Energy-Efficient Plant Monitoring via Knowledge Distillation

Large-scale vision models, especially vision transformers and multimodal foundation models, have brought breakthroughs in plant species and disease classification. But their high computational cost makes them impractical for mobile or edge devices used in precision agriculture and biodiversity monitoring. To bridge this gap, a team led by researchers from INRIA, University of Montpellier, and other institutions systematically evaluated knowledge distillation as a compression technique. They tested four representative architectures — two ConvNeXt models and two vision transformers — under multiple training regimes (from-scratch, pretrained, with and without distillation) on two challenging benchmarks: Pl@ntNet300K-v2 (for species) and Deep-Plant-Disease (for disease recognition). In total, they trained and evaluated 70 models.

The results are clear: knowledge distillation consistently improved performance across all tasks and architectures. Distilled smaller models matched the accuracy of much larger, computationally expensive models while maintaining substantially lower energy consumption. This means that accurate plant recognition can run on resource-constrained hardware without sacrificing performance. The study demonstrates a practical path to scalable, real-world deployment of automated plant monitoring systems — critical for large-scale environmental observation and smart farming applications where energy efficiency and low latency are paramount.