On the Importance of Training Data Sample Selection in Random Forest Image Classification: A Case Study in Peatland Ecosystem Mapping

Abstract

Random Forest (RF) is a widely used algorithm for classification of remotely sensed data. Through a case study in peatland classification using LiDAR derivatives, we present an analysis of the effects of input data characteristics on RF classifications (including RF out-ofbag error, independent classification accuracy and class proportion error). Training data selection and specific input variables (i.e., image channels) have a large impact on the overall accuracy of the image classification. High-dimension datasets should be reduced so that only uncorrelated important variables are used in classifications. Despite the fact that RF is an ensemble approach, independent error assessments should be used to evaluate RF results, and iterative classifications are recommended to assess the stability of predicted classes. Results are also shown to be highly sensitive to the size of the training data set. In addition to being as large as possible, the training data sets used in RF classification should also be (a) randomly distributed or created in a manner that allows for the class proportions of the training data to be representative of actual class proportions in the landscape; and (b) should have minimal spatial autocorrelation to improve classification results and to mitigate inflated estimates of RF out-of-bag classification accuracy.

Multi-year MODIS active fire type classification over the Brazilian Tropical Moist Forest Biome

“…However, the training data should be distributed evenly in geographic space to avoid generating spurious classification accuracies (Friedl, Brodley, and Strahler 1999), which is not possible in this study as the MODIS active fire detections are sparsely distributed. In addition, unlike supervised land cover classification approaches, where training sample points inherently have high spatial autocorrelation due to the way they are collected (Egorov et al 2015;Millard and Richardson 2015), the training data in this study were derived from a random subset of the MODIS active fire detections and therefore are less likely to be spatially autocorrelated. As observed in similar regional fire-related random forest studies (Archibald et al 2009;Oliveira et al 2012), spatial autocorrelation of predictor variables may occur due to various physical and biological processes but no technique to incorporate spatial dependence has been reliably demonstrated and this remains an area of active research.…”

Mapping the Dabus Wetlands, Ethiopia, Using Random Forest Classification of Landsat, PALSAR and Topographic Data

“…OOB error was used as a basis for comparison of classifications to determine optimum input parameters, years and seasons (Table 4) as described below. OOB error has been shown to be optimistic compared to independent sample validation accuracy [75,80], but when applied consistently in the same manner, it can be an efficient way to compare classifications and conduct variable selection. It was preferred over independent validation for this study given: (1) the field generated reference data set sample size was limited due to poor accessibility to all parts of the wetlands; and (2) both the field and image-based reference samples follow the general arcuate shape of the wetlands and were probably spatially auto-correlated.…”

“…Although RF is generally considered robust to overfitting [106], it is highly likely that overall classification accuracy levels reported in this study were overly optimistic since the RF accuracy assessment was obtained from the "Out-of-Bag" (OOB) accuracy estimate, which is known to represent inflated accuracy [75,80]. OOB accuracy is useful for comparison of multiple classification models as in this study but independent validation is required to determine absolute accuracy.…”

“…In addition, as a result of the Dabus Marsh arcuate configuration in the landscape, the spatial distribution of the reference data shows a highly-clustered pattern with expected spatial autocorrelation, which was not measured. Such clustering of reference sample locations can contribute to classification accuracy inflation [80,108]; however, this could not be avoided. Finally, the field surveys represented less than 5-10% of the total area, and were mainly concentrated in the northwestern portion of the Dabus Wetlands.…”

Improving Selection of Spectral Variables for Vegetation Classification of East Dongting Lake, China, Using a Gaofen-1 Image

“…2018, 10, 50 3 of 25 intensive for computation. Moreover, a machine learning algorithm-random forests (RF) has been widely used for classification of LULC types [21][22][23][24][25][26][27]. This method has the ability of optimizing both classification results and selection of remote sensing variables.…”