Impacts of sensor noise on land cover classifications: sensitivity analysis using simulated noise

Land cover maps are typically derived through classification of remotely-sensed data, usually relying on a multispectral set of views from satellite-borne sensors. A number of corrections can be applied to improve the geometric and radiometric quality of the data. We have been assessing the importance of radiometric quality using a factorial design to assess the impacts of various levels of sensor noise, and spatial autocorrelation of that noise on classified land cover maps. We compare simulated noise-affected maps to a control with no noise addition. The objective is a sensitivity analysis to quantify the effects of noisy data on image classification, both on the overall accuracy statistics and the spatial configurations of error maps. We have tested a series of noise models, and our first results show that most classifications are relatively robust except in conditions of noise with high spatial autocorrelation. We are continuing to refine our noise models to better represent typical combinations of sensor and atmospheric conditions. Background and Relevance Accurate land cover maps from classified multispectral images require high signal-tonoise ratios, which are affected by signal-dependent noise (Rangayyan et al., 1998), atmospheric influences (Song et al., 2001), and systematic sensor errors, such as striping (Pan and Chang, 1992; Torres and Infante, 2001). Much work has gone into attempts to understand and correct for these sources of noise, but there is very little research on quantifying the impact of these potential errors on classification of the data into land cover categories. Therefore, we aimed to study the stability of common classification regimes in the presence of a range of likely noise treatments. Initial experiments using this approach (Remmel and Mitchell, 2010) employed regression trees for the classification and only 4 classes, and found very little difference between treatments, likely due to the high thematic aggregation. We have moved to more detailed classes, and are examining alternative error models. This presentation demonstrates the first findings from this process. The full results of this analysis will guide methods selection for further land cover products, and justifications of the effort used to correct imagery for given applications.