Per-pixel land cover accuracy prediction: A random forest-based method with limited reference sample data

Abstract

Given the importance of accuracy in land cover (LC) maps, several methods have been adopted to predict per-pixel land cover accuracy (PLCA) of classified remote sensing images. Such a PLCA map provides spatially-explicit accuracy information and is of paramount importance for both producers and end-users of LC maps to thoroughly understand the spatial distribution of accuracy. In this study, we proposed a simple yet powerful random forest (RF) based approach for PLCA mapping with limited reference sample data. The main assumption of the proposed approach is that the LC’s misclassifications do not occur randomly, but rather exhibit some detectable characteristics which can be retrieved via the built model. With this approach, RF attempts to establish a nonlinear relationship between the accuracy and the same spectral bands used in LC classification. To confirm the proposed method as a consistent and practical approach for a variety of different settings, we evaluated it on five different classified remote sensing images derived from Landsat-8, Ikonos, and three Sentinel-2 images across different parts of Iran. In this manner, to validate the predictive capability of the RF-based method, we calculated the area under the receiver operating characteristic curve (AUROC) and several other statistical metrics, including sensitivity (SN), specificity (SP), positive predictive value (PPV), negative predictive value (NPV), and accuracy (ACC). Analysis of the average values of these metrics (AUROC = 0.88, SN = 95%, SP = 68%, PPV = 96%, NPV = 72%, and ACC = 95%) derived from the limited sample size datasets showed that the proposed model performs well in all case studies. The performance of the proposed model was further assessed through comparison against two benchmark methods, namely Gaussian kernel interpolation (GKI) and linear kernel interpolation (LKI). In conclusion, although our comprehensive evaluations revealed that RF, GKI, and LKI methods are promising approaches for PLCA mapping, RF outperformed both GKI and LKI in all of the experimental sites.