Random subset feature selection for ecological niche models of wildfire activity in Western North America

Hemlock woolly adelgid niche models from the invasive eastern North American range with projections to native ranges and future climates

“…The "PresenceAbsence" package by Freeman & Moisen (2008) was used with presence and pseudoabsence data to calculate pseudoabsence-based (psa) versions of the True Skill Statistic (TSSpsa) and Area under the Curve statistic (AUCpsa). To reduce MaxEnt model complexity and overfitting for improved model generalization, the MaxEnt beta regularization was adjusted to two and only quadratic and hinge features were used (Warren & Seifert 2011, Tracy et al 2018. MaxEnt models were calibrated to binary presence/absence format using a threshold at maximum TSSpsa (Liu et al 2013).…”

Vegetation and Soil Fire Damage Analysis Based on Species Distribution Modeling Trained with Multispectral Satellite Data

“…The main advantages of MaxEnt include: (1) It has been designed to work with presence-only data; thus, it constitutes an interesting alternative to other machine learning based classifiers [34]; (2) its probabilistic output is easy to interpret and has physical meaning [35] and (3) it is a non-parametric model (the input variables interrelations are not determine a priori) [36]. MaxEnt is widely used in ecological studies to model species distributions [37][38][39] and it is increasingly being used in remote sensing based applications like pest potential distribution [40], landslide vulnerability monitoring [41], groundwater potential distribution [42] or fire occurrence modeling [43][44][45][46], so the use of MaxEnt to deal with the effects of burn severity encourages us to check it in the present study. MaxEnt provides the percentage of contribution tables and probability distributions of each target class (three burn severity levels and burned area, in our study), from which burn severity and burned area maps can be built.…”

Forest health monitoring in transition: Evaluating insect-induced disturbances in forested landscapes at varying spatial scales

“…Pseudoabsence-based versions of the True Skill Statistic and Area under the Curve statistic (AUC) for evaluation were calculated using the "PresenceAbsence" package by Freeman and Moisen (2008). In order to reduce model complexity and overfitting, the MaxEnt beta regularization was adjusted to two and only quadratic and hinge features were used (Warren and Seifert 2011;Tracy et al 2018). MaxEnt models were calibrated to binary presence/absence format using a threshold at maximum True Skill Statistic (Liu et al 2013).…”

“…To create relatively small subsets of environmental factors predicting HWA habitat suitability in the MaxEnt niche models, all the predictors (n=119) described in section 5.4.1., were screened using the random subset feature selection algorithm (RSFSA) by Tracy et al (2018). In the feature selection method, hundreds of random subsets of the potential environmental predictors of pre-specified sizes are created.…”

“…Statistical performance between RSFSA-selected and random MaxEnt models was then evaluated. This was done by generating and ranking 3000 models for the selected subset size using held out model training and testing data (Tracy et al 2018). Instead of one final MaxEnt models, statistically 12 best RSFSA-selected models were chosen, binary calibrated, and combined by frequency consensus forming feature subset ensemble (FSE) models.…”