Lecture 9 land cover maps Flashcards
Why produce land cover maps
Land cover interacts strongly with the water cycle, carbon cycle and the climate system (e.g.) :
▪ flood and erosion modelling
▪ climate change (desertification, greening)
▪ food security (crop mapping, yield estimates)
▪ mapping environmental change e.g. deforestation, ice loss, biodiversity, anthropogenic impacts
What is image classification?
A procedure that use the remotely sensed image data to produce maps showing the location and extent of various selected land cover types or Earth surface features . Approach of going from an image - convert into another image, each colour has a different value, digital number into landcover type for each pixel.
why might landcover types be misclassified
If they have similar spectral properties – landcover types may be misclassified. may also use temporal information too.
Assumption when classify, groupings of pixels with similar values are more likely to be same landcover types. For example, forest is more likely to have similar reflectance, desert, pixels likely to have similar spectral values as sand.
What is A feature space plot?
Those graphs with the colours we looked at in practical’s.
A feature space plot is simply a scatter plot between different wavebands – most often 2D (x,y).
The colour indicates the number of pixels (i.e., the frequency) that have particular values – brighter areas indicate higher frequency
assumption, pixels with low values in both wavebands likely to cluster likely to be from same landcover type.
how to decide on group clusters
Use the Euclidian distance between clusters to group points to clusters. BC distance is smaller than AC distance, Therefore C should be grouped with cluster B
Advantages of the Digital Classification Over the Visual Interpretation
▪ Cost efficient in the analyses of large data sets ▪ Long time-series or regional scales
▪ Results can be reproduced computer = reproducible▪ Manual ‘drawing’ approach will differ slightly each time
▪ More objective than visual interpretation ▪ Based on numerical values – less subjective
▪ Effective analysis of complex multi-band (spectral) interrelationships ▪ Makes use of all the available data.
what are object based Image classification
▪ Utilise spatial information to segment imagery (e.g. texture (building with a flat roof/pointed roof), geometry)
▪ Best suited to high spatial resolution imagery – 2/3m or less spatial resolution
give two common spectral classification methods
▪ Unsupervised
▪ Supervised
What is unsupervised classification
Classification method which does not the compare pixels to be classified with any prior information. Rather, it examines a large number of unknown data points and divided them into classes based on properties of the data.
* Based on spectral groupings/clustering of data
* Considers only spectral distance measures
* Minimum user interaction
* Requires interpretation after classification (i.e., identification of what the classes represent – forest, water etc)
* User specifies number of land cover classes to find.
Doesn’t need user input, based on assumption that pixels close together have similar spectral reflectance using Euclidian distance to assign cluster. Once have image – colour automatically assigned to cluster – must decide which landcover type each cluster represents i.e., interpret the image and the result.
What are the two main methods of unsupervised classification
Two main methods 1. ISODATA 2. K-Means
Describe k means clustering
- Define of number of clusters to be generated
- Seeding (randomly) cluster mean vectors in feature space. then calculates summary statistics - mean value falls, forms central point of cluster.
- Allocates all pixels to one of the clusters based on distances (Euclidian distance). then calculates summary statistics - mean value falls, forms central point of cluster.
- Update the cluster mean on basis of assigned pixels
- New cluster means used in second allocation
- Repeat until some convergence criterion is met/ no more changes to image.
What does ISODATA mean
Iterative self-organizing data analysis technique (Algotithm)
How is ISODATA different from k-means
same, but can vary the number of clusters by splitting and merging. so if one cluster should really be two clusters they can be split, using standard deviation
What does the user have to do in unsupervised classification
determine what land cover type each class in the image refers to (i.e. black pixels are roads, green are trees etc)
What are the advantages of unsupervised classification?
- No extensive prior knowledge of the area required
- Opportunities of human error are minimised i.e. mostly relates to collecting training data in supervised classification
- Unique classes can be identified such as different tree species
