Transfer Learning Performance for Remote Pastureland Trait Estimation in Real-Time Farm Monitoring

In precision agriculture, having knowledge of pastureland forage biomass and moisture content prior to an ensiling process enables pastoralists to enhance silage production. While traditional trait measurement estimation methods relied on hand-crafted vegetation indices, manual measurements, or even destructive methods, remote sensing technology coupled with state-of-the-art deep learning algorithms can enable estimation using a broader spectrum of data, but generally require large volumes of labelled data, which is lacking in this domain. This work investigates the performance of a range of deep learning algorithms on a small dataset for biomass and moisture estimation that was collected with a compact remote sensing system designed to work in real time. Our results showed that applying transfer learning to Inception ResNet improved minimum mean average percentage error from 45.58% on a basic CNN, to 28.07% on biomass, and from 29.33% to 8.03% on moisture content. From scratch models and models optimised for mobile remote sensing applications (MobileNet) failed to produce the same level of improvement.