Satellite Image Automatic Mapper (SIAM) - A first-stage preliminary classifier for automatic two-stage Land Cover classification and Change detection

Andrea Baraldi, University of Maryland, andrea.baraldi@hermes.geog.umd.edu (Presenter)
Luigi Boschetti, University of Maryland, luigi@hermes.geog.umd.edu
Chris Justice, University of Maryland, justice@hermes.geog.umd.edu

SIAM is an automatic, multi-sensor, multi-resolution, near real-time, non-adaptive (deductive, physical model-based) decision-tree classifier based on prior spectral knowledge of surface types observed from space. Its prior spectral knowledge base comprises a reference dictionary of spectral signatures in top-of-atmosphere reflectance or surface reflectance, acquired from off-line data observations and/or existing literature. Since its knowledge base is available before looking at the specific image to be classified, SIAM belongs to the family of physical models, also called deductive inference systems. Employed as a preliminary classification first stage of a two-stage remote sensing image understanding system, SIAM enforces a shift in learning paradigm from traditional first-stage inductive Machine-Learning-from-data (e.g., image segmentation) to deductive Machine-Teaching-by-rules.

The first-stage SIAM preliminary classifier requires as input multi-spectral images radiometrically calibrated into top-of-atmosphere or surface reflectance, and brightness temperature. SIAM is pixel-based (non-contextual), i.e., it is based on spectral properties exclusively. It maps each pixel onto a discrete and finite set of spectral categories (spectral-based semi-concepts, land cover class sets) belonging to a set of six spectral super-categories (spectral end members): (I) clouds, (II) snow or ice, (III) water or shadow, (IV) vegetation, (V) bare soil or built-up and (VI) outliers.

SIAM is a multi-sensor multi-resolution classification system of systems eligible for use with all existing spaceborne optical sensors whose: (a) spectral resolution overlaps with, but is inferior to, Landsat’s and (b) spatial resolution ranges from 0.5 m (pan-sharpened WV-2) to 3 km (Meteosat SEVIRI).

Being a fully automatic (no user-defined parameter, no training samples), near real- time (less than 5 min to classify a Landsat scene with a laptop computer) system, SIAM can be used for the automatic classification of large volumes of data, and can be integrated in decision support systems designed to meet the needs for Monitoring, Reporting and Verification (MRV) of international programs.

Presentation Type:  Poster

Session:  Science in Support of Decision Making   (Wed 10:00 AM)

Associated Project(s): 

• Boschetti, Luigi: MODIS-Landsat data fusion for high spatial resolution multiannual wall to wall burned area mapping of the conterminous United States ...details

Poster Location ID: 307