Estimativa da altura do dossel e predição do acamamento em parcelas de soja via análise de imagens
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Viçosa
Abstract
As avaliações da altura de plantas e da ocorrência e severidade do acamamento em parcelas de soja são essenciais para programas de melhoramento. No entanto, a avaliação tradicional dessas características é laboriosa e muitas vezes baseada em escalas visuais. O sensoriamento remoto via veículos aéreos não tripulados (VANT) é uma alternativa precisa e eficiente para avaliações fenotípicas. Modelos de Random Forest (MRF) são algoritmos de aprendizado de máquina que possibilitam a classificação de parcelas experimentais com base em padrões em variáveis preditivas. Os objetivos do trabalho foram avaliar o potencial do uso de modelos digitais de elevação, estimados via análise de imagens, na estimação da altura do dossel nos estádios de florescimento e maturação, e avaliar o potencial de MRF em detectar e classificar parcelas de soja quanto à severidade do acamamento através de séries temporais de alturas do dossel. Dois ensaios com um total de 307 genótipos de soja foram avaliados quanto ao número de dias até o florescimento (DIASF) e até a maturação (DIASM), comprimento de plantas no florescimento (COMPF) e na maturação (COMPM) e nível de acamamento (AC). O AC foi estimado por uma escala de níveis correspondentes ao ângulo de inclinação da haste principal. Imagens aéreas foram adquiridas através de 19 voos com um VANT com sensor RGB, entre o florescimento e a maturação. Modelos digitais de altura do dossel foram estimados, e a altura do dossel estimada (ALTDI) foi determinada pelos percentis 50, 75, 90, 95, 99 e pelo intervalo de altura predominante. As ALTDI de voos próximos ao DIASF e ao DIASM foram comparadas com COMPF e com a altura do dossel calculada a partir de COMPM e AC, respectivamente. A acurácia das ALTDI foi mensurada pela raiz do erro quadrático médio (RMSE) e raiz do erro quadrático médio percentual (RMSE%). Séries temporais foram construídas com base nas ALTDIs de cada parcela e curvas logísticas foram ajustadas às séries temporais. Os parâmetros dessas curvas foram usados para treinar os MRFs para a classificação de AC, a partir de cada método de estimação de ALTDI (ME_ALTDI). Os MRFs foram avaliados através da acurácia balanceada, F1-Score e coeficiente Kappa. O percentil 50 estimou a altura do dossel no florescimento com acurácia satisfatória (RMSE = 12,9 e RMSE% = 13,5%). O percentil 99 foi superior na estimação da altura do dossel na maturação (RMSE = 28,8 e RMSE% = 18,7%), mas houve um comprometimento da acurácia da ALTDI em parcelas com plantas predominantemente eretas devido à redução da superfície das plantas causada pela senescência e queda de folhas causada pela maturação. Os MRFs treinados a partir do percentil 50 mostraram maior potencial na classificação de AC (acurácia balanceada = 0,5494, F1-Score = 0,2360 e coeficiente Kappa = 0,0876) e na detecção do acamamento (acurácia balanceada = 0,5833, F1-Score = 0,5740 e coeficiente Kappa = 0,1630), mas, devido à baixa acurácia das ALTDI próximas à maturação, não mostraram desempenho suficiente para justificar sua recomendação. A metodologia pode ter maior sucesso com sensores que representem as alturas das plantas de forma mais acurada e detalhada. Palavras-chave: Glycine max. RGB. Surface from Motion. Série temporal. Curva logística. Random Forest.
Assessing plant height and the occurrence and severity of lodging in soybean plots is essential for breeding programs. However, traditional assessment methods are labor-intensive, and often based on visual scales. Remote sensing via unmanned aerial vehicles (UAVs) offers an accurate and efficient alternative for phenotypic evaluations. Random forest models (RFM) are machine learning algorithms that enable the classification of experimental plots based on patterns in predictive variables. The objective of this study was to evaluate the potential of using digital elevation models, estimated through image analysis, for canopy height estimation at flowering and maturity stages, and to assess the potential of RFM in detecting lodging and classifying its severity in soybean plots through temporal series of canopy heights. Two trials, comprising a total of 307 soybean genotypes, were evaluated for days to flowering (DAYSF), days to maturation (DAYSM), plant length at flowering (LENGF), plant length at maturity (LENGM), and lodging severity (LS). LS was estimated on a scale of levels corresponding to the angle of inclination of the main stem. Aerial images were acquired through 19 UAV flights with an RGB sensor between flowering and maturity. Canopy height models were estimated, and the estimated canopy height (ECH) was determined using the 50th, 75th, 90th, 95th, 99th percentiles, and the predominant height range. ECH from flights executed close to DAYSF and DAYSM were compared with LF and canopy height calculated from LENGM and LS, respectively. ECH accuracy was measured using the root mean square error (RMSE) and RMSE%. Temporal series were constructed based on the ECH, and logistic curves were fitted to the series. Parameters from these curves were used to train RFM for LS classification, using each ALTDI estimator. RFMs were evaluated using balanced accuracy, F1-Score, and Kappa coefficient. The 50th percentile accurately estimated canopy height at flowering (RMSE = 12.9 and RMSE% = 13.5%). The 99th percentile was superior in estimating canopy height at maturity (RMSE = 28.8 and RMSE% = 18.7%), but ECH accuracy in plots with predominantly erect plants was compromised due to reduced plant surface area caused by leaf senescence and fall at maturity. RFM trained using the 50th percentile showed greater potential in LS classification (balanced accuracy = 0.5494, F1-Score = 0.2360, and Kappa = 0.0876) and lodging detection (balanced accuracy = 0.5833, F1-Score = 0.5740, and Kappa = 0.1630), but due to low ECH accuracy near maturity, they did not perform sufficiently well to justify their recommendation. The methodology may be more successful with sensors that are able to describe plant heights with more details and more accurately. Keywords: Glycine max. RGB. Surface from Motion. Temporal series. Logistic curve. Random Forest.
Assessing plant height and the occurrence and severity of lodging in soybean plots is essential for breeding programs. However, traditional assessment methods are labor-intensive, and often based on visual scales. Remote sensing via unmanned aerial vehicles (UAVs) offers an accurate and efficient alternative for phenotypic evaluations. Random forest models (RFM) are machine learning algorithms that enable the classification of experimental plots based on patterns in predictive variables. The objective of this study was to evaluate the potential of using digital elevation models, estimated through image analysis, for canopy height estimation at flowering and maturity stages, and to assess the potential of RFM in detecting lodging and classifying its severity in soybean plots through temporal series of canopy heights. Two trials, comprising a total of 307 soybean genotypes, were evaluated for days to flowering (DAYSF), days to maturation (DAYSM), plant length at flowering (LENGF), plant length at maturity (LENGM), and lodging severity (LS). LS was estimated on a scale of levels corresponding to the angle of inclination of the main stem. Aerial images were acquired through 19 UAV flights with an RGB sensor between flowering and maturity. Canopy height models were estimated, and the estimated canopy height (ECH) was determined using the 50th, 75th, 90th, 95th, 99th percentiles, and the predominant height range. ECH from flights executed close to DAYSF and DAYSM were compared with LF and canopy height calculated from LENGM and LS, respectively. ECH accuracy was measured using the root mean square error (RMSE) and RMSE%. Temporal series were constructed based on the ECH, and logistic curves were fitted to the series. Parameters from these curves were used to train RFM for LS classification, using each ALTDI estimator. RFMs were evaluated using balanced accuracy, F1-Score, and Kappa coefficient. The 50th percentile accurately estimated canopy height at flowering (RMSE = 12.9 and RMSE% = 13.5%). The 99th percentile was superior in estimating canopy height at maturity (RMSE = 28.8 and RMSE% = 18.7%), but ECH accuracy in plots with predominantly erect plants was compromised due to reduced plant surface area caused by leaf senescence and fall at maturity. RFM trained using the 50th percentile showed greater potential in LS classification (balanced accuracy = 0.5494, F1-Score = 0.2360, and Kappa = 0.0876) and lodging detection (balanced accuracy = 0.5833, F1-Score = 0.5740, and Kappa = 0.1630), but due to low ECH accuracy near maturity, they did not perform sufficiently well to justify their recommendation. The methodology may be more successful with sensors that are able to describe plant heights with more details and more accurately. Keywords: Glycine max. RGB. Surface from Motion. Temporal series. Logistic curve. Random Forest.
Description
Citation
LIMA, Nathan Lamounier. Estimativa da altura do dossel e predição do acamamento em parcelas de soja via análise de imagens. 2024. 90 f. Tese (Doutorado em Fitotecnia) - Universidade Federal de Viçosa, Viçosa. 2024.
