lecture 22: remote sensing image processing and interpretation by austin troy university of vermont...
TRANSCRIPT
Lecture 22:Remote Sensing Image Processing
and Interpretation
By Austin TroyUniversity of Vermont
------Using GIS--Introduction to GIS
©2005 Austin Troy
Image Pre-Processing•Once an image is acquired it is generally processed to eliminate errors
•Geometric correction
•Radiometric correction
•It is also “enhanced” to make it more viewable
Introduction to GIS
©2005 Austin Troy
Image enhancement•For improving image quality, particularly contrast
•Includes a number of methods used for enhancing subtle radiometric differences so that the eye can easily perceive them
•Two types: point and local operations
•Point: modify brightness value of a given pixel independently
•Local: modify pixel brightness based on neighborhood brightness values
Introduction to GIS
©2005 Austin Troy
Contrast enhancement(point operation)
•Most images start with low contrast; these improve it
•Level slicing reclasses DNs into fewer classes, so differences can be more easily seen; colors or grayscale values can be assigned. Like resampling down radiometric resolution. Often used where histogram shows bimodal distribution of reflectance values
•Contrast Streching is the opposite, where a smaller number of values are stretched out over full DN range
Introduction to GIS
©2005 Austin Troy
Contrast enhancement•Here is what spectral histograms look like
Introduction to GIS
Note that DN is not zero for any of them
Source: http://www.sci-ctr.edu.sg/ssc/publication/remotesense/process.htm
©2005 Austin Troy
Contrast enhancement•Gray level thresholding: all pixel values below a lower threshold are mapped to zero and those above an upper threshold are mapped to 255. All other pixel values are linearly interpolated to lie between 0 and 255
Introduction to GIS
Source: http://www.sci-ctr.edu.sg/ssc/publication/remotesense/process.htm
Grey-Level Transformation Table for performing linear grey level stretching of the three bands of the image. Red line: XS3 band; Green line: XS2 band; Blue line: XS1 band.
©2005 Austin Troy
Contrast enhancement
Introduction to GIS
Source: http://www.sci-ctr.edu.sg/ssc/publication/remotesense/process.htm
•The image on the left is hazy because of atmospheric scattering; the image is improved (right) through the use of Gray level thresholding. Note that there is more contrast and features can be better discerned
©2005 Austin Troy
Spatial Feature Enhancement(local operation)
•Spatial filtering/ Convolution: neighborhood operations (like we reviewed for raster analysis), that calculate a new value for the center pixel based on the values of its neighbors within a window (see “More Raster Analysis” lecture for more); includes low-pass (emphasizes regional spatial trends, demphasizes local variability ) and high-pass (emphasizes local spatial variability) filters
Introduction to GIS
©2005 Austin Troy
Spatial Feature Enhancement•Edge Enhancement: This is a convolution method that combines elements of both low and high-pass filtering in a way that accentuates linear and local contrast features without losing the regional patterns
•First, a high-pass image is made with local detail
•Next, all or some of the gray level of the original scene is added back
•Finally, the composite image is contrast stretched
Introduction to GIS
©2005 Austin Troy
Classification of Imagery•Classification can be used for numerous purposes, like classifying geology, water temperature, soil moisture, other soil characteristics, water sediment load, water pollution levels, lake eutrophication, flood damage estimation, groundwater location, vegetative water stress, vegetative diseases and stresses, crop yields and health, biomass quantity, net primary productivity, forest vegetation species composition, forest fragmentation, forest age (in some cases), rangeland quality and type, urban mapping and vectorization of manmade structures
•One of the most common applications of classification is land cover and land use mapping
Introduction to GIS
©2005 Austin Troy
Land Cover/ Land Use Mapping•Land cover refers to the feature present and land use refers to the human activity associated with a plot of land
•The LU/LC classes to be derived will depend on the system being used. One of the most common is the USGS Anderson Classification System (Anderson et al. 1976). This classification scheme is hierarchical, with nine very general categories at Level I, and an increasing number of classes and detail and level increases. Paper available online at http://landcover.usgs.gov/pdf/anderson.pdf
•Anderson system intermixes land use and land cover metrics, by inferring land use from land cover. Unfortunately, land cover can only tell us a limited amount about land use—think of outdoor recreation as a land use. Need additional data for these classes.
Introduction to GIS
©2005 Austin Troy
Land Cover/ Land Use Mapping•Land use and land cover classification system for use with remote sensor data (Anderson et al. 1976)
•Level I Level II
•1 Urban or Built-up Land 11 Residential
12 Commercial and Services
13 Industrial
14 Transportation, Communications, and Utilities
15 Industrial and Commercial Complexes
16 Mixed Urban or Built-up Land
17 Other Urban or Built-up Land
Introduction to GIS
©2005 Austin Troy
Land Cover/ Land Use Mapping•Level I Level II
•2 Agricultural Land 21 Cropland and Pasture
22 Orchards, Groves, Vineyards, Nurseries, and Ornamental Horticultural Areas 23 Confined Feeding Operations
24 Other Agricultural Land
•3 Rangeland 31 Herbaceous Rangeland
32 Shrub and Brush Rangeland
33 Mixed Rangeland
•4 Forest Land 41 Deciduous Forest Land
42 Evergreen Forest Land
43 Mixed Forest Land
•5 Water 51 Streams and Canals
52 Lakes
53 Reservoirs
54 Bays and Estuaries
•6 Wetland 61 Forested Wetland
62 Nonforested Wetland
Introduction to GIS
©2005 Austin Troy
Land Cover/ Land Use Mapping•Level I Level II
•7 Barren Land 71 Dry Salt Flats.
72 Beaches
73 Sandy Areas other than Beaches
74 Bare Exposed Rock
75 Strip Mines Quarries, and Gravel Pits
76 Transitional Areas
77 Mixed Barren Land
•8 Tundra 81 Shrub and Brush Tundra
82 Herbaceous Tundra
83 Bare Ground Tundra
84 Wet Tundra
85 Mixed Tundra
•9 Perennial Snow or Ice 91 Perennial Snowfields
92 Glaciers
Introduction to GIS
©2005 Austin Troy
Land Cover/ Land Use Mapping•Level 3 and 4 categories deliver even more detail.
•USGS only specifies classifications for 1 and 2. They suggest that higher level classification be designed by local planners who know the land uses, because of the narrowness of the categories
• As an example for level 3, with “urban” (level 1) “residential” (level 2) category, includes single family home (111), multifamily home (112), group quarters (113), mobile home parks (115), etc.
•LANDSAT data can be used to generate level 1 easily, level 2 with some finesse (15 to 20 m resolution recommended)
•Levels 3 and 4, IKONOS data or aerial photographs are needed. Level 4 requires much supplemental information
Introduction to GIS
©2005 Austin Troy
Land Cover/ Land Use Mapping•Here is an example of LANDSAT data classified using the Anderson System
Introduction to GIS
©2005 Austin Troy
Image classification•This is the science of turning RS data into meaningful categories representing surface conditions or classes
•Spectral pattern recognition procedures classifies a pixel based on its pattern of radiance measurements in each band: more common and easy to use
•Spatial pattern recognition classifies a pixel based on its relationship to surrounding pixels: more complex and difficult to implement
•Temporal pattern recognition: looks at changes in pixels over time to assist in feature recognition
Introduction to GIS
©2005 Austin Troy
Spectral Classification•Two types of classification:
•Supervised: the analyst designates on-screen “training areas” known land cover type from which an interpretation key is created, describing the spectral attributes of each cover class . Statistical techniques are then used to assign pixel data to a cover class, based on what class its spectral pattern resembles.
•Unsupervised:automated algorithms produce spectral classes based on natural groupings of multi-band reflectance values (rather than through designation of training areas), and the analyst uses references data, such as field measurements, DOQs or GIS data layers to assign areas to the given classes
Introduction to GIS
©2005 Austin Troy
Spectral Classification•Unsupervised:
•Computer groups all pixels according to their spectral relationships and looks for natural spectral groupings of pixels, called spectral classes
•Assumes that data in different cover class will not belong to same grouping
•Once created, the analyst assesses their utility
Introduction to GIS
Source: F.F. Sabins, Jr., 1987, Remote Sensing: Principles and Interpretation.
Spectral class 1
Spectral class 2
©2005 Austin Troy
Spectral Classification•Unsupervised:
•After comparing the reclassified image (based on spectral classes) to ground reference data, the analyst can determine which land cover type the spectral class corresponds to
•Has advantage over supervised classification: the “classifier” identifies the distinct spectral classes, many of which would not have been apparent in supervised classification and, if there were many classes, would have been difficult to train all of them. Not required to make assumptions of what all the cover classes are before classification.
•Clustering algorithms include: K-means, texture analysis
Introduction to GIS
©2005 Austin Troy
Spectral Classification•Unsupervised:
•Here’s an example
Introduction to GIS
Source: http://elwood.la.asu.edu/grsl/lter/fig5.html
©2005 Austin Troy
Spectral Classification•Unsupervised:Another example
Introduction to GIS
Source: http://mercator.upc.es/nicktutorial/Sect1/nicktutor_1-14.html
©2005 Austin Troy
Spectral Classification•Supervised:
•Better for cases where validity of classification depends on a priori knowledge of the technician
•Conventional cover classes are recognized in the scene from prior knowledge or other GIS/ imagery layers
•Therefore selection of classes is pre-determined and supervised
•Training sites are chosen for each of those classes
•Each training site “class” results in a cloud of points in n dimensional “measurement space,” representing variability of different pixels spectral signatures in that class
Introduction to GIS
©2005 Austin Troy
Spectral Classification•Supervised: Here are a bunch of pre-chosen training sites of known cover type
Introduction to GIS
Source: http://mercator.upc.es/nicktutorial/Sect1/nicktutor_1-15.html
©2005 Austin Troy
Spectral Classification•Supervised:
•The next step is for the computer to assign each pixel to the spectral class is appears to belong to, based on the DN’s of its constituent bands
• There are numerous algorithms the computer uses, including:
•Minimum distance to means classification (Chain Method)
•Gaussian Maximum likelihood classification
•Parallelpiped classification
Introduction to GIS
©2005 Austin Troy
Spectral Classification•Supervised:
•These algorithms look at “clouds” of pixels in spectral “measurement space” from training areas, and try to determine which “cloud” a given non-training pixel falls in.
•The simplest method is “minimum distance” in which a theoretical center point of point cloud is plotted, based on mean values, and an unknown point is assigned to the nearest of these. That point is then assigned that cover class.
•They get much more complex from there.
Introduction to GIS
©2005 Austin Troy
Spectral Classification•Supervised:
•Examples of two classifiers
Introduction to GIS
Source: http://mercator.upc.es/nicktutorial/Sect1/nicktutor_1-16.html
©2005 Austin Troy
Object-Oriented Classification
Introduction to GIS
Traditional classifiers don’t work as well for new generation of high resolution data, like this 2 foot Emerge Color infrared airphoto. Why? Meaningless to classify each pixel
©2005 Austin Troy
Object-oriented classification Steps:
•Segmentation of rasters into polygon objects
•Objects are defined such that they minimize within-unit heterogeneity and maximize between unit heterogeneity, subject to some user defined parameters.
•The user can control the scale parameter for acceptable level of heterogeneity. They can also control the degree to which segmentation is based on spectral or spatial characteristics, since heterogeneity is defined in terms of both. By repeating the segmentation with different scale parameters, the user can create a nested hierarchy of objects>>big objects containing smaller objects, containing smaller objects
Introduction to GIS
©2005 Austin Troy
Object-oriented classification
Introduction to GIS
©2005 Austin Troy
Object-oriented classification
Introduction to GIS
©2005 Austin Troy
Object-oriented classification Steps:
•Two levels of segmentation
Introduction to GIS
Source/More info: see Ecognition website: http://www.definiens-imaging.com/index.htm
©2005 Austin Troy
Object-oriented classification Steps:
•Following segmentation, each object is encoded with information about its tone, shape, area, context, neighborhors and spectral characteristics (e.g. mean, standard deviation, max, min or each band’s spectral reflectance)
•This information can be used for feature extraction in which objects’ properties are analyzed to look for characteristics that help to discriminate one object type from another. That is, what object information helps discriminate one from another?
Introduction to GIS
©2005 Austin Troy
Object-oriented classification
Introduction to GIS
©2005 Austin Troy
Object-oriented classification Steps:
•Then objects are classified by either defining training areas of known cover type (known as supervised fuzzy classification) or creating class descriptions organized through inheritance-based rules into a knowledge base (known as fuzzy knowledge base classification).
Introduction to GIS
©2005 Austin Troy
Object-oriented classification Steps:
•In the knowledge base approach, complex membership functions can be derived that describe characteristics that are typical or atypical for a certain class. The more a given object displays the characteristics, the more likely it is to be classified into the class to which those characteristics pertain. Characteristics can be based on spectral response summary statistics, shape characteristics, adjacency, connectivity, and overlay with certain thematic features.
Introduction to GIS
©2005 Austin Troy
Object-oriented classification
Introduction to GIS
©2005 Austin Troy
Object-oriented classification Steps
Introduction to GIS
• The classification can be hierarchical and nested, with finer classifications within coarser ones
• Small classified objects can be aggregated up to large object classes and large objects can be split into smaller ones. Can then assign different segmentations to different class hierarchy level
• Allows for high precision classifications within coarser, general classifications
©2005 Austin Troy
Object-oriented classification Steps
Introduction to GIS
• The classification can be hierarchical and nested, with
©2005 Austin Troy
Object-oriented classification Steps
Introduction to GIS
• Can use additional thematic layers to populate the knowledge base and create rules about what a certain class can be on top of, next to, or near. This can increase the accuracy of classifications, especially as you increase categorical precision and start getting into classifiying land uses in addition to land cover
• Hence, when you do training areas, you not only get average spectral responses and shape metrics for a class, but also can get average values from underlying layers to help increase classification accuracy
• Examples: farm fields as fn of slope, soils, etc; different suburban development types as function of distance to urban centers, income, crime, etc.
©2005 Austin Troy
Object-oriented classification Software
Introduction to GIS
•eCognition: one of the top Object oriented classification software packages
More info: see Ecognition website: http://www.definiens-imaging.com/index.htm
©2005 Austin Troy
Accuracy Assessment•This is one of the most important parts of image classification.
•Error rates can be very high in classification accuracies, especially with lower resolution data, and where pixels are mixed
•This is often the most time consuming part of image classification
•NLCD effort undertook effort to classify errors in each type of land cover, broken down by region of the US
•User’s accuracy for type X: Percent of pixels classified as X (e.g. “forest”) that really are forest (measures errors of commission). Producer’s accuracy: percent of pixels that were classified as other than forest but really are forest (measures errors of omission).
Introduction to GIS
©2005 Austin Troy
Accuracy Assessment:example
Introduction to GIS
Table 1 Accuracy Assessment of LULC classification
Classified data
Reference data
SumUser Acc.
%Urban Forest Other
Urban 71 13 16 100 71
Forest 0 100 0 100 100
Other 2 10 88 100 88
Sum 73 123 104 300
Producer Acc. (%)
97.3 81.3 84.6
Overall accuracy (%) 86.3