Limits...
Spectral signature generalization and expansion can improve the accuracy of satellite image classification.

Laborte AG, Maunahan AA, Hijmans RJ - PLoS ONE (2010)

Bottom Line: When applying signatures to the images they were derived from, signature expansion improved accuracy relative to the conventional method, and variability in accuracy declined markedly.In contrast, signature generalization did not improve classification.When applying signatures to images of other years (temporal extension), the conventional method, using a signature derived from a single image, resulted in very low classification accuracy.

View Article: PubMed Central - PubMed

Affiliation: International Rice Research Institute, Los Baños, Laguna, Philippines. aglaborte@gmail.com

ABSTRACT
Conventional supervised classification of satellite images uses a single multi-band image and coincident ground observations to construct spectral signatures of land cover classes. We compared this approach with three alternatives that derive signatures from multiple images and time periods: (1) signature generalization: spectral signatures are derived from multiple images within one season, but perhaps from different years; (2) signature expansion: spectral signatures are created with data from images acquired during different seasons of the same year; and (3) combinations of expansion and generalization. Using data for northern Laos, we assessed the quality of these different signatures to (a) classify the images used to derive the signature, and (b) for use in temporal signature extension, i.e., applying a signature obtained from data of one or several years to images from other years. When applying signatures to the images they were derived from, signature expansion improved accuracy relative to the conventional method, and variability in accuracy declined markedly. In contrast, signature generalization did not improve classification. When applying signatures to images of other years (temporal extension), the conventional method, using a signature derived from a single image, resulted in very low classification accuracy. Signature expansion also performed poorly but multi-year signature generalization performed much better and this appears to be a promising approach in the temporal extension of spectral signatures for satellite image classification.

Show MeSH
Methods used to create land cover class spectral signatures and their use in classification.Methods: A (conventional), B (expansion), C (single-year generalization), D (multi-year generalization), E (single-year expansion + generalization), F (multi-year expansion + generalization); and 1 (no signature extension), 2 (signature extension).
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC2865537&req=5

pone-0010516-g002: Methods used to create land cover class spectral signatures and their use in classification.Methods: A (conventional), B (expansion), C (single-year generalization), D (multi-year generalization), E (single-year expansion + generalization), F (multi-year expansion + generalization); and 1 (no signature extension), 2 (signature extension).

Mentions: Figure 2 summarizes the approaches we considered to obtain spectral signatures for land cover classes. Signatures can be derived in the conventional manner, i.e., from a single image (A), or by using multiple images (B–F). Signature expansion (B) consists of integrating two (or more) images from different seasons within a single year. The images are “stacked” and treated as additional predictor variables (spectral bands), and training data for the same year are used. In signature generalization (C–D), additional images are treated as additional observations, i.e., the number of predictor variables (bands) remains the same. By using more than one image, the overall signal to noise ratio might be higher than that associated with either of the single images. Figure 2E–F illustrates combinations of signature expansion and generalization. In all cases, signatures derived from multiple images can be used for classifying land cover for the period covered by the images used. They can also be used for temporal extension, i.e., to classify images for a different time period.


Spectral signature generalization and expansion can improve the accuracy of satellite image classification.

Laborte AG, Maunahan AA, Hijmans RJ - PLoS ONE (2010)

Methods used to create land cover class spectral signatures and their use in classification.Methods: A (conventional), B (expansion), C (single-year generalization), D (multi-year generalization), E (single-year expansion + generalization), F (multi-year expansion + generalization); and 1 (no signature extension), 2 (signature extension).
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC2865537&req=5

pone-0010516-g002: Methods used to create land cover class spectral signatures and their use in classification.Methods: A (conventional), B (expansion), C (single-year generalization), D (multi-year generalization), E (single-year expansion + generalization), F (multi-year expansion + generalization); and 1 (no signature extension), 2 (signature extension).
Mentions: Figure 2 summarizes the approaches we considered to obtain spectral signatures for land cover classes. Signatures can be derived in the conventional manner, i.e., from a single image (A), or by using multiple images (B–F). Signature expansion (B) consists of integrating two (or more) images from different seasons within a single year. The images are “stacked” and treated as additional predictor variables (spectral bands), and training data for the same year are used. In signature generalization (C–D), additional images are treated as additional observations, i.e., the number of predictor variables (bands) remains the same. By using more than one image, the overall signal to noise ratio might be higher than that associated with either of the single images. Figure 2E–F illustrates combinations of signature expansion and generalization. In all cases, signatures derived from multiple images can be used for classifying land cover for the period covered by the images used. They can also be used for temporal extension, i.e., to classify images for a different time period.

Bottom Line: When applying signatures to the images they were derived from, signature expansion improved accuracy relative to the conventional method, and variability in accuracy declined markedly.In contrast, signature generalization did not improve classification.When applying signatures to images of other years (temporal extension), the conventional method, using a signature derived from a single image, resulted in very low classification accuracy.

View Article: PubMed Central - PubMed

Affiliation: International Rice Research Institute, Los Baños, Laguna, Philippines. aglaborte@gmail.com

ABSTRACT
Conventional supervised classification of satellite images uses a single multi-band image and coincident ground observations to construct spectral signatures of land cover classes. We compared this approach with three alternatives that derive signatures from multiple images and time periods: (1) signature generalization: spectral signatures are derived from multiple images within one season, but perhaps from different years; (2) signature expansion: spectral signatures are created with data from images acquired during different seasons of the same year; and (3) combinations of expansion and generalization. Using data for northern Laos, we assessed the quality of these different signatures to (a) classify the images used to derive the signature, and (b) for use in temporal signature extension, i.e., applying a signature obtained from data of one or several years to images from other years. When applying signatures to the images they were derived from, signature expansion improved accuracy relative to the conventional method, and variability in accuracy declined markedly. In contrast, signature generalization did not improve classification. When applying signatures to images of other years (temporal extension), the conventional method, using a signature derived from a single image, resulted in very low classification accuracy. Signature expansion also performed poorly but multi-year signature generalization performed much better and this appears to be a promising approach in the temporal extension of spectral signatures for satellite image classification.

Show MeSH