Yield predict and physiological state evaluation of irrigated common bean cultivars with contrasting growth habits by learning algorithms using spectral indices

• Autor(a) principal: Coelho, Anderson Prates [UNESP]
• Data de Publicação: 2022
• Outros Autores: Faria, Rogério Teixeira de [UNESP], Lemos, Leandro Borges [UNESP], Rosalen, David Luciano [UNESP], Dalri, Alexandre Barcellos [UNESP]
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1080/10106049.2022.2096700; http://hdl.handle.net/11449/242013
• topic: artificial neural networks; NDVI; Phaseolus vulgarisL; portable chlorophyll meter; remote sensing

This study aimed to analyze and compare the accuracy of models to predict the grain yield (GY) of common bean cultivars with contrasting growth habits using spectral indices. The common bean cultivars used were IAC Imperador and IPR Campos Gerais, which have determinate and indeterminate growth habits, respectively. The plants were grown under five irrigation levels (54, 70, 77, 100, and 132% of the crop evapotranspiration) to generate variability. The normalized difference vegetation (NDVI) and leaf chlorophyll (LCI) indexes were measured at the following phenological stages: V4 (third trifoliate leaf), R5 (pre-flowering), R6 (full flowering), and R8 (grain filling). The spectral indices were used individually for each phenological stage and associated with simple and multiple regressions (SLR and MLR) and artificial neural networks (ANN) to predict GY. Then, stratified models by cultivar and general models were established using data from both cultivars. The accuracy of NDVI-based GY predictions for both models at R6 phenological stage (ANN and SLR average) was acceptable (R2 = 0.64; RMSE = 0.37 Mg ha−1; MBE = −0.14 Mg ha−1) but poor for LCI predictions. The highest accuracies were observed at reproductive phenological stages, mainly R6. The ANNs algorithm did not show superior GY prediction accuracy compared to SLR. NDVI-based remote sensing is feasible to predict and monitor common bean yield potential using cultivar-specific and general models.