Biomass estimation of spring wheat with machine learning methods using UAV-based multispectral imaging
João Gustavo Atkinson Amorim, Lincoln Vinicius Schreiber, Mirayr Raul Quadros de Souza, Marcelo Negreiros, Altamiro Susin, Christian Bredemeier, Carolina Trentin, André Luis Vian, Clódis de Oliveira Andrades Filho, D. Doering, Adriane Parraga
Abstract
Remote biomass estimation can benefit agricultural practices in several ways, especially larger areas since it does not require local measurements. The advances of the last few decades in machine learning techniques have created new possibilities for estimating aboveground biomass. A pipeline was established from image acquisition to modelling shoot biomass of two wheat cultivars used in Southern Brazil (TBIO Toruk and BRS Parrudo). A UAV was used to acquire multispectral images with high spatial resolution to calculate vegetation indices (VIs). These VIs along with machine learning approaches are used to model the measured biomass of crops in different growth phases. To correlate the wheat images with measured shoot dry biomass, the following regression models were investigated: random forest, support vector regression, and artificial neural networks. An experiment was designed and conducted at the Agriculture Experimental Station of the Federal University of Rio Grande do Sul (EEA/UFRGS) to assess wheat growth. Variability in crop growth was created for all test areas by varying nitrogen availability. To determine shoot biomass, plants were sampled at three different crop growth stages: V6 (stage of six fully developed leaves), three nodes, and flowering. Our results indicate the importance of the radiometric calibration used. Also, the features extracted from images, such as the VIs combined with machine learning models can be used in precision agriculture for predicting the spatial variability of shoot biomass. The best model for Brazilian wheat cultivars was an artificial neural network with R2 of 0.90, RMSE of 0.83t/ha, and nRMSE of 8.95%. We also found a strong correlation between ground NDVI with image-based NDVI, with an R2 of 0.84.