6. Image Classification

Transcription

1 6. Image Classification 6.1 Concept of Classification Objectives of Classification Advantages of Multi-Spectral data for Classification Variation of Multi-Spectra Data Segmentation in Feature Domain Supervised and Un-Supervised Calssification Land Cover and Land Use Existing Land Cover Class Objectives of Classification To create Maps such as Landuse Map, Forest Map, Crop Map, Shrimp pond Map, Mangrove Map, etc. Carry out quantitative interpretation using mathematical / statistical modeling. To assign corresponding class to groups with homogeneous characteristics, with the aim of discriminating multiple objects from each other within the image. The level is called class. Classification will be executed on the base of spectrally defined features, such as density, (texture etc. ) in the feature space. It can be said that classification divides the feature space into several classes based on a decision rule. Classes are for such as Land use, Land Cover, Crop Type, Forest Types, and etc. RS Image Classification Concept of Classification of Remote Sensing Multi-Spectral Data Classification Assumption - Different surface materials have defferent sepectral reflectance K-dimensional vector ( K:number of band ) divide K-dimensional feature space into few regions ( classes ) Segmentation in Feature Domain In general, the separation of all classes requires more than two spectral bands. Because the clusters occur in K-dimensions. Percent Reflectance Spectral Reflectance Vegetation Soil Clear River Water Turbid River Water Wavelength (µm) 1

2 Multi-spectral Classification The spectral signature is a K-dimensional vector whose coordinates are the measured radiance in each spectral band. If every pixel from each land cover has same radiance with in the class, only 1 band (IR) would be enough for classification for the case of water, soil and vegetation below. Variation of Multispectral data In reality, the spectral radiance of a given surface material is not characterized by a single, deterministic curve, but by a family of curves with a range of variability. VR NIR Segmentation in Multi-dimensional feature space Thus, it is very common to find big overlaps among distributions in one band information. By combining other bands, we can improve the accuracy of classification, which is a segmentation in a multi-dimensional feature space. NIR NIR Supervised and Un-Supervised Classification Supervised Classification Classify each pixel into a pre-established class. Population statistics of each class is to be identified by training areas. Each pixel will be classified into a class which has similar ( nearest ) property with the pixel. Un-supervised Classification Analyze inherent structure of the data Unconstrained by external knowledge about area When knowledge about the area is not enough Combination Un-Supervised Classification -> Ground Truth -> Supervised Classification Land Cover and Land Use Land Cover means materials which covers the ground. ( Soil ) Land use is how land is being used. ( Play ground, Harvested Paddy ) A land use class consists of some land cover classes. Urban Area ( land use ) consists of concrete, vegetation, bare land, etc.. Generally, image classification gives land cover, but not land use, because what RS is observing is spectral signature of object ( material ). Existing Land Cover Classification System There are several organizations/groups such as IGBP, UNEP/FAO, USGS, which are trying to develop land cover data set of global or continental area. The Land Cover Working Group(LCWG) of the Asian Association on Remote Sensing(AARS) also aims to develop global or continental land cover data set. It is a global land cover data set but focuses on Asian and Oceania regions in terms of ground truth. Appropriate Land classification system ( class ) are developed according to the project s objectives 2

3 Global Ecosystems Legend Value Description Global Ecosystems (Olson, 1994a, 1994b) Global Ecosystems Legend Value Description 1 Urban 2 Low Sparse Grassland 3 Coniferous Forest 4 Deciduous Conifer Forest 5 Deciduous Broadleaf Forest 6 Evergreen Broadleaf Forests 7 Tall Grasses and Shrubs 8 Bare Desert 9 Upland Tundra 10 Irrigated Grassland 11 Semi Desert 12 Glacier Ice 13 Wooded Wet Swamp 14 Inland Water 15 Sea Water 16 Shrub Evergreen.. USGS USGS Land Use/Land Cover System Legend (Modified Level 2) Value Code Description ( Level 1 Class Value Code Description 1 Urban or Buil-up Land Urban and Built-Up Land 2 Agricultural Land Dryland Cropland and Pasture Irrigated Cropland and Pasture Mixed Dryland/Irrigated Cropland and Pasture Cropland/Grassland Mosaic Cropland/Woodland Mosaic 3 Rangeland Grassland Shrubland Mixed Shrubland/Grassland Savanna LCWG, AARS The Land Cover Working Group(LCW G) of the Asian Association on Remote Sensing(AARS) also aims to develop global or continental land cover data set. It is a global land cover data set but focuses on Asian and Oceania regions in terms of ground truth. Development of Land Classification System Considering objectives of the classification, available data, cost, technical feasibility, we have to develop a classification system. Agriculture monitoring? Urban environment? Flood modeling? Flood damage assessment? Forestry? 6.2 Ground Truth Ground truth is simply observations or measurements made at or near the surface of the earth in support of an air or spacebased remote sensing survey. The location will be acquired by GPS to identify the location on RS image To understand the real situation and phenomena on the ground To provide reference data for Classification Development of Classification Procedure Estimation of statistical property of the class Verification Accuracy Assessment Ground truth may consist of several types of data acquired before, during, and after an image acquisition. Such measurements and observations may include, but are not limited to: 3

4 Measurement and observation Remote Sensing for Rice Yield Estimation in Indonesia Heading Planting GPS, algae sampling, and quadrat mapping in Yaquina Bay, Oregon spectral measurements of grasses being made with two separate spectrometers Ripening Harvest Rice main growth stages OPTICAL SATELLITE DATA Pamanukan SPOT DATA 1999/08/05 Taking Spectrophotometer readings Ciasem Patokbeusi Blanakan Binong Pusakanagara LANDSAT-TM DATA 1997/07/28 Heading Harvested Planting Blanakan Ciasem Patokbeusi Pamanukan Binong Pusakanagara Results Model Developments Develop a model to find the correlation between LAI and NDVI derived from spectrophotometer readings NDVI ((NIR-IR)/(NIR+IR)) LAI Vs. NDVI (Spectophotometer readings) y = 0.233Ln(x) R 2 = Leaf Area Index Rice Growth Monitoring Using Near Real Time RADARSAT Fine Beam SAR Data in Pathumthani Deployment sites of corner reflectors, plotted On ADEOS AVNIR image Canada Japan Thailand Within 8 hrs after reception 4

5 Field locations of reflectors and corresponding views in the (11a) St ation No. 1 image Reflector: 8.53 db Background: db (11b) Station No. 2 Reflector: 9.89 db Background: db (11c) St ation No. 3 Reflector: 9.48 db Background: db for geometric correction of RADARSAT image For Better Overlay of Radar Image and Field Survey Result GPS:Latitude E Early Vegetative Growth Date: 9/03/2003 Land Cover: Early vegetative Growth Time: 15:45 Location: Klong 3 Longitude N (11d) Station No. 4 Reflector: 9.21 db Background: db Early Reproductive Growth Date: 9/03/2003 Time: 15: Supervised Classification Steps for Supervised Classification Land Cover: early reproductive Growth GPS:Latitude E Location: Klong 3 Longitude N In order to determine a decision rule for classification, it is necessary to know the spectral characteristics or features with respect to the population of each class. Due to atmospheric effects, direct use of spectral features measured on the ground are not always available. Sampling of training data from clearly identified training areas, corresponding to defined classes is usually made for estimating the population statistics. Statistically unbiased sampling of training data should be made in order to represent the population correctly. Sampling of Training Data Get statistical characteristics for each class A representative area for each desired class must be located by an analyst. analyst knowledge, field survey, aerial photographs, existing maps Select homogeneous area Divide the class into several homogeneous sub-class Select sufficient number of pixels to estimate class statistics Unbiased selection of area Estimation of Population Statistic 5

6 Classifier There are a LOT of classifier algorithms. Such as Parallelepiped Classifier Decision Tree Classifier Minimum distance Classifier Maximum likelihood Classifier Parallelpiped Classifier The minimum and maximum DNs for each class are determined and are used as thresholds for classifying the image. Benefits: simple to train and use, computationally fast Drawbacks: pixels in the gaps between the parallelepipes can not be classified; pixels in the region of overlapping parallelepipes can not be classified. Parallepiped Classifier Decision Tree Classifier The decision tree classifier is an hierarchically based classifier which compares the data with a range of properly selected features. The selection of features is determined from an assessment of the spectral distributions or separability of the classes. There is no generally established procedure. Therefore each decision tree or set of rules should be designed by an expert. When a decision tree provides only two outcomes at each stage, the classifier is called a binary decision tree classifier (BDT). Decision Tree Classifier Minimum Distance Classifier A centroid for each class is determined from the data by calculating the mean value by band for each class. For each image pixel, the distance in n- dimensional distance to each of these centroids is calculated, and the closest centroid determines the class. Benefits: mathematically simple and computationally efficient Drawback: insensitive to different degrees of variance in spectral response data. 6

7 Minimum Distance Classifier Maximum Likelihood Classifier Most Popular methods Maximum likelihood classification uses mean and variance-covariance in class spectra to determine classification scheme. It assumes that the spectral responses for a given class has normal distribution. A pixel with the maximum likelihood is classified into the corresponding class Bayes Theory feature x --- for example, the gray level of each pixel p( x i ) : probability density function in class i p( i ) : a priori probabilities p( i x ) : a posteriori probabilities Bayes Rule p( i x ) : p( x i ) p( i ) / p( x ) If we observed feature x, what is the probability to be class i? p(x ) = p ( x i ) p( i ) Bayes Dicision Rule one dimensional, two-class classification problem a pixel belongs to class 1 if p(x 1)p(1) > p(x 2)p(2) a pixel belongs to class 2 if p(x 2)p(2) > p(x 1)p(1) 0.3 f Forest f Bayes Decision Rule Agriculture f p(f1 Forest) *p(forest)= 0.3 x 0.6 = 0.18 p(f2 Forest) *p(forest)= 0.7 x 0.6 = 0.42 p(f1 Agr) *p(agr)= 0.9 x 0.4 = 0.36 p(f2 Agr) *p(agr)= 0.1 x 0.4 = p(forest f1)=p(f1 Forest)*p(Forest) / p(f1) = 0.18 / 0.54 = 0.33 p(agr f1)=p(f1 Agr)*p(Agr) / p(f1) = 0.36 / 0.54 = 0.67 p(forest f2)=p(f2 Forest)*p(Forest) / p(f2) = 0.42 / 0.46 = 0.91 p(agr f2)=p(f2 Agr)*p(Agr) / p(f2) = 0.04 / 0.46 = 0.09 f2.04 Forest p(forest) = 0.6 p(agr) = 0.4 P(f1 Forest)=0.3 P(f2 Forest)=0.7 P(f1 Agr)=0.9 P(f2 Agr)=0.1 f1 Agri f Forest Ag Discriminant Function The Bayes Dicision Rule is restated as a pixel belongs to class 1 if D1(x) > D2(x) a pixel belongs to class 2 if D2(x) > D1(x) where Di is called discriminant function and is given by Di(x) = p( x i ) p( i ) However P(i) is unknown, we assume p(i)=p(j) Assumption of Normal Distribution If the class probability distributions are normal p( x i ) = exp ( ( x i ) ) i i = mean of x for classi 2 i = varianceof x for classi Bayes optimal discriminant function for class i is then D i (x) = ln [ p( x i ) p( i ) ] = ln [ p( i ) ] 1 2 ln [ 2 ] 1 2 ln [ i 2 ] ( x i ) 2 2 i 2 p(i) is unknown. Assumption of p(i) = p(j), D i (x) = 1 2 ln [ i 2 ] ( x i ) 2 2 i 2 7

8 Extension to K Dimension p( x i ) = 1 ( 2 ) K / 2 1/2 exp[ 1 2 ( X i )t 1 i ( X i ) ] i D i (x) = ln [ p( i ) ] K 2 ln [ 2 ] 1 2 ln [ i ] 1 2 ( X i )t i 1 ( X i ) D i (x) = 1 2 ln [ i ] 1 2 ( X i )t i 1 ( X i ) Thresholding Eliminate pixels which have low posteriori probability X : vectorof imagedata( K dimension) X= [ x 1, x 2,...,x k ] i: meanvectorfor classi i = [ m 1, m 2,...,m k ] i: variance covariancematrixfor class i i = k1 k2 1k 2k kk i : determinantof i i 1 : inversematrixof i Actual Distribution 6.4 Unsupervised Classification To determine the inherent structure of the data, unconstrained by external knowledge about the area. To produce clusters automatically, which consists of pixels with similar spectral signature Hierarchical Clustering Evaluate distance between clusters Merge a pair of clusters which have the minimum distance. Members are not reallocated to different clusters Non-Hierarchical Clustering K-mean, ISODATA method Reallocation of members Merge and Division of clusters Hierarchical clustering ISODATA method 8

9 ISODATA Unsupervised Classification example Allocation of Land Cover/Use to Clusters Un-supervised classification gives only clusters, without any interpretation; land cove/use Operator must give class name to each clusters. Usual Process Un-Supervised Classification before field survey. One land cover/use might have been divided into several clusters. More than two land cover/use classes might have been merged in one cluster. Visit area where operator cannot identify land use/cover. Based on above, carry out supervised classification 6.5 Accuracy Assessment Accuracy assessments determine the quality of the information derived from remotely sensed data (Congalton and Green, 1999). Accuracy assessment is important to produce reliable maps. Assessments can be either qualitative or quantitative. In qualitative assessments, we determine if a map "looks right" by comparing what we see in the imagery with what we see on the ground. However quantitative assessments attempt to identify and measure remote sensing-based map error. In such assessments, we compare map data with reference or ground truth data. Reference/Ground truth data collection Usually we divide ground truth into two. 50% is used for supervised classification training 50% is used for accuracy assessment Aerial photographs Other Maps Ground based data is assumed to be 100% correct in accuracy assessments, hence it's very important that the data is collected carefully. It should be collected consistently with vigilant quality control. Common quantitative error assessments Error Matrix or Confusion Matrix - assesses accuracy for each class as well as for the whole image; this includes errors of inclusion and errors of exclusion We must accept some level of error as a trade off for the cost savings of remotely sensed data (Congalton and Kass, 1999) Reference(Ground Truth) Confusion Matrix Classified Urban Crop Range Water Forest Baren Total PA EO Urban % 35.9% Crop % 25.0% Range % 46.1% Water % 34.1% Forest % 27.5% Baren % 39.8% Total CA 60.5% 74.6% 69.6% 61.9% 63.1% 51.1% EC 39.5% 25.4% 30.4% 38.1% 36.9% 48.9% TotalPixel 2601 CorectPixel OveralAccuracy =1748/ % PA ProducersAccuracy CA(UA) Consumer's(User's)Accuracy EO ErorofOmission=100%-PA EC ErorofCommission=100%-CA 9

10 Producer s s accuracy Probability of a reference pixel being correctly classified. How well has a certain type been classified? Error of omission error of omission (%) = 100 % - Prod s acc. (%) Proportion of observed features on ground that are not classified. User s s accuracy or Consumer s s Accuracy Probability that a pixel classified on the image actually represent that category on the ground. Reliability of map from user s view Error of commission error of commission (%) = 100 % - User s acc. (%) 6.6 Post - Classification Smoothing for Classified Image Noisy appearance - isolated pixels, small groups of pixels Class-homogeneous regions is important Smoothing algorithm minimum area constraint majority filter Import to GIS and Publish as a map Smoothing with minimum area constraint Alters those class-homogeneous regions that have less than specified minimum area to the class that is the majority along the original boundary. The yellow area is only 4 pixels Boundary with A:6, B:5, D:3 -> Replace with A A A A A B A A A A B A C C C B A C B B B D D D D D A A A A B A A B B B D D D D D Smoothing with majority filter If there is a majority class in 3x3 window, replace central pixel with the majority class Import Classification to GIS The result of image classification is in raster format that will be transformed to vector shape file format for data input of a GIS. A A A A C B C A A Class#Pixels A 6 B 1 C 2 Class A is the majority! Replace Central Pixel to A A A A A A B C A A A A B C C B Class#Pixels A 3 B 3 No majority! No Change A A B C C B C B A C 3 C B A 10

11 Import Classification to GIS Map Development Example Habitat type Phi Phi Coral Reef map L and and deep water area Live Coral on Substratum zone Dead Reef zone Reef Dense zone Dead Reef flat zone Reef on Substratum zone Coral Rock zone Substrata S andy zone Substrata Rubble zone Substrata Rock zone Substrata S ilt zone Habitat type map of Phi Phi Don Island. Map Development Example Map Development Example End of Lecture! 11