A credit assignment approach to fusing classifiers of multiseason hyperspectral imagery

Bachmann, Charles M.; Bettenhausen, M.; Fusina, R.A.; Donato, T.F.; Russ, Andrew L.; Burke, Joseph; Lamela, Gia; Rhea, W. Joseph; Truitt, Barry R.; Porter, John H.

doi:10.1109/tgrs.2003.818537

Cited by 27 publications

(18 citation statements)

References 36 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…By fusing slightly different observations of the same spatial scene -each of which generates a slightly differen spatial patternswe follow a similar approach. As pointed out earlier, e.g., by Bachman et al: "In remote sensing applications, variance reduction also can be achieved by using multisensor or multi-temporal data to produce a pool of classifiers with decorrelated error distributions" [50]. However, while some recommend very generally to pool predictions [51] and that "inaccurate classifiers should not be excluded .…”

Section: Multitemporal Fusion and Bias-variance Tradeoffmentioning

confidence: 99%

Multitemporal Fusion for the Detection of Static Spatial Patterns in Multispectral Satellite Images—With Application to Archaeological Survey

Menze¹,

Ur²

2014

IEEE J. Sel. Top. Appl. Earth Observations Remote Sensing

View full text Add to dashboard Cite

Abstract-We evaluate and further develop a multitemporal fusion strategy that we use to detect the location of ancient settlement sites in the Near East and to map their distribution, a spatial pattern that remains static over time. For each ASTER images that has been acquired in our survey area in north-eastern Syria, we use a pattern classification strategy to map locations with a multispectral signal similar to the one from (few) known archaeological sites nearby. We obtain maps indicating the presence of anthrosol -soils that formed in the location of ancient settlements and that have a distinct spectral pattern under certain environmental conditions -and find that pooling the probability maps from all available time points reduces the variance of the spatial anthrosol pattern significantly. Removing biased classification mapsi.e. those that rank last when comparing the probability maps with the (limited) ground truth we have -reduces the overall prediction error even further, and we estimate optimal weights for each image using a non-negative least squares regression strategy. The ranking and pooling strategy approach we propose in this study shows a significant improvement over the plain averaging of anthrosol probability maps that we used in an earlier attempt to map archaeological sites in a 20 000 km 2 area in northern Mesopotamia, and we expect it to work well in other surveying tasks that aim at mapping static surface patterns with limited ground truth in long series of multispectral images.

show abstract

Section: Multitemporal Fusion and Bias-variance Tradeoffmentioning

confidence: 99%

Multitemporal Fusion for the Detection of Static Spatial Patterns in Multispectral Satellite Images—With Application to Archaeological Survey

Menze¹,

Ur²

2014

IEEE J. Sel. Top. Appl. Earth Observations Remote Sensing

View full text Add to dashboard Cite

show abstract

“…Adding an extra element to the representation which indicates the value of K at which curvature onset occurs, using our adaptive K estimation procedure, will have the effect of sub-dividing the manifold into sets of varying degree of curvature ( Figure 5). We demonstrate this approach for the problem of scene anomaly finding using PROBE airborne HSI data of Smith Island, VA, originally described in [6] [7]. Scene anomalies, in manifold representation or in the new submanifold derivations, tend to project off of the background data distribution (Figure 4).…”

Section: Re-examining Data Curvaturementioning

confidence: 99%

Manifold coordinate representations of hyperspectral imagery: Improvements in algorithm performance and computational efficiency

Bachmann

Ainsworth

Fusina

et al. 2010

2010 IEEE International Geoscience and Remote Sensing Symposium

View full text Add to dashboard Cite

Several years ago, NRL first demonstrated a computationally efficient framework for optimizing a set of intrinsic manifold coordinates for high-dimensional data such as hyperspectral imagery (HIS). Since that time, NRL has continued to improve the fidelity of the representation to describe the details of the nonlinear structure in even greater detail. In addition, working closely with CelesTech, several improvements have been made that increase the computational efficiency, accelerating processing speeds from hours to minutes.

show abstract

“…To illustrate ARESEPE, we have chosen a set of airborne hyperspectral data drawn from a larger scale study described in [2] and [3]. In particular, we evaluate land-cover classification models derived from a PROBE2 scene acquired on October 18, 2001 of Smith Island, VA, a barrier island in The Nature Conservancy's Virginia Coast Reserve (VCR).…”

Section: Example: Probe2 Airborne Hyperspectral Imagerymentioning

confidence: 99%

“…In particular, we evaluate land-cover classification models derived from a PROBE2 scene acquired on October 18, 2001 of Smith Island, VA, a barrier island in The Nature Conservancy's Virginia Coast Reserve (VCR). 1 Land-cover modeling using a HyMAP scene of this island from May 2000 was described in [2], while [3] addresses multiseason models that have been derived for this island and which show improvement over single-season models, including the set used for illustration in this paper. The database of labeled spectral samples was divided into a Training Set (3632 samples), a Cross-Validation Test Set (1971 samples) used to determine the best model during the optimization, and a Sequestered Test Set (2834 samples) that served as an independent test of generalization capability.…”

Section: Example: Probe2 Airborne Hyperspectral Imagerymentioning

confidence: 99%

Improving the performance of classifiers in high-dimensional remote sensing applications: an adaptive resampling strategy for error-prone exemplars (ARESEPE)

Bachmann

2003

IEEE Trans. Geosci. Remote Sensing

Self Cite

View full text Add to dashboard Cite

In the past, "active learning" strategies have been proposed for improving the convergence and accuracy of statistical classifiers. However, many of these approaches have large storage requirements or unnecessarily large computational burdens and, therefore, have been impractical for the large-scale databases typically found in remote sensing, especially hyperspectral applications. In this paper, we develop a practical on-line approach with only modest storage requirements. The new approach improves the convergence rate associated with the optimization of adaptive classifiers, especially in high-dimensional remote sensing data. We demonstrate the new approach using PROBE2 hyperspectral imagery and find convergence time improvements of two orders of magnitude in the optimization of land-cover classifiers.Index Terms-Active learning, active sampling, barrier islands, hyperspectral, land-cover classification, Virginia Coast Reserve.

show abstract

A credit assignment approach to fusing classifiers of multiseason hyperspectral imagery

Cited by 27 publications

References 36 publications

Multitemporal Fusion for the Detection of Static Spatial Patterns in Multispectral Satellite Images—With Application to Archaeological Survey

Multitemporal Fusion for the Detection of Static Spatial Patterns in Multispectral Satellite Images—With Application to Archaeological Survey

Manifold coordinate representations of hyperspectral imagery: Improvements in algorithm performance and computational efficiency

Improving the performance of classifiers in high-dimensional remote sensing applications: an adaptive resampling strategy for error-prone exemplars (ARESEPE)

Contact Info

Product

Resources

About