Random and spatially autocorrelated sensor noise effects on image classification
One factor limiting the accuracy of land cover maps derived from classified, remotely-sensed imagery is the quality of the spectral data used in the classification process. Satellite data is routinely pre-processed to improve both its geometric and radiometric qualities. We implement a factorial design that assesses the individual and joint effects of simulated sensor noise on specific spectral bands and along continua of intensity and spatial configuration; an image with no added simulated noise is our control. Our focus is on the radiometric component of image quality, as we assume that for our single-image controlled experiment, the multispectral bands are all perfectly aligned and that topographic relief insignificantly affects the geometric properties of our data. For each simulated noisy image we produce a detailed land cover classification using identically-defined classification tree decisions and observe the spatial changes relative to the classification of the control image. We assess the classification accuracy between all noisy cases and the control using traditional error matrices and measures of overall thematic agreement. The objective is to perform a full sensitivity analysis that quantifies the effect of noisy data on image classification, both in terms of the aspatial class area tabulations and their spatial configurations. We link the classification differences with uncertainty metrics as a guide to improving the selection of classifiers and pre-processing techniques.
|Keywords||Autocorrelated noise, CART, Composition, Configuration, Random noise, Uncertainty|
|Conference||9th International Symposium on Spatial Accuracy Assessment in Natural Resources and Environmental Sciences, Accuracy 2010|
Remmel, T.K. (Tarmo K.), & Mitchell, S. (2010). Random and spatially autocorrelated sensor noise effects on image classification. Presented at the 9th International Symposium on Spatial Accuracy Assessment in Natural Resources and Environmental Sciences, Accuracy 2010.