About ICREST

Steering Committee - Dr. Andrew Blanchard, Dept. of Electrical Engineering, Dr. Xinhua Zhuang, Dept. of Computer Engineering and Computer Science, Dr. K. Palaniappan, Dept. of Computer Engineering and Computer Science, Dr. Curt Davis, Dept. of Electrical Engineering, Dr. James D. Hipple, Dept. of Geography, Dr. Aderbal C. Correa, Dept. of CEE, Dr. Dave Diamond, Director MoRAP
Datasets - primarily for the state of Missouri
Survey - state and local government use of remote sensing datasets

Automatic BTGA Land Cover Classification

Goal:

To develop an automatic landcover classification algorithm for the state of Missouri using multispectral Landsat TM data and additional information collected by MoRAP. Objectives include:

Reduce amount of manual supervision in the classification procedure (ideally none).
Reduce the total time needed to provide a statewide classification.
Develop tools for high performance visualization and compression of large remote sensing datasets.
Avoid or minimize overlapping scene clustering.
Determine generalizability of approach beyond the state of Missouri.
Determine accuracy versus performance versus cost tradeoffs.

Deliverables:

Develop automatic landcover analysis algorithm using tree classifiers, mixture density class probability modeling, robust relaxation algorithms, and hierarchical classification methods.
Develop analysis and visualization tools:

Progressive image (SLCCA) and video (VSLCCA) compression
Volumetric compression of multispectral data
Portable OpenGL version of DISS
Terrain geometry compression and landcover visualization using FlyBy

Automatic clustering result for Landsat scenes covering a single ecological region
Comparing performance of automatic algorithm with MoRAP?s final classification result evaluated in terms of landcover categories and ecological regions
Quantify performance (i.e., confusion matrices) with or without ground truth data
Evaluate algorithm performance on simulated 1m data using aerial photo

Personnel:

Joshua Fraser: DISS visualization tool development for manipulating remote sensing landcover data. Missouri DEM data access and creation. Algorithm development for remote sensing data analysis.
Feng Zhu: Algorithm development for remote sensing landcover data analysis.
Jozsef Vass: Image, video, and volume data compression using wavelets. Algorithm development for remote sensing landcover data analysis.
Jia Yao: Progressive transmission of image data. Content-based searching of image databases.
Scott Musler: Geometry compression of DEM data. Visualization and rendering algorithms.

Recent Progress:

4/30/99: Acquired two scenes of original Landsat data, phase 1 and phase 2 intermediate steps, and ground verification points from MoRAP.
5/10/99: Meeting to discuss Gaussian density approach for training. Delegated work for preparing datasets and coding the algorithm.
5/11/-5/17/99: Learned format of ERDAS GIS files (.lan, .sta, .pro). Read scenes acquired from MoRAP. Got phase 1 cluster label meanings.
5/11/99: Hired Feng Zhu.
5/24/99: Acquired MoRAP ground truth observations in Access format. Feng and Joshua met with Kan He to further clarify the MoRAP landcover and landuse algorithm.
6/1/99: NASA Stennis Space Center site visit
11/4/99: Feng completed implementation of Bayesian BTGA and begin testing on the program with data from scene 33.

Conferences, Meetings and Presentations:

April 19-20, 1999, attended by Joshua Fraser. GPS training using Z-Surveyor equipment.
May 13-17, 1999, attended by Joshua Fraser. NASA presentation pertaining to visualization of Earth and Space Science High Resolution Imagery and HDTV for Space Week at James Lovell Discovery World Museum, Milwaukee, WI.
June 21-25, 1999, attended by K. Palaniappan and Joshua Fraser. "Construction and Visualization Techniques for Weather and Climate Models", NSF Geospace Environment Modeling Workshop (GEM?99), Snowmass, CO, Dr. John Freeman, Rice University.
July 12-16, 1999, attended by X. Zhuang. MPEG-4 Workshop, Vancouver, BC.

Summary

Landsat TM data with Phase II landcover classification data (four scenes total) acquired and preprocessed for automatic classification algorithm training and testing
BTGA implementation capable of using 12 features and classifying 47 classes completed and in the process of being tested and improved
Piecewise linear classification boundaries organized in a binary tree structure with weights learned using a genetic algorithm
Two scenes (one spatial region, two temporal observations) with 33.7million observations used to design and test classifier
Entire Missouri dataset (overlapping scenes) has about 15 times more data
Classification accuracy of 72.7% achieved using 4.6% of the data for training, 12 features, 8 landcover information categories (with 47 distinct classes)
Classification accuracy increases as amount of training data increases
Classifier training time increases linearly with the number of observations and is becoming prohibitive - more than 10 hours using 4.1% (1.36 million) of the data
Size of the classification tree increases as amount of training data increases (classifier adaptation and complexity with more data)

Future Work

Improving classification accuracy using non-linear classifiers, hierarchical classification, handling reject regions, expanded feature set (such as image texture features), image space segmentation, feature space clustering.
Use the Bayesian minimum error classifier for baseline comparisons
Expanding the size of the training data (require additional computational capabilities).
Testing the classifier on the entire Missouri dataset.
Testing the classifier using ecological region information.

Satellite Imagery and Landcover Data Characteristics

Landsat TM imagery with MORAP Phase II classification data two regions in Spring 1992 (05/03/92) and Fall 1992 (09/24/92) acquired and preprocessed for algorithm development and testing.

Each of the four scenes are 7226 x 6524 pixels in image size, with 30 m x 30 m ground resolution and have 6 TM spectral bands which are used as the feature set. These regions are in the vicinity of Columbia, Missouri. Approximately fifteen overlapping TM scenes are required to cover the entire state of Missouri with the mosaic image size being 19,398 x 17,004 pixels.

Testing and training has been done using a single scene which contains 33,784,355 Phase II classified pixels. The ith labelled sample test data is a row vector: (f1,i, f2,i, ...., f12,i, ci)