Smaller is better? Unduly nice accuracy assessments in roof detection using remote sensing data with machine learning and k-fold cross-validation

Deriving the thematic accuracy of models is a fundamental part of image classification analyses. K-fold cross-validation (KCV), as an accuracy assessment technique, can be biased because existing built-in algorithms of software solutions do not handle the high autocorrelation of remotely sensed images, leading to overestimation of accuracies. We aimed to quantify the magnitude of the overestimation of KCV-based accuracies and propose a method to overcome this problem with the example of rooftops using a WorldView-2 (WV2) satellite image, and two orthophotos. Random split to training/testing subsets, independent testing and different types of repeated KCV sampling strategies were used to generate input datasets for classification. Results revealed that applying the random splitting of reference data to training/testing subsets and KCV methods had significantly biased the accuracies by up to 17%; overall accuracies (OAs) can incorrectly reach >99%. We found that repeated KCV can provide similar results to independent testing when spatial sampling is applied with a sufficiently large distance threshold (in our case 10 m). Coarser resolution of WV2 ensured more reliable results (up to a 5–9% increase in OA) than orthophotos. Object-based pixel purity of buildings showed that when using a majority filter for at least of 50% of objects the final accuracy approached 100% with each sampling method. The final conclusion is that KCV-based modelling ensures better accuracy than single models (with better pixel purity on the object level), but the accuracy metrics without spatially filtered sampling are not reliable.