Land Cover Classification Defining the pieces that make up the puzzle
What is a Classified Image Image has been processed to put each pixel into a category Image has been processed to put each pixel into a category Result is a vegetation map, land use map, or other map grouping related features Result is a vegetation map, land use map, or other map grouping related features Categories are defined by the intended use of the map Categories are defined by the intended use of the map Can be few or many categories, depending on the purpose of the map and available resources Can be few or many categories, depending on the purpose of the map and available resources
Land cover classification steps Define why you want a classified image, how will it be used? Define why you want a classified image, how will it be used? Decide if you really need a classified image Decide if you really need a classified image Define the study area Define the study area Select or develop a classification scheme (legend) Select or develop a classification scheme (legend) Select imagery Select imagery Prepare imagery for classification Prepare imagery for classification Collect ancillary data Collect ancillary data Choose classification method and classify Choose classification method and classify Adjust classification and assess accuracy Adjust classification and assess accuracy
How will you use the classified information? Are statistics okay or is mapped output necessary? Are statistics okay or is mapped output necessary? Is it going to be used primarily as a visual tool? Is it going to be used primarily as a visual tool? Will it be used as input to a model or for some other numerical analysis? Will it be used as input to a model or for some other numerical analysis? Are the results going to be used as a management tool? Are the results going to be used as a management tool? Do you really need to create a classified image? Do you really need to create a classified image?
Define the study area Reaching consensus among participants can be challenging Reaching consensus among participants can be challenging Often need to balance practical issues with desired output. For example, is it worth purchasing and processing extra imagery to include a small portion of the study area? Often need to balance practical issues with desired output. For example, is it worth purchasing and processing extra imagery to include a small portion of the study area? Should adjacent areas be included? Can adjacent areas be included using a buffer or by deliberate selection. Should adjacent areas be included? Can adjacent areas be included using a buffer or by deliberate selection.
Choosing a classification scheme A classification scheme defines the legend A classification scheme defines the legend There are several classification schemes available: There are several classification schemes available: Classification name URL Anderson National Land Cover Data FAO Land Cover Classification System Classification schemes can be hierarchical or non-hierarchical Classification schemes can be hierarchical or non-hierarchical Other attributes can be mapped including: Other attributes can be mapped including: Vegetation structure Vegetation structure Land cover disturbance Land cover disturbance Vegetation age (for example, primary and secondary) Vegetation age (for example, primary and secondary) Distribution of taxa Distribution of taxa Land use Land use Crown closure Crown closure
Other classification comments Each class should be well defined and documented so users of the final map know what the individual classes represent Each class should be well defined and documented so users of the final map know what the individual classes represent Rules for dealing with mixed classes, such as transition and mosaic vegetation should be developed Rules for dealing with mixed classes, such as transition and mosaic vegetation should be developed A minimum mapping unit can be defined to explicitly define the smallest feature that will be mapped as a single class A minimum mapping unit can be defined to explicitly define the smallest feature that will be mapped as a single class Defining classes is often an iterative process. A balance must be struck between what is desired based on the maps purpose and the classes that can be accurately and economically delimited Defining classes is often an iterative process. A balance must be struck between what is desired based on the maps purpose and the classes that can be accurately and economically delimited Above all, be consistent and document your methods Above all, be consistent and document your methods
Select imagery Selecting appropriate imagery is usually a subjective task and experience is very helpful Selecting appropriate imagery is usually a subjective task and experience is very helpful Image choice is often limited by budget and image availability Image choice is often limited by budget and image availability Spatial detail of the final classified map is limited by the input imagery Spatial detail of the final classified map is limited by the input imagery After assessing your imagery alternatives it may be necessary to redefine your goals, study area, and classification scheme After assessing your imagery alternatives it may be necessary to redefine your goals, study area, and classification scheme
Preprocessing – Radiometric corrections Goal is to calculate ground reflectance (ratio of the intensity of light reflected from a surface over the intensity of incident light) Goal is to calculate ground reflectance (ratio of the intensity of light reflected from a surface over the intensity of incident light) Satellite measures radiance at the sensor not surface reflectance Satellite measures radiance at the sensor not surface reflectance Due to irregularities in the earths surface and atmospheric scattering and absorption it is impossible to directly measure surface reflectance Due to irregularities in the earths surface and atmospheric scattering and absorption it is impossible to directly measure surface reflectance Many of the “easy” correction methods are not very effective Many of the “easy” correction methods are not very effective Radiometric correction is becoming more accessible to novice users Radiometric correction is becoming more accessible to novice users
Preprocessing – Geometric corrections Goal is to warp the image to match a map Goal is to warp the image to match a map Geometric processing includes: Geometric processing includes: Systematic corrections – removes distortions from the sensor and movement of the satellite and earth Systematic corrections – removes distortions from the sensor and movement of the satellite and earth Geo-referencing – Uses control points to warp the image to a map but in mountainous terrain distortions are still problematic Geo-referencing – Uses control points to warp the image to a map but in mountainous terrain distortions are still problematic Ortho-rectification – Corrects for distortions caused by terrain by using a DEM Ortho-rectification – Corrects for distortions caused by terrain by using a DEM
Collect ancillary data If additional data is available that can improve the classification it should be included. Some possible datasets include: If additional data is available that can improve the classification it should be included. Some possible datasets include: DEMs and their derived datasets (slope and aspect) DEMs and their derived datasets (slope and aspect) Climate data such as rainfall and temperature Climate data such as rainfall and temperature Vector overlays such as roads, rivers, and populated places Vector overlays such as roads, rivers, and populated places In many cases adequate datasets are not useful because they are too coarse, too expensive, or simply not available In many cases adequate datasets are not useful because they are too coarse, too expensive, or simply not available Incorporating ancillary data is not always straightforward and depends on the selected classification method Incorporating ancillary data is not always straightforward and depends on the selected classification method
Manual vs. Automated Classification Manual Manual Uses photo interpretation methods to delineate classes Uses photo interpretation methods to delineate classes Uses visual cues such as tone, texture, shape, size, shadows, and location Uses visual cues such as tone, texture, shape, size, shadows, and location Uses heads-up digitizing Uses heads-up digitizing Very subjective Very subjective Automated Automated Uses computer algorithms to group clusters based on similar characteristics Uses computer algorithms to group clusters based on similar characteristics Rules for classifying usually developed subjectively but these are applied objectively Rules for classifying usually developed subjectively but these are applied objectively Hybrid Hybrid Mixture of automated and manual methods Mixture of automated and manual methods Can be done by manually editing an automated classification result Can be done by manually editing an automated classification result
Supervised vs. Unsupervised Classification Supervised Supervised The analyst provides information, usually in the form of samples from the image, to train the algorithm to define classes The analyst provides information, usually in the form of samples from the image, to train the algorithm to define classes Requires prior information about the land cover Requires prior information about the land cover Unsupervised Unsupervised The computer uses an algorithm to group similar pixels into classes The computer uses an algorithm to group similar pixels into classes User provides parameters such as the number of classes to produce, and rules defining how classes should be merged and split as the algorithm runs User provides parameters such as the number of classes to produce, and rules defining how classes should be merged and split as the algorithm runs Usually an iterative algorithm Usually an iterative algorithm The user must associate each class with a particular land cover type The user must associate each class with a particular land cover type Hybrid Hybrid Uses unsupervised methods to create training data and then use a supervised algorithm for the final classification Uses unsupervised methods to create training data and then use a supervised algorithm for the final classification
The Sea of Algorithms There are dozens of algorithms that can be used There are dozens of algorithms that can be used Algorithm selection depends on what is available, the image type, available training data, ancillary layers available, and experience Algorithm selection depends on what is available, the image type, available training data, ancillary layers available, and experience Claims suggesting superior accuracy from one algorithm over another should be viewed with a gain of salt Claims suggesting superior accuracy from one algorithm over another should be viewed with a gain of salt Some popular algorithms include: Some popular algorithms include: ISODATA unsupervised classification ISODATA unsupervised classification Supervised statistical classification Supervised statistical classification Parallelepiped Parallelepiped Minimum distance Minimum distance Maximum likelihood Maximum likelihood Mahalanobis distance Mahalanobis distance Artificial neural net Artificial neural net Binary decision tree Binary decision tree Image segmentation (not really a classification tool) Image segmentation (not really a classification tool)
NDVI = (Near Infrared - Red) (Near Infrared + Red) Normalized Difference Vegetation Index (NDVI)
Identify several training sites for each category (Numerous small sites are preferable to a few large ones.) Identify several training sites for each category (Numerous small sites are preferable to a few large ones.) The size of each site generally should be 10 to 40 acres. Training areas for each category should contain a total of at least 100 pixels. Training areas for each category should contain a total of at least 100 pixels. Select representative sites across the image. Training areas (all sites for each category) must be uniform. (The histogram should be unimodal.) Training areas (all sites for each category) must be uniform. (The histogram should be unimodal.) Guidelines for Training Areas