7 Image Classification Flashcards
Image classification 1/2
–One of the digital image interpretation techniques
•Image visualization and image indices are two other techniques
–Converts Remote Sensing images to thematic maps
• Land cover, land use, soil type etc. can be used for further analysis or as input into models.
–converts n number of bands into a single band image.
•Data reduction
–Each pixel in the image is labeled with a class
Image classification 2/2
- Image classification is based on the different spectral characteristics of different materials on the Earth’s surface
- Image classification is operated in feature space (band space)
- Let’s assume each ground resolution cell contains a uniform land cover
- Some classes form compact clusters, while others form dispersed clusters
- Basic assumption for image classification: a specific part of the feature space corresponds to a specific class
Principle of image classification
•A pixel is assigned to a class based on its feature vector
•In image classification –The classes are defined in the feature space –Each feature vector is assigned to a class where it fits best.
- Some classes occupy their own area in the feature space,
- while others are overlapped
- It is important to select bands in which different classes are well distinguished
Two types of image classification
•Supervised classification
–Using training data
–Interaction with the operator
–clusters are defined by the user during the training process
•Unsupervised classification
–No training data needed
–Clusters are automatically defined by the algorithm
Process of image classifications
Major steps
–Selection and preparation of RS data
•Depending on what needs to be classified, the most appropriate sensor, dates of acquisition, and the most appropriate bands are selected
–Partitioning of feature space
•Supervised classification
•Unsupervised classification
–Selection of classification algorithm
•How the pixels are assigned to the classes
–Running the actual classification
•Assign each pixel in the image to a class
–Validation of the results
•The quality of classification result is assessed by comparing it to a reference data
Major steps
Selection and preparation of RS data 1/2
–For particular applications, images need to be acquired by specific sensors at specific dates e.g. classification of snow cover or certain types of land cover, and classification of flood or landslides.
–Cloud cover should be considered in data selection
–In some applications, illumination of sun should be taken into account.
Major steps
Selection and preparation of RS data 2/2
–For a selected image, a selection of specific bands may be made.
–Some bands might be too correlated (e.g. red and green bands for vegetation)
–Redundant information might disturb the classification process.
–In hyperspectral images neighboring bands are often too correlated.
–It is particularly important to select bands in which different classes of interest are well distinguished
–e.g . select n first components of PCA (in hyper-spectral image, it’s common to apply PCA transformation, and then select n first number of band. e.g. choose first 10 bands which contain the most information and not much correlated. -> this is how we avoid redundant infos)
Major steps
Partitioning
- Partitioning the feature space is an important step in the classification
- In supervised classification, the operator identifies sample areas (training data) to define the spectral characteristics of classes.
•The operator needs to find pixels representing classes of interest in the image
–From general knowledge of the scene
–Other sources of assumed higher accuracy
–From field observations
Major steps
Selecting training data
•Training data selected by the operator should form representative data for each class.
–A minimum number of observations per cluster is required (30x n where n is the number of bands
–Variability of class within the image should be taken into account.
–Each cluster should not overlap with other clusters, otherwise it is not possible to reliably separate different classes
– If selected classes are significantly overlapped, adding spectral bands or images might help.
Classification algorithms
• Classification algorithm is defined based on the purpose of classification, characteristics of the image, and training data
•Classifiers –Nearest neighbor –Box (Parallelepiped –Minimum Distance to Mean –Mahalanobis distance –Maximum Likelihood
•It is decided if an ‘Unknown’ class is allowed.
Classification algorithms
Nearest neighbor classifier
–One of the simplest classification methods
–Distance to all training vectors are estimated
–The candidate pixel is assigned to the class with nearest sample
Classification algorithms
Box classifier (Parallelepiped)
–One of the simplest classification methods
–A number of boxes are defined based on the number of classes
–Upper and lower limits of each class are defined based on
•Minimum/maximum values
•Mean and standard deviation
–Candidates that do not fall inside any box will be assigned to ‘unknown class’
Classification algorithms
Box classifier (Parallelepiped) Advantage and Disadvantage
–Advantage of Box classifier
•Very simple
–Disadvantage of Box classifier
•Overlap between the boxes
Classification algorithms
Minimum distance classifier
–Cluster centers are calculated
–The Euclidean distances from each candidate vector to all cluster centers are calculated
–The candidate pixel is assigned to the closest cluster center
Euclidian distance between vectors x and y:
d = [ (x-y) * (x-y)^T ]^0.5
Classification algorithms
Minimum distance classifier Disadvantage
- Points might be assigned to cluster centers located at large distances.
- Usually a threshold is defined to avoid this problem.
- It does not take into account the class variability.
- Small/large and dense/disperse clusters are treated equally.
