Modelling species presence‐only data with random forests

Ecography, Год журнала: 2021, Номер 44(12), С. 1731 - 1742

Опубликована: Окт. 27, 2021

The random forest (RF) algorithm is an ensemble of classification or regression trees and widely used, including for species distribution modelling (SDM). Many researchers use implementations RF in the R programming language with default parameters to analyse presence‐only data together ‘background' samples. However, there good evidence that does not perform well such ‘presence‐background' modelling. This often attributed disparity between number presence background samples, also known as 'class imbalance', several solutions have been proposed. Here, we first set context: sample should be large enough represent all environments region. We then aim understand drivers poor performance when models are fitted alongside show overlap' (where both classes occur same environment) important driver performance, class imbalance. Class overlap can even degrade presence–absence data. explain, test evaluate suggested solutions. Using simulated real presence‐background data, compare other weighting sampling approaches. Our results demonstrate clear improvement RFs techniques explicitly manage imbalance used. these either limit enforce tree depth. Without compromising environmental representativeness sampled background, identify approaches fitting ameliorate effects allow excellent predictive performance. Understanding problems allows new insights into how best fit models, guide future efforts deal

Язык: Английский

---

MaxEnt brings comparable results when the input data are being completed; Model parameterization of four species distribution models

Ecology and Evolution, Год журнала: 2023, Номер 13(2)

Опубликована: Фев. 1, 2023

Abstract Species distribution models (SDMs) are practical tools to assess the habitat suitability of species with numerous applications in environmental management and conservation planning. The manipulation input data deal their spatial bias is one advantageous methods enhance performance SDMs. However, development a model parameterization approach covering different SDMs achieve well‐performing has rarely been implemented. We integrated tuning for four commonly‐used SDMs: generalized linear (GLM), gradient boosted (GBM), random forest (RF), maximum entropy (MaxEnt), compared predictive geographically imbalanced‐biased rare complex mountain vipers. Models were tuned up based on range model‐specific parameters considering two background selection methods: weighting schemes. fine‐tuned was assessed recently identified localities species. results indicated that although version all shows great predicting training (AUC > 0.9 TSS 0.5), they produce classifying out‐of‐bag data. GBM RF higher sensitivity showed more performances. GLM, despite having high test data, lower specificity. It only MaxEnt comparable identifying both procedures. Our highlight while prone overfitting GLM over‐predict nonsampled areas capable producing predictable (extrapolative) (interpolative). discuss assumptions each conclude could be considered as method cope modeling approaches.

Язык: Английский

---

Reducing global land-use pressures with seaweed farming

Nature Sustainability, Год журнала: 2023, Номер 6(4), С. 380 - 390

Опубликована: Янв. 26, 2023

Язык: Английский

---

Integrating citizen science and spatial ecology to inform management and conservation of the Italian seahorses

Ecological Informatics, Год журнала: 2023, Номер 79, С. 102402 - 102402

Опубликована: Дек. 1, 2023

Citizen science and spatial ecology analyses can inform species distributions, habitat preferences, threats in elusive endangered such as seahorses. Through a dedicated citizen survey submitted to the Italian diving centers, we collected 115 presence records of two seahorses occurring along coasts: Hippocampus hippocampus H. guttulatus. From this dataset, used 85 seahorse valitaded identify ecological features these poorly known quantify effects human activities on their suitability through geographic information systems distribution modelling. Our results indicated continuous suitable area for both coasts, with single major gap central Adriatic Sea (Emilia-Romagna Marche regions). They co-occurred most range, particularly southern Tyrrhenian niches resulted be significantly similar, although not equivalent. The least-cost paths were concentrated Italy (Apulia, Calabria, Sicily), suggesting that more data is needed improve resolution available information, especially northern Italy. Human influenced 35% 41% guttulatus, respectively, while only 25% 30% potential are protected by Italy's existing conservation system, accordance global average In particular, represents critical where occurrence lower anthropic impact higher. Considering all regions, fishing effort main activity impacting species. These findings will support implementation efficient actions. We encourage application interaction facilitate assessment sustainable management organisms.

Язык: Английский

---

Top ten hazards to avoid when modeling species distributions: a didactic guide of assumptions, problems, and recommendations

Ecography, Год журнала: 2024, Номер 2024(4)

Опубликована: Янв. 31, 2024

Species distribution models, also known as ecological niche models or habitat suitability have become commonplace for addressing fundamental and applied biodiversity questions. Although the field has progressed rapidly regarding theory implementation, key assumptions are still frequently violated recommendations inadvertently overlooked. This leads to poor being published used in real‐world applications. In a structured, didactic treatment, we summarize what our view constitute ten most problematic issues, hazards, negatively affecting implementation of correlative approaches species modeling (specifically those that model by comparing environments species' occurrence records with background pseudoabsence sample). For each hazard, state relevant assumptions, detail problems arise when violating them, convey straightforward existing recommendations. We discuss five major outstanding questions active current research. hope this contribution will promote more rigorous these valuable stimulate further advancements.

Язык: Английский

---

Exploring forest fire susceptibility and management strategies in Western Himalaya: Integrating ensemble machine learning and explainable AI for accurate prediction and comprehensive analysis

Environmental Technology & Innovation, Год журнала: 2024, Номер 35, С. 103655 - 103655

Опубликована: Май 5, 2024

Forest fires pose a significant threat to ecosystems and socio-economic activities, necessitating the development of accurate predictive models for effective management mitigation. In this study, we present novel machine learning approach combined with Explainable Artificial Intelligence (XAI) techniques predict forest fire susceptibility in Nainital district. Our innovative methodology integrates several robust — AdaBoost, Gradient Boosting Machine (GBM), XGBoost Random Deep Neural Network (DNN) as meta-model stacking framework. This not only utilises individual strengths these models, but also improves overall prediction performance reliability. By using XAI techniques, particular SHAP (SHapley Additive exPlanations) LIME (Local Interpretable Model-agnostic Explanations), improve interpretability provide insights into decision-making processes. results show effectiveness ensemble model categorising different zones: very low, moderate, high high. particular, identified extensive areas susceptibility, precision, recall F1 values underpinning their effectiveness. These achieved ROC AUC above 0.90, performing exceptionally well an 0.94. The are remarkably inclusion confidence intervals most important metrics all emphasises robustness reliability supports practical use management. Through summary plots, analyze global variable importance, revealing annual rainfall Evapotranspiration (ET) key factors influencing susceptibility. Local analysis consistently highlights importance rainfall, ET, distance from roads across models. study fills research gap by providing comprehensive interpretable modelling that our ability effectively manage risk is consistent environmental protection sustainable goals.

Язык: Английский

---

Projecting Future Wetland Dynamics Under Climate Change and Land Use Pressure: A Machine Learning Approach Using Remote Sensing and Markov Chain Modeling

Remote Sensing, Год журнала: 2025, Номер 17(6), С. 1089 - 1089

Опубликована: Март 20, 2025

Wetlands in the Yellow River Watershed of Inner Mongolia face significant reductions under future climate and land use scenarios, threatening vital ecosystem services water security. This study employs high-resolution projections from NASA’s Global Daily Downscaled Projections (GDDP) Intergovernmental Panel on Climate Change Sixth Assessment Report (IPCC AR6), combined with a machine learning Cellular Automata–Markov (CA–Markov) framework to forecast cover transitions 2040. Statistically downscaled temperature precipitation data for two Shared Socioeconomic Pathways (SSP2-4.5 SSP5-8.5) are integrated satellite-based (Landsat, Sentinel-1) 2007 2023, achieving high classification accuracy (over 85% overall, Kappa > 0.8). A Maximum Entropy (MaxEnt) analysis indicates that rising temperatures, increased variability, urban–agricultural expansion will exacerbate hydrological stress, driving substantial wetland contraction. Although certain areas may retain or slightly expand their wetlands, dominant trend underscores urgency spatially targeted conservation. By synthesizing data, multi-temporal transitions, ecological modeling, this provides insights adaptive resource planning management ecologically sensitive regions.

Язык: Английский

---

Paleorecords Reveal Biological Mechanisms Crucial for Reliable Species Range Shift Projections Amid Rapid Climate Change

Ecology Letters, Год журнала: 2025, Номер 28(2)

Опубликована: Фев. 1, 2025

ABSTRACT The recent acceleration of global climate warming has created an urgent need for reliable projections species distributions, widely used by natural resource managers. Such have been mainly produced distribution models with little information on their performances in novel climates. Here, we hindcast the range shifts forest tree across Europe over last 12,000 years to compare reliability three different types models. We show that most climatically dissimilar conditions, process‐explicit (PEMs) tend outperform correlative (CSDMs), and PEM are likely be more than those made CSDMs end 21st century. These results demonstrate first time often promoted albeit so far untested idea explicit description mechanisms confers model robustness, highlight a new avenue increase projection future.

Язык: Английский

---

Forest tree species distribution for Europe 2000–2020: mapping potential and realized distributions using spatiotemporal machine learning

PeerJ, Год журнала: 2022, Номер 10, С. e13728 - e13728

Опубликована: Июль 25, 2022

This article describes a data-driven framework based on spatiotemporal machine learning to produce distribution maps for 16 tree species ( Abies alba Mill., Castanea sativa Corylus avellana L., Fagus sylvatica Olea europaea Picea abies L. H. Karst., Pinus halepensis nigra J. F. Arnold, pinea sylvestris Prunus avium Quercus cerris ilex robur suber and Salix caprea L.) at high spatial resolution (30 m). Tree occurrence data total of three million points was used train different algorithms: random forest, gradient-boosted trees, generalized linear models, k-nearest neighbors, CART an artificial neural network. A stack 305 coarse covariates representing spectral reflectance, biophysical conditions biotic competition as predictors realized distributions, while potential modelled with environmental only. Logloss computing time were select the best algorithms tune ensemble model stacking logistic regressor meta-learner. An trained each species: probability uncertainty produced using window 4 years six per species, distributions only one map produced. Results cross validation show that consistently outperformed or performed good individual in both tasks, models achieving higher predictive performances (TSS = 0.898, R 2 logloss 0.857) than ones average 0.874, 0.839). Ensemble Q. achieved 0.968, 0.952) 0.959, 0.949) distribution, P. 0.731, 0.785, 0.585, 0.670, respectively, distribution) 0.658, 0.686, 0.623, 0.664) worst. Importance predictor variables differed across green band summer Normalized Difference Vegetation Index (NDVI) fall diffuse irradiation precipitation driest quarter (BIO17) being most frequent important distribution. On average, fine-resolution (250 m) +6.5%, +7.5%). The shows how combining continuous consistent Earth Observation series state art can be derive dynamic maps. predictions quantify temporal trends forest degradation composition change.

Язык: Английский

---

Predicted range shifts of invasive giant hogweed (Heracleum mantegazzianum) in Europe

The Science of The Total Environment, Год журнала: 2022, Номер 825, С. 154053 - 154053

Опубликована: Фев. 23, 2022

Язык: Английский

---