Random forests and stochastic gradient boosting for predicting tree canopy cover: comparing tuning processes and model performance

Abstract

1As part of the development of the 2011 National Land Cover Database (NLCD) tree canopy 2 cover layer, a pilot project was launched to test the use of high resolution photography coupled 3 with extensive ancillary data to map the distribution of tree canopy cover over four study regions 4 in the conterminous US. Two stochastic modeling techniques, Random Forests (RF) and 5Stochastic Gradient Boosting (SGB), are compared. The objectives of this study were first to 6 explore the sensitivity of RF and SGB to choices in tuning parameters. Second, to compare the 7 performance of the two final models by assessing the importance of, and interaction between, 8 predictor variables, the global accuracy metrics derived from an independent test set, as well as 9 the visual quality of the resultant maps of tree canopy cover. The predictive accuracy of RF and 10 SGB was remarkably similar on all four of our pilot regions. In all four study regions, the 11 independent test set MSE was identical to 3 decimal places, with the largest difference in Kansas 12where RF gave an MSE of 0.0113 and SGB gave an MSE of 0.0117. With correlated predictor 13 variables, stochastic gradient boosting had a tendency to concentrate variable importance in 14 fewer variables, while Random Forest tended to spread importance out amongst more variables. 15 RF is simpler to implement than SGB, as RF both has fewer parameters needing tuning, and also 16 was less sensitive to these parameters. As stochastic techniques, both RF and SGB introduce a 17 new component of uncertainty: repeated model runs will potentially result in different final 18 predictions. We demonstrate how RF allows the production of a spatially explicit map of this 19 stochastic uncertainty of the final model. 20

Effects of Burn Severity and Environmental Conditions on Post-Fire Regeneration in Siberian Larch Forest

“…Consequently, there were a total of 1350 samples with 70% of them used as a training dataset and 30% used for validation. Similar to other stochastic techniques, the RF model may introduce uncertainty to the final predictions since model runs can include highly correlated variables and results in spurious variation [74,75]. To reduce model uncertainty, multicollinearity among predictors was first examined using a variance inflation factor (VIF).…”

“…All proposed predictor variables (Table 2) were independent and thus included in the RF model. Second, the process of RF model tuning [74] was applied to the 70% training samples where the ntrees (i.e., number of trees) parameter in the model was selected based on building different RF models with increasing numbers of trees ranging from 100 to 2000 in 100 tree increments. Based on the 30% testing dataset, the numbers of trees that stabilized the role of predictor variables in determining response variables was then selected as an optimal ntrees parameter in the final RF model.…”

Quantifying Live Aboveground Biomass and Forest Disturbance of Mountainous Natural and Plantation Forests in Northern Guangdong, China, Based on Multi-Temporal Landsat, PALSAR and Field Plot Data

“…The weaknesses of the RF method are that the decision rules are unknown (black box), it is computationally intense and it requires input parameters [69,70]. We used RF regression tree implemented through the R package ModelMap [71] by calling the R package random forest [72] to identify important predictor variables, to model the relationship between the predicted variables and AGB, and to apply the model over the study area for mapping AGB based on single or composited Landsat surface reflectance images and mosaic PALSAR data.…”

Improving sub-pixel imperviousness change prediction by ensembling heterogeneous non-linear regression models

“…This approach was previously used for sub-pixel imperviousness assessment based on Landsat (Walton, 2008); (Bernat and Drzewiecki, 2014) and MODIS images (Deng and Wu, 2013;Tsutsumida, 2016). It was also chosen by U.S. Forest Service to produce the 2011 NLCD percent tree canopy cover dataset (Freeman et al, 2016).…”

“…In case of stochastic gradient boosting, the models in each iteration are based only on the randomly selected fraction of training data (Kuhn and Johnson, 2013). This method was used to estimate aboveground biomass based on regression with Landsat and SPOT or ALOS PALSAR derived predictors (Carreiras, 2013) and predicting tree canopy cover based on Landsat-5 images (Freeman, 2016). To my best knowledge, it has not been used for sub-pixel imperviousness evaluation so far.…”