This is a repository for the botanical composition model developed in the paper "Estimating the botanical composition of bahiagrass–rhizoma peanut pastures using aerial multispectral imagery and deep learning".
If you use the code or model, please cite our paper as follows:
Bretas, I. L., Zhao, C., Dubeux Jr, J. C., Xin, Y., Tang, Z., Trumpp, K. R., ... & de Souza, C. H. (2026). Estimating the botanical composition of bahiagrass–rhizoma peanut pastures using aerial multispectral imagery and deep learning. Crop Science, 66(2), e70253.
Our research is highlighted in the CSA news on May 26, 2026: https://www.sciencesocieties.org/publications/csa-news/2026/june/new-approach-to-estimate-botanical-composition?q=publications/csa-news/2026/june/new-approach-to-estimate-botanical-composition/
Estimating the botanical composition in mixed pastures is challenging due to resource-intensive sampling methods overlooking grassland spatial variability. We aimed to (I) classify bahiagrass (Paspalum notatum Flügge) and rhizoma peanut (Arachis glabrata Benth.) in grass–legume pastures using aerial multispectral imagery and machine learning, (II) estimate the proportion of grass and legumes on an area basis, and (III) estimate the botanical composition on a dry matter basis to guide grazing management. Unmanned aerial vehicle-based images were taken in the early and late warm seasons of 2023 and 2024. Field campaigns were conducted for ground-truth species labeling (n = 554) across nine paddocks under three treatments (fertilized grass, unfertilized grass, and grass–legume mixed) using a high-precision positioning system. Ninety-six points were harvested and manually sorted for ground truth botanical composition estimation on a dry matter basis. The classification dataset was split into training and testing using three different approaches to assess spatial generalization performance. A random forest (RF) was employed as the base model, and a convolutional neural network (CNN) was developed using fivefold cross-validation on the training set. Our CNN model outperformed the RF model, achieving an average cross-validated overall accuracy of 93% and 88% for grass–legume classification on the grid-based and repeated spatial hold-out testing set. Combining land cover classification, compressed height, and vegetation indices showed promising potential to predict total biomass (R2 = 0.85) and grass biomass (R2 = 0.83) in bahiagrass–rhizoma peanut pastures using stepwise multiple linear regression. This approach also enabled the estimation of legume proportion by difference.