[Paper] Meta-Learning Guided Pruning for Few-Shot Plant Pathology on Edge Devices

Source: arXiv - 2601.02353v1

Overview

Deep learning can spot plant diseases from leaf photos with impressive accuracy, but the models are usually too bulky to run on cheap edge devices that farmers actually have on the field. This paper tackles both the size problem and the scarcity of labeled data by marrying neural‑network pruning with few‑shot meta‑learning, delivering a lightweight model that still performs well with only a handful of training examples.

Key Contributions

• Disease‑Aware Channel Importance Scoring (DACIS): a novel metric that ranks convolutional channels by how much they help differentiate plant diseases, guiding aggressive yet safe pruning.
• Three‑stage PMP pipeline: Prune → Meta‑Learn → Prune (PMP) that first removes obvious redundancies, then adapts the slimmed model to new diseases with few examples, and finally fine‑tunes the pruning for maximal compression.
• Real‑world validation: Experiments on the PlantVillage and PlantDoc benchmarks show a 78 % reduction in model parameters while retaining 92.3 % of the original accuracy.
• Edge‑device deployment: The compressed model runs at ≈7 fps on a Raspberry Pi 4, demonstrating feasible real‑time inference for low‑cost hardware.

Methodology

1. Initial Pruning (Stage 1)

   • Start with a standard CNN (e.g., ResNet‑18).
   • Compute DACIS for each channel: the score aggregates gradients of the loss w.r.t. channel activations across disease classes, highlighting channels that are “disease‑aware.”
   • Remove channels with the lowest DACIS scores, yielding a slimmer backbone.

2. Meta‑Learning on Few Shots (Stage 2)

   • Use Model‑Agnostic Meta‑Learning (MAML) on the pruned network.
   • During meta‑training, the model sees many episodes each containing a few labeled leaf images from a subset of diseases.
   • The objective is to learn a set of parameters that can be quickly adapted (in 1–5 gradient steps) to a new disease with only a handful of examples.

3. Fine‑Grained Pruning (Stage 3)

   • After meta‑training, re‑evaluate DACIS on the adapted model.
   • Perform a second, more aggressive pruning pass, now informed by the meta‑learned representations, and optionally fine‑tune the remaining weights.

The pipeline is lightweight enough to be executed once offline (e.g., in a research lab), after which the final compressed model can be shipped to field devices.

Results & Findings

• Inference speed: On a Raspberry Pi 4 (4 GB RAM), the compressed model processes ~7 fps, compared to ~2 fps for the unpruned network.
• Few‑shot adaptation: With only 5 labeled images per new disease, the meta‑learned model reaches within 2 % of the fully supervised performance.
• Ablation: Removing DACIS and using naïve L1‑norm pruning drops accuracy by an additional ~6 %, confirming the importance of disease‑aware scoring.

Practical Implications

• Field‑ready diagnostics: Smallholder farmers can attach a cheap camera module to a Raspberry Pi and get near‑real‑time disease alerts without internet connectivity.
• Rapid rollout for emerging diseases: Because the model can be fine‑tuned with just a few new samples, agricultural extension services can quickly disseminate updates for novel pathogens.
• Cost savings: Eliminating the need for cloud inference reduces data‑plan expenses and latency, making the solution viable in bandwidth‑constrained rural areas.
• Open‑source potential: The pipeline can be applied to any image‑based classification task on edge hardware (e.g., pest detection, soil‑type recognition), encouraging community‑driven extensions.

Limitations & Future Work

• Dataset bias: The experiments rely on curated datasets (PlantVillage, PlantDoc) that may not capture the full variability of field lighting, occlusions, or leaf orientations. Real‑world field trials are needed to confirm robustness.
• Meta‑learning overhead: While inference is cheap, the meta‑training phase still requires GPU resources, which could be a barrier for organizations lacking compute budgets.
• Hardware specificity: Performance numbers are tied to Raspberry Pi 4; further optimization (e.g., quantization, TensorRT) could be explored for even lower‑cost microcontrollers.
• Extension to multi‑modal data: Future work could integrate additional sensors (e.g., hyperspectral imaging) to improve disease discrimination, especially for early‑stage infections that are visually subtle.

Authors

• Shahnawaz Alam
• Mohammed Mudassir Uddin
• Mohammed Kaif Pasha

Paper Information

• arXiv ID: 2601.02353v1
• Categories: cs.CV, cs.LG
• Published: January 5, 2026
• PDF: Download PDF