Transfer learning and stacking ensembles for biomass estimation from smartphone imagery in pasture-based dairy systems

Accurate estimation of pasture biomass is essential for optimising feed management in pasture-based dairy systems. Traditional methods remain labour-intensive and impractical for frequent monitoring. Remote sensing via satellites or unmanned aerial vehicles offers alternatives but is constrained by cloud interference and infrastructure costs. This study investigates smartphone imagery as a scalable solution, evaluating 21 machine learning configurations for predicting pasture biomass across 15 commercial dairy farms in New South Wales, Australia. Ground-level red-green-blue images (n = 3291) were paired with rising plate meter measurements and weather data to develop estimation models using transfer learning, ensemble methods, and multi-architecture feature integration. Under random within-dataset partitioning, stacking ensemble combining pre-trained convolutional neural network features with hand-crafted vegetation indices and weather variables achieved R 2 = 0.561, MAE = 351 kg DM ha −1 , and RMSE = 437 kg DM ha −1 , representing a 48% improvement over baseline hand-crafted features. Transfer learning using frozen ImageNet features outperformed deep learning trained from scratch. Multi-architecture integration captured complementary information, with DenseNet121 features dominating importance rankings. Weather data integration provided 5.9% performance improvement. Seasonal performance varied substantially, with spring achieving highest accuracy (R 2 = 0.701, MAE = 308 kg DM ha −1 ) during peak ryegrass growth and winter showing reduced performance (R 2 = 0.456, MAE = 375 kg DM ha −1 ). However, leave-one-farm-out cross-validation revealed severe spatial generalisation constraints (mean R 2 = −0.009, MAE = 399 kg DM ha −1 , RMSE = 488 kg DM ha −1 ), exposing strong farm-specific dependencies that limit predictive accuracy for unseen farm environments. These findings indicate that the proposed approach performs best when target farms contribute to the training dataset; generalisation to completely new farm environments will require either substantially expanded training datasets or farm-specific adaptation strategies.