ArcGIS Pro: Machine Learning Classification for Impervious Surfaces

[Music] in this demo I'll show you how to use machine learning classification to extract features from multispectral imagery and then we'll use those classification outputs in additional analyses for example today we'll identify impervious surfaces like roads roofs and sidewalks many local government institutions use impervious surfaces to calculate the stormwater bill for a property we'll be using an object-oriented feature extraction method and the ArcGIS Pro classification wizard to accomplish this here we have a multispectral image just outside of Louisville Kentucky it was provided by the Louisville Jefferson County information consortium this type of image is ideal for performing detailed feature extraction of impervious surfaces it's a six inch resolution aerial ortho photograph and it contains a near infrared band we can change the band combination of the image to highlight certain features like vegetation roads here a color infrared band combination makes it very easy to pick out vegetated areas in this neighborhood this color infrared band combination looks like it will work well for what we're trying to identify since the man-made features look very different from the vegetation now that we have the correct band combination we can begin the classification process to do this we'll use the classification wizard in ArcGIS Pro the classification wizard guides you through a series of steps to complete your analysis very quickly and easily the first step of the classification wizard is configuring your classification process here you can choose between supervised and unsupervised methods and also between pixel or object based classification types for this example we will complete a supervised object based classification object based classification uses segmented images rather than classifying the individual pixels segmentation is the process where pixels that are close together and have similar spectral characteristics are grouped into a segments the segments that exhibit certain shapes and spectral and spatial characteristics can then be further grouped into objects during this classification process we'll be classifying the segments not the pixels for this project we've created our own classification scheme in other projects if you don't already have one you can create one using pre-existing training samples generate a schema from a classified raster or you can use the default schema we'll load our classification schema along with some previously collected training samples that we can then add to and then we'll click Next to move on to the next step of the classification wizard in this step we will create the segmented image on which will actually run the classification for this step you can tweak the segmentation parameters to fit your desired output if you don't know what settings you want to use you can experiment with the segment mean shift raster function before running the classification wizard here we already know what parameters we want to use so we can set them for spectral detail a higher value equates to more segments so for our scene a spectral detail of 12 will capture roofs into a few classes and merge subtle changes where shadows exist from changes in the roof pitch for the spatial detail smaller values will create spatially smoother outputs for our scene a lower value will help distinguish features such as a roof from a driveway or a driveway from a road segment for the minimum segment size in pixels you would set this value higher for a higher resolution data like images with one foot or six inches per pixel for our scene which is at a resolution of six inches per pixel a value of ten is going to merge changes from features like chimneys on the roofs into one segment we will also leave the show segment boundaries only box unchecked this will create objects if you check this box it will actually display the segment's as contours once you've set your segmentation parameters you can click Next to generate your segments and image and move on to the next step of the classification wizard now that we have our segments to image we can create training samples over features in the image will collect impervious features like roofs driveways and roads as well as pervious features like trees grass and exposed dirt we'll use a training sample manager to create and save these training samples the training samples manager pane of the image classification wizard will open with the classification schema and the training samples that you selected in the first step so we're already ready to create more training sample and let's start with some shingle roofs we'll be using the segment picker tool to select whole segments for our training data rather than drawing polygons manually and we'll be using our new segment to image as our selection layer when you use the segment picker tool one click will select the shape for best results we recommend that you zoom in closely when you use this tool because the segments are drawn at the resolution they are displayed on the screen here we've selected a roof segment and at this resolution it appears that the outline matches the segment boundary but if we zoom in to a higher resolution we notice that the outline doesn't match the outline of the segment very well but now azumed in if we click this roof segment again we now have two different lines outlining the roof and if you look closely you'll see the second selection highlighted here in blue better matches the outline of the segment we can delete the first sample that we collected in our training sample manager and samples are added to this list in the order that you create them so the first of these two shingle roofs listed at the bottom is the one that doesn't match the outline correctly and needs to be deleted every shape that you draw goes into this list as a separate class but you can merge the classes together in the training samples manager pane for today we've created a complete training sample set so we don't need to continue to collect more training samples now that we've created a segmented image and we have some training samples we have the required inputs to move to the next step into classifying our image the next step in classifying your image is training your classifier for this classification we're going to use a support vector machine or SVM classifier to analyze our data and help recognize patterns in our data set but the image classification wizard can also do a maximum likelihood classification or a random trees classification for the maximum number of samples per class we didn't create more than 500 training samples so we will just use the default of 500 this will use all the training samples when we're working with a segmented raster data set we can also select any variables that we want to include in our training in this example we will use all of the characteristics listed here once we have our input set we can train the model once the model has been trained then you can run the classifier you now that we have our classified data let's quickly compare the classified output with our original color infrared image the next step in the classification wizard gives you the chance to group classes into their parent classes now that we have our classified image we can reclassify it into only pervious and impervious surfaces since what the town really wants to know here is what percentage of the property is covered with impervious surfaces you the final step of the image classification wizard is the reclassify ur tool which is where you have a chance to go in and make any edits to your classified data set so over here in the bottom right-hand corner of our image we have a spot where there is a median full of grass that was not classified as a pervious surface you can see that here when we do the swipe tool we'll make it a little transparent so we can see where we're working and we want to reclassify this to pervious so we'll select pervious as our new class and select the reclassify within a region tool and then we just draw an outline of the area that we want to reclassify double click to finish our polygon in that area will now be reclassified as pervious once we're done with any reclassifications you can click run and you finished your image classification wizard so here we can look at our final classification of pervious and impervious surfaces compared to our original color infrared image so pervious features are shown in green and impervious features are shown in black we can use this classification output to analyze property parcels and calculate the percent of each parcel that is covered in impervious surfaces to do this we will again use the reclassify tool and now we will create a new layer showing only the impervious surfaces we will set all the impervious features to a value of one and the pervious features to no data and this will create a new layer that only contains impervious features let's change the symbology here so we can view the impervious layer over the imagery a little better so now all of our impervious features are highlighted in red to determine the area of impervious surfaces in each parcel we'll need to add the parcel dado that was provided by the county to our map you we'll need to change the symbology on this layer so we can see our data better so let's change it to an outline only you now that we have the parcel data in our map we can use the tabulate area geoprocessing tool to calculate the square footage of impervious surfaces in each parcel the tabularium new table that gets added to the map and this table contains the square footage of impervious surfaces in each parcel so if we open the new table we can see that each parcel ID is now associated with an impervious value in square feet the impervious value is calculated in square feet because the parcel airs coordinate system is in feet this table could be used by itself to determine the storm water bill for each parcel but we can also join this table to the parcel feature class to provide more display options to make it easier to compare the amount of impervious surfaces from one parcel to another we'll calculate the percentage of each parcel that includes impervious surfaces so we'll add a new percent impervious field to the parcel attribute table you and now if we look at our partial attribute table we can see the new percent impervious field is included now that we've created this new field we can calculate the percentage of each parcel that's covered with impervious surfaces we'll do this using the calculate field tool you and you can see here our percent impervious field has updated to include these newly calculated values parcels that have no impervious surfaces will still have a null value like this one you can see here to get a better view of the amount of impervious surfaces in each parcel we can now change the parcel layer symbology to reflect that percent impervious field which you just calculated you and here we have our final map which shows the percent of each parcel that contains impervious surfaces parcel is highlighted in red contain a large amount of impervious surfaces and parcels highlighted in green contain a very small amount of impervious surfaces all of this was accomplished with only one 4-band image and a parcel map the ArcGIS Pro classification wizard makes image classification simple so that you can make the most of your data you

Channel: ArcGIS

Views: 9,482

Rating: 4.9783783 out of 5

Keywords: Esri, ArcGIS, GIS, Geographic Information System, ArcGIS Pro, impervious surfaces, machine learning, feature extraction, classification, multispectral imagery, object-oriented feature extraction, imagery, ArcGIS Image Analyst extension

Id: iy3HBkaXYbw

Length: 14min 25sec (865 seconds)

Published: Mon May 18 2020