Modelagem espacial de áreas susceptíveis a danos econômicos de Glycaspis brimblecombei em plantios de Eucalyptus spp. no território brasileiro: cenários presente e futuro
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Viçosa
Abstract
O Glycaspis brimblecombei (Hemiptera: Aphalaridae) destaca-se como uma das principais pragas exóticas do eucalipto no Brasil, anualmente infestando milhares de hectares e impactando a produtividade dos plantios. Sua rápida dispersão e adaptação têm sido evidentes, porém, ainda não está claro quais variáveis ambientais contribuem para a distribuição potencial de áreas suscetíveis à ocorrência de danos econômicos. As atuais medidas de controle se mostram insuficientes, indicando a necessidade de desenvolvimento de novas estratégias de gestão climática inteligente. Nesse contexto, este estudo teve como objetivo modelar a distribuição geográfica das áreas suscetíveis a danos econômicos causados por G. brimblecombei no território brasileiro, em cenários climáticos do presente e do futuro. No processamento inicial dos dados, foram empregados 56 preditores ambientais, (36 variáveis climáticas, 19 variáveis bioclimáticas e 1 variável topográfica). Para cada preditor, foram extraídos os valores de máximo, mínimo, mediana, desvio padrão (SD) e quartis 0,05, 0,25, 0,75 e 0,95. A seleção das variáveis preditoras foi conduzida em um processo subdividido em três etapas complementares: remoção por variância próxima a zero (1), remoção baseada na correlação e avaliação (2) da importância das variáveis (3). Sete algoritmos de machine learning amplamente utilizados foram aplicados para mapear a distribuição da espécie: Random Forest (RF), C5.0, Support Vector Machine Radial Sigma (SVM), k-Nearest Neighbors (KNN) e Model Averaged Neural Network (AVNNET), Gradient Boosting Machine (GBM) e Partial Least Squares (PLS). A avaliação do desempenho incluiu métricas de acurácia, acurácia balanceada, F1- score, sensibilidade, especificidade e Kappa. O modelo mais eficiente foi então empregado na modelagem dos cenários atuais e futuros. Os cenários climáticos IPSL- CM6A-LR, MIROC6, HadGEM3-GC31-LL e UKESM1-0-LL foram utilizados para simulações nos cenários SSP-2.45 e SSP-5.85 nos períodos de 2041-2060, 2061- 2080 e 2081–2100. As variáveis que mais influenciaram nas previsões das distribuições potenciais incluíram o quartil 0,05 da precipitação do trimestre mais seco (q05.BIO17), máxima temperatura máxima de maio (max.tmax_05), máxima temperatura máxima de setembro (max.tmax_09), máxima, temperatura máxima do mês de novembro (max.tmax_11), o mínimo da precipitação anual (min.BIO12), máxima temperatura mínima de novembro (max.tmin_11) e mínima precipitação de novembro (min.prec_11). Embora todos os setes modelos tenham apresentado previsões de alta precisão, o RF demonstrou o melhor desempenho na maioria das métricas avaliativas. No cenário presente, as regiões Sudeste, Sul e Centro-Oeste concentram as maiores áreas suscetíveis a danos econômicos, enquanto o Norte e o Nordeste exibem as menores extensões. Para os cenários futuros, o modelo prevê uma redução nas áreas potenciais para ocorrência de danos econômicos, as áreas que prevalecem suscetíveis se concentram no Sudeste e Sul do território. Palavras-chave: Aprendizado de máquina; Danos econômicos; Pragas florestais; Modelo de distribuição de espécie; Mudanças climáticas.
The Glycaspis brimblecombei (Hemiptera: Aphalaridae) stands out as one of the main exotic pests of eucalyptus in Brazil, annually infesting thousands of hectares and impacting plantation productivity. Its rapid dispersion and adaptation have been evident, yet it is still unclear which environmental variables contribute to the potential distribution of areas susceptible to economic damage occurrence. Current control measures are proving insufficient, indicating the need for the development of new strategies for intelligent climate management. In this context, this study aims to model the geographical distribution of areas susceptible to economic damage caused by G. brimblecombei in the Brazilian territory, under present and future climate scenarios. In the initial data processing, 56 environmental predictors were employed (36 climatic variables, 19 bioclimatic variables, and 1 topographic variable). For each predictor, maximum, minimum, median, standard deviation (SD), and quartiles 0.05, 0.25, 0.75, and 0.95 values were extracted. The selection of predictor variables was conducted in a process subdivided into three complementary stages: removal by near-zero variance (1), removal based on correlation and evaluation (2) of variable importance (3). Seven widely used machine learning algorithms were applied to map the species distribution: Random Forest (RF), C5.0, Support Vector Machine Radial Sigma (SVM), k-Nearest Neighbors (KNN), Model Averaged Neural Network (AVNNET), Gradient Boosting Machine (GBM), and Partial Least Squares (PLS). Performance evaluation included accuracy metrics, balanced accuracy, F1-score, sensitivity, specificity, and Kappa. The most efficient model was then employed in modeling current and future scenarios. The climate scenarios IPSL-CM6A-LR, MIROC6, HadGEM3-GC31-LL, and UKESM1-0-LL were used for simulations in the SSP-2.45 and SSP-5.85 scenarios for the periods 2041-2060, 2061-2080, and 2081–2100. The variables that most influenced the potential distribution predictions included the 0.05 quartile of precipitation in the driest quarter (q05.BIO17), maximum maximum temperature of May (max.tmax_05), maximum maximum temperature of September (max.tmax_09), maximum maximumtemperature of November (max.tmax_11), minimum annual precipitation (min.BIO12), maximum minimum temperature of November (max.tmin_11), and minimum precipitation of November (min.prec_11). Although all seven models showed predictions of high accuracy, RF demonstrated the best performance in most evaluative metrics. In the present scenario, the Southeast, South, and Midwest regions concentrate the largest areas susceptible to economic damage, while the North and Northeast exhibit the smallest extents. For future scenarios, the model predicts a reduction in potential areas for economic damage occurrence, with the remaining susceptible areas concentrated in the Southeast and South of the territory. Keywords: Machine learning; Economic damage; Forest pests; Species distribution model; Climate change.
The Glycaspis brimblecombei (Hemiptera: Aphalaridae) stands out as one of the main exotic pests of eucalyptus in Brazil, annually infesting thousands of hectares and impacting plantation productivity. Its rapid dispersion and adaptation have been evident, yet it is still unclear which environmental variables contribute to the potential distribution of areas susceptible to economic damage occurrence. Current control measures are proving insufficient, indicating the need for the development of new strategies for intelligent climate management. In this context, this study aims to model the geographical distribution of areas susceptible to economic damage caused by G. brimblecombei in the Brazilian territory, under present and future climate scenarios. In the initial data processing, 56 environmental predictors were employed (36 climatic variables, 19 bioclimatic variables, and 1 topographic variable). For each predictor, maximum, minimum, median, standard deviation (SD), and quartiles 0.05, 0.25, 0.75, and 0.95 values were extracted. The selection of predictor variables was conducted in a process subdivided into three complementary stages: removal by near-zero variance (1), removal based on correlation and evaluation (2) of variable importance (3). Seven widely used machine learning algorithms were applied to map the species distribution: Random Forest (RF), C5.0, Support Vector Machine Radial Sigma (SVM), k-Nearest Neighbors (KNN), Model Averaged Neural Network (AVNNET), Gradient Boosting Machine (GBM), and Partial Least Squares (PLS). Performance evaluation included accuracy metrics, balanced accuracy, F1-score, sensitivity, specificity, and Kappa. The most efficient model was then employed in modeling current and future scenarios. The climate scenarios IPSL-CM6A-LR, MIROC6, HadGEM3-GC31-LL, and UKESM1-0-LL were used for simulations in the SSP-2.45 and SSP-5.85 scenarios for the periods 2041-2060, 2061-2080, and 2081–2100. The variables that most influenced the potential distribution predictions included the 0.05 quartile of precipitation in the driest quarter (q05.BIO17), maximum maximum temperature of May (max.tmax_05), maximum maximum temperature of September (max.tmax_09), maximum maximumtemperature of November (max.tmax_11), minimum annual precipitation (min.BIO12), maximum minimum temperature of November (max.tmin_11), and minimum precipitation of November (min.prec_11). Although all seven models showed predictions of high accuracy, RF demonstrated the best performance in most evaluative metrics. In the present scenario, the Southeast, South, and Midwest regions concentrate the largest areas susceptible to economic damage, while the North and Northeast exhibit the smallest extents. For future scenarios, the model predicts a reduction in potential areas for economic damage occurrence, with the remaining susceptible areas concentrated in the Southeast and South of the territory. Keywords: Machine learning; Economic damage; Forest pests; Species distribution model; Climate change.
Description
Citation
FERNANDES, André Pereira. Modelagem espacial de áreas susceptíveis a danos econômicos de Glycaspis brimblecombei em plantios de Eucalyptus spp. no território brasileiro: cenários presente e futuro. 2024. 142 f. Dissertação (Mestrado em Ciência Florestal) - Universidade Federal de Viçosa, Viçosa. 2024.
