Deep Learning-based RGB Image Modelling for Multi-component Pasture Biomass Estimation in Precision Grazing

Background: Accurate estimation of pasture biomass is significant for managing grazing, feed distribution and overall livestock productivity sustainably. The classical methods of cutting and weighing are the most trustworthy worst-case scenario but they are also foul slow, demanding much labour and are not useful for very large areas. Methods: This research intended to construct a minimalist deep-learning framework that can accurately estimate various pasture biomass components from inexpensive RGB images and corresponding field data. The material used for the experiment included 1,785 tree-top images taken from Australian pastures, all linked to ground measurements of biomass fractions, vegetation indices and plant-height data. The proposed architecture brings together depth wise-separable convolutions, channel-attention mechanisms and Fourier feature projections to sink the spatial detail and capture the vegetation patterns. A stratified train–validation split and controlled augmentations were used to avoid data leakage and to boost the model’s robustness. Result: The validation obtained very consistent accuracy across the five fractions of biomass with a weighted determination coefficient of 0.81. The developed framework is a sustainable, non-invasive method for estimating pasture biomass and thus it gives huge advantages in precision grazing and sustainable livestock management.