Welcome to a continuation of last weeks look at Land Use / Land Cover.
This week I am specifically looking at the classification accuracy of last weeks generated Land Use Land Cover Example map of Pascagoula Mississippi. The overall objectives of this week were to explore the different types of accuracy and how to compare them with the data presented. This builds on last weeks map and is based on taking samples at locations of each classification type and determining if the sample is accurately portrayed by the assigned classification. these samples are typically generated in situ meaning at the physical location being examined, or ex situ using some other external means such as higher resolution imagery. All of the points on the map below were looked at in an ex situ manner using Google Maps satellite view. There are three different types of accuracy that can be calculated utilizing the sample points. These are the overall accuracy which is the totally correctly classified sample points, users accuracy which is the probability that a sample point is actually of the appropriate classification. If this is not the case it generates an error of commission which is when a point is committed to an incorrect class. And the last error is the producers error, which is the overall probability that any location on the map is correctly classified. Lets look at the example map.
There are 35 points on the map with a total accuracy of 66%. 12 of the points have been miss-classified. These points are in red. Correctly classified points are in green. A random stratified approach was taken to selecting sample point locations. Each category has a minimum of one point with categories that cover much larger areas having more sample points allocated. None more than 5 points total for proportionality. You can start to see themes on which points are inappropriately classified. For example, the light purple class 61 was determined to be entirely false throughout. Upon closer imagery inspection this region should all be classified as a 62, essentially going from forested wetland to emergent wetland.
Swapping trains of thought and looking at user accuracy we have a couple examples where categories 11 and 12 both had 4/5 (80%) user accuracy rating. However 12 did have 3 points that should have been classified to it. This means there was a possible total of 8 sites, 4 classified appropriately 1 not and 3 discovered during accuracy testing totaling 50% classification accuracy. This lab was a good look at the kinds of errors that can be made when classifying areas for land use and land cover and how to start avoiding them. Thank you.
This week I am specifically looking at the classification accuracy of last weeks generated Land Use Land Cover Example map of Pascagoula Mississippi. The overall objectives of this week were to explore the different types of accuracy and how to compare them with the data presented. This builds on last weeks map and is based on taking samples at locations of each classification type and determining if the sample is accurately portrayed by the assigned classification. these samples are typically generated in situ meaning at the physical location being examined, or ex situ using some other external means such as higher resolution imagery. All of the points on the map below were looked at in an ex situ manner using Google Maps satellite view. There are three different types of accuracy that can be calculated utilizing the sample points. These are the overall accuracy which is the totally correctly classified sample points, users accuracy which is the probability that a sample point is actually of the appropriate classification. If this is not the case it generates an error of commission which is when a point is committed to an incorrect class. And the last error is the producers error, which is the overall probability that any location on the map is correctly classified. Lets look at the example map.
There are 35 points on the map with a total accuracy of 66%. 12 of the points have been miss-classified. These points are in red. Correctly classified points are in green. A random stratified approach was taken to selecting sample point locations. Each category has a minimum of one point with categories that cover much larger areas having more sample points allocated. None more than 5 points total for proportionality. You can start to see themes on which points are inappropriately classified. For example, the light purple class 61 was determined to be entirely false throughout. Upon closer imagery inspection this region should all be classified as a 62, essentially going from forested wetland to emergent wetland.
Swapping trains of thought and looking at user accuracy we have a couple examples where categories 11 and 12 both had 4/5 (80%) user accuracy rating. However 12 did have 3 points that should have been classified to it. This means there was a possible total of 8 sites, 4 classified appropriately 1 not and 3 discovered during accuracy testing totaling 50% classification accuracy. This lab was a good look at the kinds of errors that can be made when classifying areas for land use and land cover and how to start avoiding them. Thank you.
No comments:
Post a Comment