Mapping Spatial Distribution of Land Cover Classification Errors Maria João Pereira, Amílcar Soares CERENA – Centre for Natural Resources and Environment 1

Introduction 2 Learning Selection of training areas Determine multivariate relation Generalization Spatial and temporal stationarity of multivariate relation Accuracy Assessment Validation data set Confusion matrixClassificaon

Land Cover Maps 3 Classification errors mismatches between actual ground-based and image derived class

Confusion Matrix 4 Table 1. Confusion matrix. Class labels: A – coniferous forest; B – deciduous forest; C – grassland; D – permanent tree crops; E– non-irrigated land; F – irrigated land; G – artificial areas; H – water; I – maquis and mixed forest.

Geostatistics 5  indicator kriging with locally varying means to integrate the image classifier’s posterior probability vectors and reference data (Kyriakidis & Dungan, 2001)  SIS with prediction via collocated indicator cokriging for updating cover type maps and for estimation of the spatial distribution of prediction errors (Magnussen and De Bruin, 2003)

Objective 6 Mapping the spatial distribution of classification errors based on stochastic simulation and that takes into account :  the spatial continuity of each land cover class errors.  Varying errors’ patterns over the classification area Classification error

Rationale 7 for each thematic class different errors occur depending on sensors and ground conditions Assumption Class A Class B Classification error

Method 8 1. Calculate the trend of the errors mi 2. Calculate local error e(x) conditioned to the mean error of the predicted class for that location and to the neighboring error values

Method 9 SIS with varying local means Map the distribution of classification errors Map the associated uncertainty indicator kirging estimation local errors means for each thematic class

Mapping local mean error of thematic classe i 10 Indicator kriging experimental data errors e i (x 0 ) kriging weights Number of neibghour data

Mapping local mean error of thematic classe i 11

Mapping the spatial dispersion of classification error e(x) Define a random path visiting each node u of the grid 2. For each location u along the path 1. Search conditioning data (point data and previously simulated values) and compute point-to-point covariances 2. Build and solve the kringing system conditioned to local varying means 3. Define local ccdf with its mean and variance given by the kriging estimate and variance 4. Draw a value from the ccdf and add the simulated value to data set 3. Repeat to generate another simulated realization

Mapping the spatial dispersion of classification error e(x) 13 Mean image Mean image

Results Mean Variance 14

Final remarks 15  Geostatistics provides na adequacte framework to assess spatial accuracy  In areas with field data, its influence prevails over the error trend mi(x) and vice-versa;  The method succeeded to map the spatial distribution of classification errors accounting for :  the spatial continuity of each land cover class errors.  Varying errors pattern over the classification area

Thank you! 16 Project Landau - Contract Ref. PTDC/CTE-SPA/103872/2008