Land Cover / Land Use Classification

University of Antwerp

Remote Sensing Courses

Land Cover / Land Use Classification Seminar Practicum IV Vegetation classification using LANDSAT 7 ETM+ imagery Tuesday, May 7th, 2015

Frank Veroustraete and Roeland Samson

6 May 2015

1



Contents

I. Short introduction into the Problem area II. Classification algorithms

III. Exercises

6 May 2015

2



Introduction into the problem area A few Key Points Knowledge about Land Use and Land Cover and especially their changes (LUCC) has become increasingly important to try solving problems like : 1. UNCONTROLLED URBAN DEVELOPMENT 2. DEGRADING ENVIRONMENTAL QUALITY 3. LOSS OF PRIME AGRICULTURAL LAND 4. DESTRUCTION OF (IMPORTANT) WETLANDS 5. LOSS OF FISH AND WILDLIFE HABITATS 6. °°° Excerpt from Anderson, et. al. (1976). A Land Use And Land Cover Classification System for use with Remote Sensing Data.

6 May 2015

3



Introduction into the problem area A few key points - Land Use versus Land Cover

6 May 2015

•

Land Use refers to human activity in landscapes:

•

Land Cover refers to Landscape properties:

•

Agriculture

•

Crops

•

Commerce

•

Water

•

Settlement

•

Forest

•

Recreation

•

Buildings

4



Introduction into the problem area A few key points - Standardisation in LUCC determination

6 May 2015



Increased use of Land Use /Land Cover Databases in the 1960’s, led to an increased requirement to legislate for governmental administrations and planners to obtain the most recent standardized and accurate LUC data.



LUC data were often not shared with other government and planning bodies.



Collected LUC data were often too specific for a project and not usefull for other projects or at later dates.



Different LUC classification algorithms, made it difficult to utilize and compare land use maps from different studies.

5



Introduction into the problem area A few key points - Standardisation in LUCC determination Problems in Land Use /Land Cover Standardization

6 May 2015



Incomplete global data coverage at high resolution (for pixels less than 50 m);



Different definitions of LUC classes in map legends;



Incompatible (non comparable) LUC classes in map legends.



Different classification methods applied by different administrations;



Variable Classification Dataset production dates. 6



Introduction into the problem area A few key points - Interpretation

6 May 2015



The minimal area which can be classified for a particular LUC class is dependent on the scale and resolution of the remote sensing satellite/sensor duo.



LUC classification requires validation based on the interpretation of a multitude of attributes of the imagery LUC units: colour, texture, shadow, patterns, associations, shapes, size, ...



Other data, such as topographic maps, road maps, and field studies are applicable as a reference, when their scale is small enough with respect to the resolution of the sensor imagery.

7



Introduction into the problem area A few key points - Spatial resolution LUC Levels are dependent on the Spatial Resolution of the imagery. Level  Level I:  Level II:  Level III:  Level IV:

6 May 2015

Resolution (m) 80.00 2.50 0.90 0.45

8



Introduction into the problem area A few key points - Typical LUC Classes at Level I Level I: Global to Continental scale

1 Urban or Built-up Land 2 Cropland 3 Rangeland 4 Forest Land 5 Water Bodies 6 Wetland 7 Barren Land 8 Tundra 9 Perennial Snow or Ice ...

6 May 2015

9



Introduction into the problem area A few key points - Typical LUC Classification Levels at Level II Level II: Biomes at Regional scale Examples: Agriculture • Croplands and Pastures • Orchards, Groves, Vineyards • Confined Feeding Land • Other Agricultural Land

Examples: Water • Streams • Lakes • Reservoirs • Bays, Estuaries and Mangroves

6 May 2015

10



Introduction into the problem area A few key points - LUC Classses at Level II – An example map Level II: Biome/Region Corine Land Use Map - Resolution 250x250 m - Classification with MODIS imagery.

1. 2. 3. 4. 5.

6 May 2015

Urban (Red) Deciduous forest (light green) Needle Forest (dark green) Agriculture (pale yellow) Inland Water body’s (blue)

11



Introduction into the problem area A few key points - LUC Single Class Probability Classification Level II: Biome/Region Forest Map of European Forestry Institute (EFI)

- Resolution 1x1 km - Classification based on NOAA-AVHRR data

6 May 2015

12



Introduction into the problem area A few key points - LUC Classification Level III

•

Level III: Biome/Ecosystem

Example: Boreal biome • Boreal forest • Northern Hardwood forest • Grassland • Deciduous forest

6 May 2015

13



Introduction into the problem area A few key points - LUC Level IV

6 May 2015

14



Introduction into the problem area A few key points - LUC Level V

6 May 2015

15



Introduction into the problem area A few key points - LUC Level VI

6 May 2015

16



Classification Algorithms Typical Land Use Classification method - “k-means clustering” 1. The k-means algorithm is an algorithm to cluster n objects based on their attributes into k classes (with k < n). 2. It is similar to the expectation-maximization algorithm for mixtures of Gaussians in that oth attempt to find the centers of natural clusters in a dataset. 3. The assumption is made that object attributes form a vector space. 4. The objective it tries to achieve is to minimize total intra-cluster variance, or, the Squared Error Function V below:

where there are k clusters Si ( ) and, where μi is the centroid or mean point of all data points (See next page)

6 May 2015

17



Classification Algorithms Typical Land Use Classification method - “k-means clustering”

6 May 2015

18



Classification Algorithms Typical Land Use Classification method - “k-means clustering” Common applications of k-means clustering

6 May 2015

1.

Mostly used as an exploratory LUC data analysis tool;

2.

In one-dimension, a good way to classify real-value variables into k non-uniform (Boolean) collections;

3.

Also used to classify acoustic data in speech recognition to convert waveforms into k categories - known as Vector Quantization;

4.

Used to attribute colour palettes on Graphical Display Devices.

19



Classification Algorithms Typical Land Use Classification method - “k-means clustering” The most common formalisation of the algorithm uses iterative refinement known as the heuristic Lloyd’s algorithm.

6 May 2015

1.

Lloyd's algorithm starts by partitioning the input points into k initial pointsets, either at random or using some heuristic algorithm.

2.

It then calculates the mean point, or centroïd of each point set. It constructs a new cluster partition by associating each point with the closest centroïd.

3.

Then the centroids are recalculated for the new clusters. The algorithm is repeatedly applied with steps 1 and 2 described until convergence is obtained e.g., when the points no longer switch clusters (or alternatively when the centroids do not longer change).

4.

Lloyd's algorithm and k-means are often used as synonyms. In reality Lloyd's algorithm is a heuristic for solving the k-means problem, with certain combinations of starting points and centroids.

20



Classification Algorithms Typical Land Use Classification – Maximum Likelihood Classification

It is one of the most powerful LUC classifiers in use! 1.

Based on statistics (mean; variance/covariance), a (Bayesian) Probability Function is calculated from the inputs of classes selected in training sites.

2.

An operator judges each pixel is then to be attributed to a class to which it most probably belongs.

• An example of the Morro Bay TM data is shown , using first the six reflected radiation bands and in a second classification with the longer wavelength thermal band added. • The result is a pair of believable classification maps whose patterns (the classes) seem to closely depict the reality of vegetation units. • Keep in mind that several classes are not always normal components of the actual ground scene, e.g., cloud shadows. A classification using a smaller number of classes gives a somewhat different end product.

6 May 2015

21



Classification Algorithms Typical Land Use Classification – Maximum Likelihood Classification In the computation of reflectance distributions, determining the means and covariances for the classes that fall within the training sites, a statistical (Bayesian) Probability Function is calculated for the class data. Class reflectance distribution, for a two-dimensional case, gives rise to elliptical boundaries which define the equiprobability envelope for each (vegetation) class. A scatter plot shows the outer envelope bound for each class. Band 2 represents NIR. Band 3, Red. 6 May 2015

22



Classification Algorithms Typical Land Use Classification – Maximum Likelihood Classification • A Supervised

Classification (by an operator) here at right for Morro Bay in California, uses the Maximum Likelihood Classifier acting on all seven Landsat ETM+ bands. • Multi-band classes are derived statistically and each unknown pixel is assigned to a class using the Maximum Likelihood Classifier .

6 May 2015

23



Classification Algorithms SRFs of the bands in the VIS/NIR and SWIR for LANDSAT ETM+ Bands Landsat ET M+ 1 0.9 0.8 0.7

B1 B2 B3 B4 B5 B6

SRF

0.6 0.5 0.4 0.3 0.2 0.1 0 410

610

862 1062 1262 1462 1662 1862 2062 2263 wavelength

6 May 2015

24



Classification Exercises 1) Start the ENVI software package.

2) Load raster image p199r024_7t20010523_z31_nn6bands.tif. The image is from Western Belgium (the coast to Antwerp).

3) Make a true colour composite with bands 3, 2 and 1. 4) Make a NDVI image with bands 4 and 3. Attribute a colour

table. Perform level slicing. 5) Perform an unsupervised 10 classes k-means clustering

classification of Landsat7, using all bands of the image. 6) Perform a supervised Maximum Likelihood Classification of

Landsat7, using all bands of the image. 7) Define the training set.

6 May 2015

25


Thanks

6 May 2015


for your attention

26