Using Continuous Output Neural Nets to Estimate Pasture Biomass from Digital Photographs in Grazing Lands

Abstract

Accurate estimates of pasture biomass in grazing lands are currently a time-consuming and resource-intensive task. The process generally includes physically cutting, bagging, labelling, drying, and weighing grass samples using multiple “quadrats” placed on the ground. Quadrats vary in size but are typically in the order of 0.25 m2 (i.e., 0.5 m × 0.5 m) up to 1.0 m2. Measurements from a number of harvested quadrats are then averaged to get a site estimate. This study investigated the use of photographs and ‘machine learning’ to reduce the time factor and difficulty in taking pasture biomass measurements to potentially make the estimations more accessible through the use of mobile phone cameras. A dataset was created from a pre-existing archive of quadrat photos and corresponding hand-cut pasture biomass measurements taken from a diverse range of field monitoring sites. Sites were clustered and one was held back per model for testing. The models were based on DenseNet121. Individual quadrat errors were large but more promising results were achieved when estimating the site mean pasture biomass. Another two smaller additional datasets were created post-training which were used to further assess the ensemble; they provided similar absolute errors to the original dataset, but significantly larger relative errors. The first was made from harvested quadrats, and the second was made using a pasture height meter in conjunction with a mobile phone camera. The models performed well across a variety of situations and locations but underperformed when assessed on some sites with very different vegetation. More data and refinement of the approach outlined in the paper will reduce the number of models needed and help to correct errors. These models provide a promising start, but further investigation, refinement, and data are needed before becoming a usable application.