Semi-automated classification of exposed bedrock cover in British Columbia's Southern Mountains using a Random Forest approach

Scarpone, Christopher; Schmidt, Margaret G.; Bulmer, Chuck; Knudby, Anders

doi:10.1016/j.geomorph.2017.02.013

Cited by 16 publications

(13 citation statements)

References 46 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The BART modelling demonstrated that although many variables can be used as inputs, the majority of variability explained will largely be determined from a subset of all variables, whenceforth only a marginal decrease in accuracy will be observed [ 58 ]. The preliminary model decreased from 366 to only 31 variables in the final model, and the only exhibited a proportionally small decrease from 0.89 to 0.73, with very minimal differences in variable importance among the top 10 variables shown herein.…”

Section: Discussionmentioning

confidence: 99%

“…Variable importance plots were generated from the BART model, which displays a quantitative metric of a variable’s relative influence on model predictions, compared to all other variables [ 53 ]. We also generated Partial Dependence Plots (PDPs), which are graphical outputs that illustrate the marginal effect of each independent variable on the response variable [ 56 – 58 ]. A PDP only displays the marginal effect of each independent variable in relation to the influence of all other independent variables, and should be interpreted as exploratory [ 53 ].…”

Section: Methodsmentioning

confidence: 99%

See 1 more Smart Citation

A multimethod approach for county-scale geospatial analysis of emerging infectious diseases: a cross-sectional case study of COVID-19 incidence in Germany

et al. 2020

Self Cite

View full text Add to dashboard Cite

Background: As of 13 July 2020, 12.9 million COVID-19 cases have been reported worldwide. Prior studies have demonstrated that local socioeconomic and built environment characteristics may significantly contribute to viral transmission and incidence rates, thereby accounting for some of the spatial variation observed. Due to uncertainties, non-linearities, and multiple interaction effects observed in the associations between COVID-19 incidence and socioeconomic, infrastructural, and built environment characteristics, we present a structured multimethod approach for analysing cross-sectional incidence data within in an Exploratory Spatial Data Analysis (ESDA) framework at the NUTS3 (county) scale. Methods: By sequentially conducting a geospatial analysis, an heuristic geographical interpretation, a Bayesian machine learning analysis, and parameterising a Generalised Additive Model (GAM), we assessed associations between incidence rates and 368 independent variables describing geographical patterns, socioeconomic risk factors, infrastructure, and features of the build environment. A spatial trend analysis and Local Indicators of Spatial Autocorrelation were used to characterise the geography of age-adjusted COVID-19 incidence rates across Germany, followed by iterative modelling using Bayesian Additive Regression Trees (BART) to identify and measure candidate explanatory variables. Partial dependence plots were derived to quantify and contextualise BART model results, followed by the parameterisation of a GAM to assess correlations. Results: A strong south-to-north gradient of COVID-19 incidence was identified, facilitating an empirical classification of the study area into two epidemic subregions. All preliminary and final models indicated that location, densities of the built environment, and socioeconomic variables were important predictors of incidence rates in Germany. The top ten predictor variables' partial dependence exhibited multiple non-linearities in the relationships between key predictor variables and COVID-19 incidence rates. The BART, partial dependence, and GAM results indicate that the strongest predictors of COVID-19 incidence at the county scale were related to community interconnectedness, geographical location, transportation infrastructure, and labour market structure.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

A multimethod approach for county-scale geospatial analysis of emerging infectious diseases: a cross-sectional case study of COVID-19 incidence in Germany

et al. 2020

Self Cite

View full text Add to dashboard Cite

show abstract

“…For example, Milodowski et al (2015) created a computational algorithm for calculating a roughness metric which performed well predicting rock exposures on rapidly eroding hillslopes in California and Idaho, USA. Scarpone et al (2017) increased accuracy from 48% of bedrock exposures in manually delineated legacy land cover maps to 88% with the use of a random forest model using 17 predictor variables predictions in southern British Columbia, Canada.…”

Section: Comparison With Other Regionsmentioning

confidence: 99%

“…Extracting bedrock outcrop features, however, is especially challenging in landscapes where outcrops are scattered and where associated thin soils constitute a thin mantle over shallow bedrock, resulting in a topographically smooth appearance in a shaded relief map rather than the common rough topographic expression of exposed outcrops. Since the areal extent of bedrock outcrops serves as a predictor of various biotic and abiotic properties, a few studies have explored automating the delineation of bedrock outcrops by leveraging the spectral characteristics of exposed rock using high-resolution panoramic photographs (DiBiase et al, 2012) and image classification techniques (Scarpone et al, 2017). Additionally, Milodowski et al (2015) created a roughness metric based on the rough expression and slope associated with bedrock demonstrating that the local variability of surface normal vectors can be used as a topographic signature to identify rock exposure.…”

Section: Introductionmentioning

confidence: 99%

Predictive modeling of bedrock outcrops and associated shallow soil in upland glaciated landscapes

et al. 2020

View full text Add to dashboard Cite

Identifying the areal extent of bedrock outcrops and shallow soils has important implications for understanding spatial patterns in vegetation composition and productivity, stream chemistry gradients, and hydrologic and soil properties of landscapes. Manual methods of delineating bedrock outcrops and associated shallow soils are still commonly employed, but they are expensive to implement over broad areas and often limited by representation of polygon units. Few studies have automated the delineation of bedrock outcrops. These focused on delineation approaches in landscapes with rapidly eroding hillslopes and sparse vegetation. The objectives of this study were to assess the accuracy of visually interpreting high-resolution relief maps for locating bedrock outcrops and associated shallow soil (BOSS) < 50 cm deep in a heavily forested landscape, to use visually interpreted point locations to train predictive models, and to compare predictions with manually delineated polygons in upland glaciated landscapes. Visual interpretation of Lidar-derived 1 m shaded relief maps at Hubbard Brook Experimental Forest (HBEF), USA resulted in a 79% accuracy of interpreting deep soil locations and 84% accuracy in distinguishing BOSS. We explored four probabilistic classifications of BOSS using multiple Lidar-derived topographic metrics as predictive variables. All four methods identified similar predictors for BOSS, including slope and topographic position indices with a 15, 100 and 200 m circular analysis window, respectively. Although all classifiers yielded similar results with little difference in interpretation, a generalized additive model had slightly higher accuracy predicting BOSS presence, yielding 85% overall accuracy using independent validation data across the primary study area, and 86% overall accuracy in a second validation area.

show abstract

“…Once the accuracy is influenced by the choice of features, the use of many rank indices is reasonable in order to assure that the most accurate subset will be obtained (Novakovic et al, 2011). Another important aspect of identifying the main variables is the time saved in the acquisition and preparation of the database and computational efficiency, if there is interest in applying such models to larger and similar areas (Scarpone et al, 2017;Yu et al, 2016).…”

Section: Model Evaluationmentioning

confidence: 99%

Soil type spatial prediction from Random Forest: different training datasets, transferability, accuracy and uncertainty assessment

Machado

Silva

Curi

et al. 2019

Sci. agric. (Piracicaba, Braz.)

View full text Add to dashboard Cite

Different uses of soil legacy data such as training dataset as well as the selection of soil environmental covariables could drive the accuracy of machine learning techniques. Thus, this study evaluated the ability of the Random Forest algorithm to predict soil classes from different training datasets and extrapolate such information to a similar area. The following training datasets were extracted from legacy data: a) point data composed of 53 soil samples; b) 30 m buffer around the soil samples, and soil map polygons excluding: c) 20 m; and d) 30 m from the boundaries of polygons. These four datasets were submitted to principal component analysis (PCA) to reduce multidimensionality. Each dataset derived a new one. Different combinations of predictor variables were tested. A total of 52 models were evaluated by means of error of models, prediction uncertainty and external validation for overall accuracy and Kappa index. The best result was obtained by reducing the number of predictors with the PCA along with information from the buffer around the points. Although Random Forest has been considered a robust spatial predictor model, it was clear it is sensitive to different strategies of selecting training dataset. Effort was necessary to find the best training dataset for achieving a suitable level of accuracy of spatial prediction. To identify a specific dataset seems to be better than using a great number of variables or a large volume of training data. The efforts made allowed for the accurate acquisition of a mapped area 15.5 times larger than the reference area.

show abstract

Semi-automated classification of exposed bedrock cover in British Columbia's Southern Mountains using a Random Forest approach

Cited by 16 publications

References 46 publications

A multimethod approach for county-scale geospatial analysis of emerging infectious diseases: a cross-sectional case study of COVID-19 incidence in Germany

A multimethod approach for county-scale geospatial analysis of emerging infectious diseases: a cross-sectional case study of COVID-19 incidence in Germany

Predictive modeling of bedrock outcrops and associated shallow soil in upland glaciated landscapes

Soil type spatial prediction from Random Forest: different training datasets, transferability, accuracy and uncertainty assessment

Contact Info

Product

Resources

About