spectral appearance normally changes so for example consider the red and near infrared spectral characteristics of a single pixel displayed in a two-dimensional feature space so what i have done is if you look at this diagram in black up on the top left i have created this in e-cognition in which i have mapped uh the red image the red image along the horizontal and the near infrared image along the vertical and this has plotted out the points according to the different classes in the image so red is urban and i hope you can see that we can put a straight line or a vector that can best approximate the red distribution which is these the distribution of these points is representing the urban class now if i take this image at a later time and let's say the urbanity has increased or decreased then this net total average slope of all of these red pixels is going to change and that allows for a very effective method of change detection that is known as spectral change vector analysis so so it appears that the land cover associated with this particular pixel has changed from date one to date two because a pixel resides at a substantially different location in feature space on date two feature space simply means plotting a graph between one layer of a multi spectral raster versus another and the vector describing the direction and magnitude from change from date one and date two is known as a spectral change vector so once again i hope you can visualize this for the red points up here on this graph that represents urbanity but then you you also have the dark green that represents forests and you have nearly almost a vertical line through it then you have the light green points that represent pasture and you could put a line through it that has a negative slope and right in the corner and near the origin you can see the water pixels and you could also put a straight line through it with some kind of an average slope that best fits those points as well so i hope looking at this graphic over here that i have constructed in e-cognition helps you visualize the basic concept of the idea of spectral change vector analysis there is also another advanced change detection algorithm that is known as the cross correlation change detection and this makes use of an existing date one digital land cover map and a date two unclassified multi spectral data set the date two multi spectral data set does not need to be atmospherically corrected or converted to percent reflectance and several passes have to be made through the data sets to implement the cross correlation change detection and i just want you to be aware of this possibility we will not be getting into the details of this course and this should give you an idea of some fundamental techniques that are used for looking at change detection using remotely sensed data if you have any questions or comments please post them in the general question and comments discussion for this lesson thank you