Image analysis – Learning systems – Trainable classifiers or pattern recognizers
Reexamination Certificate
2000-01-24
2001-08-21
Boudreau, Leo (Department: 2721)
Image analysis
Learning systems
Trainable classifiers or pattern recognizers
C382S155000, C382S156000, C382S157000, C706S015000, C706S020000
Reexamination Certificate
active
06278799
ABSTRACT:
FIELD OF THE INVENTION
The present invention relates to a hierarchical artificial neural network (HANN) for automating the recognition and identification of patterns in data matrices.
BACKGROUND
The present invention was developed with a meteorological application. It has particular, although not exclusive, application to the identification of severe storm events (SSEs) from spatial precipitation patterns, derived from conventional volumetric radar imagery and will be discussed in that connection in the following. It is to be understood, however, that the invention has other applications as will be appreciated by those knowledgeable in the relevant arts. It may be applied wherever a pattern in a data matrix is to be recognized and identified, regardless of the orientation, position or scale of the pattern. One application of particular interest is in the analysis of multiple time frames of financial data series, for example stock prices and the like, where it provides a visualization in an n-dimensional space.
Severe storm events (SSEs) include tornadoes, downbursts (including macrobursts—damaging straight line winds caused by downbursts—wind shear, microbursts), large hail and heavy rains. These events, particularly tornadoes, may form quickly, vanish suddenly, and may leave behind great damage to property and life. It is therefore of importance to be able to provide some prediction and warning of the occurrence of these events.
Weather systems are known to be chaotic in behaviour. Indeed, chaos theory was originally introduced to describe unpredictability in meteorology. The equations that describe the temporal behaviour of weather systems are nonlinear and involve several variables. They are very sensitive to initial conditions. Small changes in initial conditions can yield vast differences in future states. This is often referred to as the “butterfly effect.” Consequently, weather prediction is highly uncertain. This uncertainty is likely to be more pronounced when attempting to forecast severe storms, because their structure, intensity and morphology, are presented over a broad spectrum of spatial and temporal scales.
In a storm warning system, problems of prediction originate at the level of storm identification. The uncertainty in initial conditions manifests itself in two distinct forms:
(i) the internal precision and resolution of storm monitoring instruments; and
(ii) the speed at which a storm can be pinpointed.
Furthermore, the recognition of storm patterns based on local observations is not always possible, since the patterns are inherently temporal in nature, with a sensitive dependence on previous states that may not have been observed.
Real-time recognition and identification of SSE patterns from weather radar imagery have been an instrumental component of operational storm alert systems, serving the military, aerospace, and civilian sectors since the early 1950's. This research theme continues to be among the most difficult, complex, and challenging issues confronting the meteorological community. While weather services around the globe have been improving methods of storm surveillance to facilitate the identification and forecasting of SSEs, the resulting increase in both the size and diversity of the resultant data fields have escalated the difficulty with assimilating and interpreting this information.
Factors at the heart of the problem include:
(i) The life cycle of SSEs is very short, in the order of 10 to 30 minutes. They are often of shorter duration than the opportunity to capture, dissect, and analyze the event on radar, let alone interpret the information.
(ii) Unlike real or physical entities, radar patterns do not manifest themselves in a life-like form, but are mere artifacts that resemble the type of reflectivity return expected from bona fide precipitation distributions accompanying SSEs. The relationship between SSEs and these abstractions is analogous to the correspondence between fire and smoke. Just like smoke can prevail after a fire ceases existence, so can a storm pattern be observed in the wake of a SSE. This time lag interferes with the perception of current conditions.
(iii) The features which do assist in the discrimination of SSE patterns rarely display themselves on a single radar image level, but are present at every level on a three dimensional grid. This complication is attributed to the fact that the severity of a storm is a function of buoyancy, the potential energy available to lift a parcel air and initiate convection. Since buoyancy is maximized during SSEs, the convective currents initiated give rise to non-uniform precipitation distributions at various altitudes. Furthermore, since feature structure (pattern boundaries) in the high dimensional data of radar imagery is usually quite sparse, most of the data is redundant. As such, it will likely require an extensive amount of visual processing to extract a sufficient number of features to secure class separability.
(iv) Distinctive SSE signatures: bow; line; hook; and (B)WER, have been universally accepted as indicators of specific storm features: squall lines; strong rotating updrafts; downbursts; and storm tilt. However, their tremendous spatial and temporal variability through translation, rotation, scale, intensity and structure, give rise to non-linear and multiple attendant mappings in the radar image domain, often resulting in two very different events being perceived as one and the same pattern.
(v) Often some of the most severe SSEs, tornadoes and macrobursts, do not visually present themselves on radar reflectivity (Z) imagery, since they occur in the virtual absence of precipitation. Any weak Z patterns displayed are usually buried in noise: radar clutter; side-lobe distortion; and range folding, causing subtle but distinguishing features to be obscured and overlooked.
(vi) The human brain is not conditioned to recognize SSE patterns. This is a complex task at least as difficult to learn as facial and object identification, and speech recognition.
As difficult as the human act of SSE recognition may seem, the more perplexing issue is to translate this process into the algorithmic and machine domain. To date, most approaches to this problem have relied on traditional artificial intelligence (AI) technology, with emphasis on two paradigms: (i) statistical methods; and (ii) artificial rule based experts. W. R. Moninger, “The Artificial Intelligence Shootout: A comparison of Severe Storm Forecasting Systems,” Proc. 16th Conf. on Severe Local Storms, Kananaskis Park, Alta., Canada, Amer. Meteor. Soc., pp. 1-6, 1990 provides a comparative analysis of the implementation of such models in thunderstorm identification systems. K. C. Young, “Quantitative Results for Shootout-89,” Proc. 16th Conf. on Severe Local Storms, Kananaskis Park, Alta., Canada, Amer. Meteor. Soc., pp. 112-115, 1990. elaborates on this study with some quantitative results.
These systems are unnatural in terms of their pattern encoding mechanisms. They make false assumptions about the underlying processes in question and require explicit knowledge, massive amounts of memory or extensive processing to encode, recall, and maintain information.
Statistical methods either make Gaussian assumptions or require a priori information about the underlying distribution of the pattern classes. Since there is insufficient information to fully express the relationships between radar patterns and SSEs, this technique produces unsatisfactory results.
Artificial experts, which rely on the use of explicit rules to emulate the qualitative reasoning and subjective analysis skills of a trained expert, are not appropriate because the nonlinear behaviour of SSEs gives rise to non-explicit descriptions of these relationships.
What is needed is a system that is capable of learning what it needs to know about a particular problem, without prior knowledge of an explicit solution, one which can be incrementally trained to extract and generate its own pattern features from exposure to real time quantitative radar data (stimu
Battison Adrian D.
Boudreau Leo
Mariam Daniel G.
LandOfFree
Hierarchical data matrix pattern recognition system does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Hierarchical data matrix pattern recognition system, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Hierarchical data matrix pattern recognition system will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2436238