Abstract: |
The application of remote sensing image classification to derive land covers maps is widely used, because it is a simple and fast procedure. However, these maps are many times disregarded for land use planning and management due to the difficulty to assess accuracy, as well as the lack of reference methods to tackle the problem. Presently land cover classification accuracy assessments are based on error or confusion matrix, which is a simple cross-tabulation of the mapped class against that observed in the reference data at a set of validation pixels providing a summary of commission (type I) errors and omission (type II) errors. Spatial characterization of classification errors is of prime importance for the further use of classified thematic maps: characterization of spatial uncertainty areas, evaluation of the classified themes which can be considered reliable or with the need of local field samples, etc. Geostatistics framework is appropriate to model spatial variation of the classification uncertainty. Previous works proposed the use of indicator kriging to local varying means and sequential indicator simulation with prediction via collocated indicator cokriging. However, two main problems remain unsolved: the incorporation of distinct spatial error patterns for each thematic class due to its radiometric features and previous methodologies do not take into account patch sizes contribution to uncertainty. In the present work, these two issues are address through the use of patch size weighted spatial covariance estimation in conjunction within the framework of Direct Sequential Simulation algorithm. Early tests of the methodology applied to a synthetic application shown promising results. Further testing and validation is undergoing over thematic maps concerning different Portuguese landscapes for which ground truth data is available. |
