Digital Image Processing

is remotely sensed imagery valuable by itself

NO

what must happen to remote imagery to answer questions

it needs to be processed

what is the correlation between inputs and results

quality data = quality inputs = quality results

what are steps in digital image processing

1. preprocessing

2. image enhancement

3. image transformation

4. image classification and extraction

what are raw satellite images

images that have no corrections for atmosphere, topography, sensor irregularities and noise

what are the three levels of preprocessed data

level 0 = raw image

level 1 = some radiometric and geometric corrections

level 2 = some atmospheric corrections

what are radiometric corrections

correcting for sensor irregularities, unwanted sensor noise or atmospheric noise

what does preprocessing do

corrects data to represent the reflected or emitted radiation

what does radiometric corrections do

removes the influence of the atmosphere

what are two types of sensor irregularities

1. striping

2. dropped lines

what are three ways to correct atmospheric irregularities

1. ground target collection

2. pseudo invariant calibration sites

3. manufactured ground targets

how are geometric corrections done

linking unknown points with known ground control points

what are image mosaics

composite images consisting of at least two imagaes

what are benefits of image mosaics

1. increase the area covered

2. removes clouds

3. reduces off nadir angles

what must be considered for image mosaic analysis

time

why is image enhancement used

to improve the appearance of imagery to assist with visual interpretation

what are satellites designed to capture data from

a range of targets and conditions

what does image enhancement do

manipulates pixel values to display the optimum brightness rang and contrast for targets in the image

what is contrast enhancement

changing the original values so that more of the available range is used

what does contrast enhancement result in

increase in contrast between features

what is contrast

the difference in luminance or colour that makes object details distinguishable

what is the image histogram

the graphical representation of the brightness values that comprise an image

what is linear contrast stretch

identifies the minimum and maximum values of a histogram and stretches the data to fill the new range

what is histogram equalization stretch

assigns more display values (range) to frequently occurring portions of the histogram

why is histogram equalization stretching not linear

because it applies more values to areas with more points in an image

what is spatial filters

highlight or suppress specific features in an image based on spatial frequency

what is spatial frequency

variation in tone that appears in an image

contrast rough vs smooth spatial frequency

rough - textured area where changes in tone are ABRUPT

smooth - areas with little variation in tone (texture)

what describes high spatial frequency? low spatial frequency?

high = rough texture

low - smooth texture

contrast high vs low spatial frequency

high

• data tends to look sharp and contain more detail

low

• data tends to look smooth or blurry

what type of spatial filtering is used to remove elements from an image

low spatial frequency

contrast low vs high pass filters

low pass - smooths the data by reducing local variation and removing noise

high pass - sharpens the appearance of fine detail in an image

what are high pass filters used for

to highlight boundaries between features

what values does low pass filtering use from the window

mean or median values

what type of pass filtering removes extreme values from the data

low pass filtering

low or high pass filtering

low pass

what type of pass filters remove low frequency variation in cells

high pass filters

low or high pass filtering

high pass

what does edge detection filters do

highlights linear features such as roads, fields or boundaries

how do edge detection filters work

identify points in an image where brightness values change SHARPLY and use these points to create line segements

what is a directional filter

used to highlight features in a specific direction

what are directional filters used for

to detect linear features in geologic structures

what filtering is this

edge detection filtering

what type of filtering is this

directional filtering

what is image transformation

the generation of a new image from two or more sources and highlights a feature of intrest

what does image transformations involve

the combined processing of multiple spectral bands

where can the multiple spectral bands come from for image transformations

a single image OR multitemporal images

what is image subtraction

identifies changes that occur in the time between two images

what does image subtraction require

images to be geometrically registered

what is spectral rationing

ratioing data from two different spectral bands to enhance variation in the spectral reflectance between two spectral ranges

what do ratios reduce

the influence of variation in illumination

what are examples of spectral ratioing

NDVI and NBR

true or false: different multispectral bands are highly correlated and contain similar information

TRUE

what is principal component analysis

reduces redundancy and correlation between bands by compressing as much info as possible to reduce the number of bands

what are components

the new bands that result from statistical procedures (combo of correlated data in the image)

what is the benefit of PCA

maximizes the amount of information from the original data into FEWER bands

what is the relationship between features of interest and features that can be extracted

the middle ground (intersection) of the two will be the features available to work with

what are the three types of image extraction

1. pixel based classification

2. object oriented segmentation

3. deep learning

   1. pixel classification

   2. object detection

   3. picture classification

what does pixel based classification result in

thematic map

what is pixel based classification

feature extraction based on the assumption that different features in an image have different spectral reflectance signatures

what is the preferred method doing pixel based classification

using reflectance (BUT DN values are also possible)

what is required for good pixel based classification

good spectral and radiometric calibration of sensors

contrast spectral and information classes in pixel based classification

spectral - groups of pixels that are uniform with respect to their brightness values in different spectral channels

information - categories of interest that are trying to be identified in the image (crop type, rock type…)

is matching spectral classes to information classes easy

NO

what is unsupervised pixel based classification

the clustering algorithms group pixels based on their properties but the classes DO NOT have labels

does the machine assign labels to the separated classes or the user in unsupervised pixel based classification

USER

what are pros and cons of unsupervised pixel based classification

pros

1. machine determines the different classes

2. simple method of classification

cons

1. unlabeled classes are not always easy to identify

2. results may be over or under classified

3. less accurate in complex landscapes

what is supervised pixel based classification

the analyst identifies features of each type of cover (water, trees, concrete…)

how does supervised pixel based classification work

the analyst identifies training samples for each type of cover and the classification process is used to assign class names to each pixel in the image

what is the output of a supervised pixel based classification

thematic raster

pros and cons of supervised pixel based classification

pros

1. considerable control over inputs

2. effective for distinguishing similar classes

3. high accuracy if data is selected well

cons

1. applied to the pixel level = output is noisy

2. accuracy is dependent on equality of training point s

3. subjective to human bias

what are object oriented approaches

groups pixels into representative polygons and the image is divided into homogeneous regions

how does object oriented approaches contrast from pixel based classifications

the classes are assigned to the ENTIRE object and NOT to individual pixels

can object oriented approaches be used in conjunction with only supervised classification

no - can be used with unsupervised as well

pros and cons of object oriented approaches

pros

1. incorporates shape, texture and contextual information

2. reduces noise that occurs using individual pixels

cons

1. computationally intensive

2. not effective for coarse resolution

3. requires a model

what is deep learning

based on neural networks, originating from research in AI

what is the primary technique for doing Deep learning

Convolutional neural networks (CNNs)

how is Deep learning done

labeled samples train the neural network by exposing the CNN to the samples and reinforcing connections that are correct and weaken those that are incorrect

what is the output of feature learning

raster with class for each pixel

pros and cons of deep learning

pros

1. incorporates spatial and spectral cues

2. pixel classification results are promising

cons

1. requires large number of samples for training

2. requires experienced data scientists

3. vulnerable to over and under training

what is data integration

the combining or merging of data from multiple sources to extract better/more information