Searching for Patterns in Remote Sensing Image Databases Using Neural Networks

The work summarized here was presented at the 1995 IEEE Geoscience And Remote Sensing Symposium in Florence, Italy. This work was supported by a NASA High Performance Computing and Communications grant and also by the Research Institute for Advanced Computer Science (RIACS) at NASA Ames Research Center, Moffet Field, Ca. Don't miss our snazzy graphic illustrating this project. Don't miss our second, brand new snazzy graphic (made Sept. 14, 1995) either!.

The full text can be found in:

++ Paola, J. D., and Schowengerdt, R. A., Searching for Patterns in Remote Sensing Image Databases Using Neural Networks, Proceedings, 15th Annual International Geoscience and Remote Sensing Symposium, Florence, Italy, July 10-14, 1995, pp. 443-445.

Abstract

We have investigated a method, based on a successful neural network multispectral image classification system, of searching for single patterns in remote sensing databases. While defining the pattern to search for and the feature to be used for that search (spectral, spatial, temporal, etc.) is challenging, a more difficult task is selecting competing patterns to train against the desired pattern. Schemes for competing pattern selection, including random selection and human interpreted selection, are discussed in the context of an example detection of dense urban areas in Landsat Thematic Mapper imagery. When applying the search to multiple images, a simple normalization method can alleviate the problem of inconsistent image calibration. Another potential problem, that of highly compressed data, was found to have a minimal effect on the ability to detect the desired pattern. The neural network algorithm has been implemented using the PVM (Parallel Virtual Machine) library and nearly- optimal speedups have been obtained that help alleviate the long process of searching through imagery.

For a very brief visual overview of this research, see this graphic.

Experiment

Neural networks have proven their worth as supervised multispectral classifiers in many previous experiments. With the advent of EOS and other remote sensing platforms, a major challenge in the near future will be the task of searching large remote sensing image databases for patterns of interest in particular applications. These patterns might be spectral, spatial, temporal, or any combination thereof.

In the example explored here, a neural network is used to search some sample imagery for dense urban (i.e., downtown) areas. A 3x3 window was used in each of the six non-thermal Landsat Thematic Mapper (TM) bands as input to the neural network. This example is essentially a spectral problem, with the 3x3 windows providing a measure of texture. The training image is a scene of Tucson, Arizona acquired on April 1, 1987.

Neural network configuration for single pattern searching:

Competing Pattern Selection

To train the neural network we need not only samples of the desired search pattern, but also samples of patterns to exclude. Our goal was to make this aspect of the procedure as simple as possible to the user without sacrificing too much accuracy.

Four strategies were used for selecting the "competing patterns":

Use the training data for the other classes of a multi-class classification
Generate uniformly distributed random values for the image bands
Use a grid of values from the training image
Combine methods 2 and 3 - gridded image samples supplemented by generated values

Results on the Training Image

The original Tucson image:

A 12 class multispectral image classification (the dense urban class is light blue):

For all search images the raw network output is displayed, with darker values representing more likely matches to the dense urban training pattern.

Competing pattern selection Method 1 yields the following search image:

Method 2 disappoints us with this result:

The feature space is simply too large to be adequately covered by independent uniformly generated random patterns. Perhaps the use of correlated RV's could make this method more appealing.

Method 3 and method 4 provide nearly identical search results.

Results on Other Imagery

Clearly, the last two methods are the best. The urban areas of the Tucson image are highlighted nicely while the surrounding areas are suppressed in the continuous-valued neural network output. The next step is to apply the network as trained above on other imagery.

The first example is a TM scene of Oakland, Ca. from August 12, 1983:

Method 3 produces the following map, , which is notable for one major flaw, the water of the San Francisco Bay is incorrectly identified as "dense urban". Since the competing patterns were selected from the Tucson image, and this image had no areas resembling this very different kind of land cover, the network has not excluded it from the identification.

The network trained with method 4, however, produces this highly accurate result:

Perhaps the best training method is to use gridded patterns from not just the image(s) used for training but from completely different images as well.

The Oakland image was not calibrated (beyond being a Level 1 Landsat product) to the Tucson image in any way. The network was able to overcome the variations brought about by the atmosphere, sun angle, and changing sensor characteristics in this case. For the next image, of Washington, D.C., however, some simple calibration was necessary. The D.C. image, , was much darker than the Tucson image. A simple calibration was done in which the mean and variance of each band were adjusted to match those of the Tucson image.

As with the successful Oakland image search, the network trained using "competing pattern" method 4 was applied to the uncalibrated and calibrated Washington, D.C. images.

Before calibration, no urban areas were detected:

After calibration, we obtain a very nice result:

Conclusion

A general pattern matching algorithm is not expected to achieve consistently high accuracy for all the varied imagery used in remote sensing applications. Fortunately, for this application, the goal is not to achieve the highest possible accuracy, but to provide a good estimate of candidate matches that can be used to guide further investigation. The flexibility of the neural network allows for adaptation to many different types of imagery and pattern signatures, while providing moderate accuracy in pattern matching. In addition to the dense urban area detection discussed here, we have attempted other searches, including more subtle TM classes such as 'grassland' and 'pine-oak woodland', as well as other land-cover classes using temporal NDVI data. These patterns are more difficult to detect, particularly in imagery not used for training. This result stresses the need for a well-calibrated dataset of global imagery in the EOS era in order to achieve widely applicable content-based browsing of the type investigated here.

This page created August 2, 1995 by Justin D. Paola

Return to DIAL

home page.

Digital Image Analysis Laboratory
Department of Electrical and Computer Engineering,
University of Arizona,
Tucson, Az 85721
(520) 621-4554