A data-driven approach to quality assessment for hyperspectral systems

Abstract

The increasing use of products based on airborne hyperspectral data for decision-making calls for a thorough quality assessment. Due to the complexity of the corresponding processing chain, as well as the variety of physical processes involved, such a task is usually only performed in specific cases and on specific parts of the processing chain. In particular, the quality assessment of data-products is still an open issue. A generic quality assessment method – based on an cross-comparison of errors – is proposed in this paper. Airborne hyperspectral – also called imaging spectroscopy – data is commonly acquired by means of whisk- or push-broom sensors, and requires several strips – or flight-lines – to cover the full area of interest. A comparison of the discrepancies between overlapping parts of these flight-lines is used to retrieve an assessment of the measurement reproducibility. This mapping can be performed on pre-processed data which avoids the need to separately investigate all input parameters and their associated models, hence bypassing the ‘curse of dimensionality’. The first step involves retrieving the pairs of pixels corresponding to the same areas imaged from overlapping flight-lines. Even when an ortho-rectification of the data has been carried out, various phenomena, such as errors in the underlying digital elevation model, lead to flight-line mis-registrations. For heterogeneous land-covers, a pixel to pixel registration step has therefore to be performed in order to allow a cross-comparison of pixels: a suitable methodology is proposed along with its validation. The second step corresponds to the relative errors analysis itself. A set of quantitative quality indicators – corresponding to different types of land-products – is presented. These methods are illustrated with an example along with a discussion. This approach can be used on any reasonably well contrasted scene to retrieve a quality assessment for any raster product independently of its data type as well as for the reflectance data itself.