Nate Herold

by Friday, October 28, 2011 @ .

I am with the Coastal Remote Sensing Program at the NOAA Office for Coastal Management, where I lead the office's land cover mapping activities. I manage the Coastal Change Analysis Program (C-CAP), a nationally standardized database of land cover and change information developed for the nation’s coastal regions. C-CAP products are developed from remotely sensed imagery and can be used to track changes in the landscape through time. I came to NOAA after several years working in the private sector, where I was involved in the production of land cover products worldwide. I have a bachelor’s in soil science from Penn State University and a master’s in geography from George Mason University.

Read More >

As a producer of moderate resolution land cover data, I am often asked questions about the spatial resolution of our data.  I have gotten used to the fact that 30 meter pixels of our C-CAP data are not always seen as ultra-sexy and the reaction that they “are not good enough.”  And they aren’t in some instances, but then again, sometimes they are (and sometimes it doesn’t matter, as they are the only/best thing available).  What I tell people, though, is that the resolution itself isn’t enough to determine whether these products are good enough or not.  I tell them there are two other key pieces of information to consider: 1) what is the Minimum Mapping Unit (MMU) of the product (or what is the smallest feature that is being reliably mapped)? And, 2) what is it you are trying to map? 

Minimum Mapping Unit:

There are two ways to think about MMU:

  1. What is technically possible based upon the image data that the land cover is being derived from?
  2. What is reality based upon the methods and techniques used in producing that product? 

The smallest possible feature that could be mapped would be equal to one pixel.  For a 30 meter image source this would be 30 meter by 30 meter area (or approximately 1/4th of an acre).  For a one meter image this would be a 1 square meter area. But, it is generally agreed that the smallest observable feature that can reliably be identified would need to be four contiguous pixels in size (60 by 60 meters or 2 by 2 meters, respectively for the examples above). This is because a quarter-acre-sized feature may not fall entirely within one given pixel but may instead be split among as many as 4 pixels, therefore making up only a minority of any one of those pixels and not being the dominant feature reflected in any. 

Many products (especially those derived from higher resolution imagery) are not based on individual pixels.  These products are often based on polygons derived from the imagery in some way (use of image segments / object or manually derived through photo-interpretation / heads-up digitizing).  When this is the case, this reality trumps what might have been possible based solely on pixel size. 

In reality then, you could end up with a land cover product that while derived from higher resolution imagery, may have a less detailed MMU then that possible.  If this is the case, small linear features and the boundary line between features may be mapped in more detail, but the size of smallest general features mapped would not necessarily be better. 

The tipping point for this comparison between higher resolution and moderate resolution might be approximately 1 acre (i.e. the area approximately equal to four 30 meter pixels, as discussed above).  If the high res data aren’t resolving more detail than that, then you would be highly likely to see these same features in a moderate resolution product as well.

The figure on the right (top) is an example of a land cover data set developed from digital orthophotography with a 0.25 meter pixel resolution and a minimum mapping unit of 0.25 acres.  Note the similarity of most features to those seen in the corresponding 30 meter land cover product (seen below it).  You may also note that there is some added detail in the linear tidal creeks that are mapped as water in blue.

 So, what are you trying to map?

Or more to the point, how big is it?  Smaller than the MMU?  As the mob would say, “fuhgeddaboudit (forget about it),” or at least it isn’t going to be likely. You just can’t map what you can’t see.  Bigger than the MMU?  Don’t worry, be happy…  at least until you turn to consider the accuracy of the product…  but that is detail for another time.

FOOTNOTES:

* For products that are not based solely on pixels, the Minimum Mapping Unit should be clearly defined and included as part of the metadata.  

**For pixel based products, you can determine the appropriate resolution necessary for your applications based upon the smallest feature you want to resolve (i.e. MMU).  The pixel size must be half the smallest dimension of the feature in question. For instance if you want to find a car (which would be ~10 feet  x 6 feet), then the smallest dimension of that car is 6 feet, so your pixel size must be no larger 3 x 3 feet to reliably identify and map cars.

6 Responses to “Resolution Vs. Minimum Mapping Unit: Size Does Matter”

  1. Kirk Waters

    I will appreciate if you could respond the following point
    Which map scale we can expect from a satellite pixel size? For example the TM or ETM+ has roughly 30m ground resolution, then, what could be the finest scale of the extracted map?
    Are there any role or a table for the converting the satellite pixel size to map scale?
    khajedin@yahoo.com

  2. Kirk Waters

    It is probably important note that pixel (spatial) resolution and scale are not interchangeable terms, and that they mean different things. With raster imagery, you need to deal with both.

    Scale is the relative difference in size or distance between features in the image and features as the actually exist on the ground. 1:100,000 (one to one hundred thousand) means that every one inch (or any unit) in the map represents 100,000 units on the ground (in reality). Pixel resolution is the size of the area on the ground that each pixel within the image is representing (i.e. a 30m raster data set portrays the ground that it images in 30m x 30m pixel units).

    Raster data are not typically attributed to a specific scale, though the resolution is an indication of the potential detail in an image. The actual detail you will see will also depend on the scale at which you view the image (i.e. how closely you look at the pixels). That said, there are general rules-of thumb associated with the “right” or “appropriate” map scale at which you are utilizing the maximum level of image detail, given the images spatial resolution.

    See the following blog from ESRI on the topic (http://blogs.esri.com/esri/arcgis/2010/12/12/on-map-scale-and-raster-res…). Essentially, you can multiple the pixel size (in meters) by 2,000 to identify the most detailed rule-of-thumb scale level for viewing. 30m data, for example, can be viewed at as detailed a level as 1:60,000… Though, it may be important to note that many people typically view data at slightly less detailed levels. 1:100,000 for such 30m data is standard.

  3. Kirk Waters

    I am contemplating between the use of 2 types of satellite images for mapping mangroves and coral reefs and these are ASTER images and Landsat images. Since they both have different spatial resolutions , how can I clearly determine the number of pixels that I can identify to map let’s say an area covering 62.5 square kilometers?Thanks in advance. buklind@gmail.com

  4. Kirk Waters

    Linda,

    In a lot of ways Landsat and ASTER are very similar data sources. As you mention, ASTER does have a slight advantage in terms of spatial resolution (15 meters on at least 4 of its bands). And, it also has a number of additional infrared bands that Landsat does not (which can also be an advantage)… though most of those are 30 meter in resolution. It is, however, more of a “research” sensor and does not collect data as systematically as Landsat does, so there can be issues in getting complete (or up-to-date) coverage of areas. Something to check in on as a first step, perhaps.
    See http://asterweb.jpl.nasa.gov/data.asp for various ways to access ASTER data

    As far as the number of exact pixels, that would depend on the shape of the area, etc. but roughly you are looking at 69,000+ 30m pixels (and 4 times that amount of 15m pixels). But, the bigger question (that I wrote about in my blog) might be how large are the individual areas/clumps of mangrove that you want to map… Landsat (and most of the ASTER bands) wouldn’t be much use in capturing areas that are less than about 1 acre in size. Have you considered aerial imagery as the alternative to Landsat, or ASTER? We have many aerial data set available within the Digital Coast. And, in your case, you are not looking at a huge total study area, and fairly specific features that you want to map. I don’t know what type of mapping method you may be thinking of, but you might be better off with higher resolution imagery which would supply you with much more contextual detail… that would likely improve your mapping accuracy (no matter what MMU you choose to map to). This is especially true when dealing with the coral areas, and the complicating factor of viewing those through the water column.

    Anyway, this is far from a simple answer, and depends on a number of factors… If you would like to discuss more, feel free to contact me directly.

  5. Kirk Waters

    Hello Nate,

    Would you consider it possible to have a MMU smaller than an acre using Landsat if the pixels are accurately unmixed? For a school project I have used Landsat data and linear spectral unmixing to detect changes on forest to bare/impervious surface. I have an overall accuracy of 85% detecting changes that are smaller than a pixel and I finding it quite difficult to figure out/justify a MMU. Thanks for this blog and any advice you may be able to offer. -Sean

  6. Kirk Waters

    Sean,

    Short answer: Most of the time MMU is based on the resolution of the imagery, or the output raster that you have created. The idea being that you can’t map to more detail than the units you are mapping with (i.e. you can’t go to 4 decimal places if you only have space to record 2). For landsat imagery, and the land cover that I produce from it (http://coast.noaa.gov/digitalcoast/data/ccapregional), the MMU would be .22 acres (30m X 30m). That is the absolute smallest division that a map with that resolution can make.

    Longer answer: Single feature maps, like change VS. no-change masks, or products that map % surfaces can map features at sub pixel levels (like the NLCD Percent Impervious products – http://coast.noaa.gov/digitalcoast/data/nlcd-impervious) can be a little more tricky of an answer. The % surface, in particular is providing more detail but the MMU is still 30m. In these products a pixel can be said to be 30% impervious, but those classes are still associated with the full pixel area… and you can’t say much about what might make up the other 70% of that pixel (or at least not in the same product), or where within the pixel those things occur.

    Same is true for change masks. For both, you are mapping one thing at the expense of another. You may be able to detect and map areas of change smaller than your MMU… And I’m not surprised that bare and impervious surfaces would show up well in your analysis, as bright change features like these tend to be over-represented (more visible) than other features within the imagery… but if only half that pixel is change then the other half is not really represented in your product anymore (because you have called the whole area change). That might be perfect, because that is what you care about, you just need to recognize that fact.

    Hope that helps.