Без темы
<<  Enquiry-Based Learning (EBL) in Humanities at Manchester Equation Editor Tips  >>
Multispectral image classification
Multispectral image classification
i.e. separate in feature space
i.e. separate in feature space
i.e. separate in feature space
i.e. separate in feature space
Supervised classification
Supervised classification
Supervised classification
Supervised classification
Supervised classification: MDM
Supervised classification: MDM
Supervised classification: parallelepiped (‘box’)
Supervised classification: parallelepiped (‘box’)
Supervised classification: parallelepiped (‘box’)
Supervised classification: parallelepiped (‘box’)
Supervised classification: Gaussian maximum likelihood
Supervised classification: Gaussian maximum likelihood
Supervised classification: Gaussian maximum likelihood
Supervised classification: Gaussian maximum likelihood
Classification Accuracy
Classification Accuracy
Postclassification filtering
Postclassification filtering
Postclassification filtering
Postclassification filtering
Картинки из презентации «Environmental Remote Sensing GEOG 2021» к уроку английского языка на тему «Без темы»

Автор: Simon Brown. Чтобы познакомиться с картинкой полного размера, нажмите на её эскиз. Чтобы можно было использовать все картинки для урока английского языка, скачайте бесплатно презентацию «Environmental Remote Sensing GEOG 2021.ppt» со всеми картинками в zip-архиве размером 1644 КБ.

Environmental Remote Sensing GEOG 2021

содержание презентации «Environmental Remote Sensing GEOG 2021.ppt»
Сл Текст Сл Текст
1Environmental Remote Sensing GEOG 17maximum likelihood. Now we use probability
2021. Lecture 4 Image classification. rather than distance in feature space
2Purpose. categorising data data Which class is each pixel “most likely” to
abstraction / simplification data belong to?? 17.
interpretation mapping for land cover 18Supervised classification: Gaussian
mapping use land cover class as a maximum likelihood. Now pixel 1 correctly
surrogate for other information of assigned to corn class Much more
interest (ie assign relevant sophisticated BUT is computationally
information/characteristics to a land expensive compared to distance methods.
cover class). 2. 18.
3Multispectral image classification. 19Supervised classification: decision
Very widely used method of extracting tree. Classify in steps, where the
thematic information Use multispectral classifier has only to be able to
(and other) information Separate different distinguish between two or more classes at
land cover classes based on spectral each step can combine various types of
response, texture, …. i.e. separability in classifiers as appropriate using such
“feature space”. 3. methods. 19.
4Basis for 'classifying'. method: 20Classification Accuracy. How do we
pattern recognition use any/all of the tell if classification is any good?
following properties in an image to Classification error matrix (aka confusion
differentiate between land cover classes:- matrix or contingency table) Need “truth”
spectral spatial temporal directional data – sample pixels of known classes How
[time / distance-resolved (LIDAR)]. 4. many pixels of KNOWN class X are
5Spatial pattern recognition use incorrectly classified as anything other
spatial context to distinguish between than X (errors of omission)? So-called
different classes e.g. measures of image Type 2 error, or false negative Divide
texture, spatial context of 'objects ' correctly classified pixels in each class
derived from data. Temporal pattern of truth data by COLUMN totals (Producer’s
recognition the ability to distinguish Accuracy) How many pixels are incorrectly
based on spectral or spatial classified as class X when they should be
considerations may vary over the year use some other known class (errors of
variations in image DN (or derived data) commission)? So-called Type 1 error, or
over time to distinguish between different false positive Divide correctly classified
cover types e.g. variations in VI over pixels in each class by ROW totals (User’s
agricultural crops. 5. Accuracy). 20.
6Directional pattern recognition 21Classification Accuracy. 21.
surface with different structures will 22Can use original training data to test
tend to give different trends in BUT…. …this only tells us how well the
reflectance as a function of view and classifier can classify the training areas
illumination angles Spectral pattern Ideally, use an independent set of samples
recognition most widely used distinguish to give a better 'overall' accuracy
between different land cover classes from estimate. 22.
differences in the spectral reflectance 23Unsupervised Classification
(or more typically, image DN) in different (clustering). Little input from user
wavebands. 6. required (few assumptions) BUT means
7i.e. separate in feature space. Use results hard to interpret (may not
different spectral response of different represent classes we recognise) cluster
materials to separate e.g. plot red v NIR pixels in feature space based on some
DN values…. 7. measure of their proximity interpretation
8Approaches to Classification. We need of results / assigned classes can be
some form of automated (rule-based) useful, e.g. in picking up variations
classification algorithm to allow us to within what would otherwise be
distinguish one surface type from another distinguished as a single class e.g.
Supervised Classification Unsupervised stressed/unstressed crop in a single
Classification. 8. field) clusters can be of little intrinsic
9Supervised classification. training value in themselves e.g. sunlit trees,
stage (significant user input/expertise) shaded trees is perhaps not a useful
Identify areas of cover types of interest discrimination if one simply wants to
(map, ground survey, spectral classify 'trees', and so clusters may have
characteristics) in bands of an image. to be combined. 23.
From Lillesand, Kiefer and Chipman (2004). 24Unsupervised Classification: K-means.
9. A large number of clustering algorithms
10Supervised classification: training exist K-means input number of clusters
stage. areas of interest delineated by desired algorithm typically initiated with
user spectral information on the cover arbitrarily-located 'seeds' for cluster
types is gathered for these areas Training means each pixel then assigned to closest
data (subset of whole) These are “classes” cluster mean revised mean vectors are then
we will place all remaining pixels in computed for each cluster repeat until
according to their DN values Can plot in some convergence criterion is met (e.g.
feature space – do we see clusters? 10. cluster means don't move between
11Supervised classification: iterations) computationally-expensive
classification stage. Need rule(s) to because it is iterative. 24.
decide into which class we put given pixel 25Unsupervised classification: ISODATA
e.g. Minimum distance to means (MDM) for (Iterative self-organising data analysis)
each land cover class, calculate the mean algorithm. Same as K-means but now we can
vector in feature space (i.e. the mean vary number of clusters (by splitting /
value in each waveband) Put every pixel merging) Start with (user-defined number)
into nearest class/cluster define a limit randomly located clusters Assign each
beyond which a pixel remains unclassified pixel to nearest cluster (mean spectral
a simple and fast technique but has major distance) Re-calculate cluster means and
limitations… 11. standard deviations If distance between
12Supervised classification. Feature two clusters < some threshold, merge
space clusters E.g. 2 channels of them If standard deviation in any one
information Are all clusters separate? 12. dimension > some threshold, split into
13Supervised classification: MDM. Find two clusters Delete clusters with small
closest cluster mean for each pixel Simple number of pixels Re-assign pixels,
and quick BUT what about points 1, 2? i.e. re-calculate cluster statistics etc. until
MDM insensitive to variance of clusters changes of clusters < some fixed
Can we improve? 13. threshold. 25.
14Supervised classification: 26ISODATA example: 2 classes, 2 bands.
parallelepiped (‘box’). Assign boundaries DN Ch 2. 26.
around the spread of a class in feature 27Hybrid Approaches. useful if large
space i.e. take account of variance variability in the DN of individual
typically use minimum/maximum of DN in a classes use clustering concepts from
particular class to define limits, giving unsupervised classification to derive
a rectangle in 2D, box in 3D (if we have sub-classes for individual classes,
> 2 bands) etc. pixels outside of these followed by standard supervised methods.
regions are unclassified (which is good or can apply e.g. K-means algorithm to (test)
bad, depending on what you want!!) subareas, to derive class statistics and
problems if class regions overlap or if use the derived clusters to classify the
high covariance between different bands whole scene requirement that all classes
(rectangular box shape inappropriate) can of interest are represented in these test
modify algorithm by using stepped areas clustering algorithms may not always
boundaries with a series of rectangles to determine all relevant classes in an image
partially overcome such problems simple e.g. linear features (roads etc.) may not
and fast technique takes some account of be picked-up by the textural methods
variations in the variance of each class. described above. 27.
14. 28Postclassification filtering. The
15Supervised classification: result of a classification from RS data
parallelepiped (‘box’). Simple boxes can often appear rather 'noisy' Can we
defined by min/max limits of each training aggregate information in some way?
class. But overlaps……..? …so use stepped Simplest & most common way is majority
boxes. 15. filtering a kernel is passed over the
16Supervised classification: Gaussian classification result and the class which
maximum likelihood. assumes data in a occurs most commonly in the kernel is used
class are (unimodal) Gaussian (normal) May not always be appropriate; the
distributed class then defined through a particular method for spatial aggregation
mean vector and covariance matrix of categorical data of this sort depends
calculate the probability of a pixel on the particular application to which the
belonging to any class using probability data are to be put e.g. successive
density functions defined from this aggregations will typically lose scattered
information we can represent this as data of a certain class, but keep
equiprobability contours & assign a tightly-clustered data. 28.
pixel to the class for which it has the 29Postclassification filtering. Majority
highest probability of belonging to. 16. filter. 29.
17Supervised classification: Gaussian
Environmental Remote Sensing GEOG 2021.ppt
cсылка на страницу

Environmental Remote Sensing GEOG 2021

другие презентации на тему «Environmental Remote Sensing GEOG 2021»

«Времена в английском языке» - Указание момента времени. The past perfect tense. Modal verbs. Чувства. The present perfect continuous tense. The future Indefinite Tense. Употребление и перевод. London. Употребление. Английский язык. The past continuous tense. The present continuous tense. The past Indefinite tense. Отрицательная форма глаголов.

«Jimmy Carter» - To find out what contribution Jimmy Carter made ??to the development of America. 1977 - Panama Canal Treaty. Conclusion. Jimmy Carter. Major Events during his presidency. Career. Presidential facts. Sources used. Born on October 1, 1924 Plains, Georgia. Education. Contens. Aim. Soldier, Peanut Farmer, Professor, Politician, Statesman 39th president.

«Easy English» - «Формирование элементарных навыков общения на иностранном языке детей в системе «Детский сад - начальная школа». Сквозная программа по раннему обучению английского языка детей в детском саду и первом классе. Учебно-методический комплект «Easy English». Эксперимент по проблеме. Учимся говорить по-английски в детском саду и первом классе начальной школы.

«Chinese New Year» - The Lunar New year. The Origin Of Chinese New Year. Taboos. All about Chinese New Year. Big dinner. The Dragon And The Lion Dance. Chinese red envelopes. The First day of the Lunar New year. Chinese Lunar Calendar. Red-paper couplets. Very busy during. Couplets.

«День Валентина» - После смерти Валентина стали считать святым покровителем влюбленных. Сладкий. Почитаем стихи, посвященные Дню святого Валентина. Стихи, посвященные празднику. Познакомимся с историей празднования Дня святого Валентина. Проверим свои знания. Word List. Greetings. It’s nice to have a friend like you! Задание «Найди слова».

«Английский язык в 1 классе» - Особенности процесса обучения английскому языку с 1 класса. Спасибо за внимание. Технология обучения иностранному языку. Актуальность темы. Обучение говорению. Обучение английскому языку в первом классе. Обучение произношению. Умения учителя иностранного языка. Обучение аудированию. Потребности учащихся 1 класса.

Без темы

661 презентация

Английский язык

29 тем