Classification of soybean genotypes for industrial traits using UAV multispectral imagery and machine learning

• Autor(a) principal: Santana, Dthenifer Cordeiro [UNESP]
• Data de Publicação: 2023
• Outros Autores: Teodoro, Larissa Pereira Ribeiro, Baio, Fábio Henrique Rojo, Santos, Regimar Garcia dos [UNESP], Coradi, Paulo Carteri, Biduski, Bárbara, Silva Junior, Carlos Antonio da, Teodoro, Paulo Eduardo [UNESP], Shiratsuchi, Luaciano Shozo
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1016/j.rsase.2023.100919; http://hdl.handle.net/11449/246601
• topic: Computational intelligence; High-throughput phenotyping; Precision agri-culture; Spectral bands; Vegetation indices

Soybean genotypes have distinct physicochemical characteristics, mainly regarding the oil and protein contents in the grains. The use of high-throughput phe-notyping technologies allied to data processing by machine learning algorithms facili-tates and can make it faster and more precise to obtain information about the charac-teristics of the grains. Thus, the objective of the study was to identify machine learning algorithms and inputs with better performance for classifying genotypes clustered based on industrial traits. The experiment was implemented in a randomized block design with two replicates. 103 F2 soybean populations were evaluated. Red, green, near-infrared, and infrared spectral bands and the vegetation indices NDVI, NDRE, GNDVI, SAVI, MSAVI, MCARI, EVI, and SCCCI were measured using UAV multispectral imagery. The industrial traits evaluated were: crude protein content, oil yield, and ash and fiber contents. Data were subjected to Pearson correlation analysis expressed by a correlation network. A genotype clustering based on industrial traits was performed using PCA and k-means algorithm, and then the clusters formed were used as output variables of the ML models, while three input configurations were tested: only spectral bands (B), only vegetation indices (VIs), and B + VIs. ML algorithms tested were: artificial neural net-work (ANN), decision tree algorithms J48 (J48), REPTree (DT), and RandomTree (Rt), random forest (RF), Support Vector Machine (SVM), and logistic regression (LR, used as control). Statistical metrics used to evaluate the accuracy of the models were per-centage of correct classification (CC) and F-score. ML algorithms that achieved the highest classification accuracies were ANN, DT and SVM. As for the inputs tested, the best results were obtained using only spectral bands.