Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Shahnawaz Alam, Mohammed Mudassir Uddin, Mohammed Kaif Pasha

TL;DR

This work tackles deploying plant-disease detectors on edge devices under limited labeled data by integrating pruning with few-shot learning. The DACIS scoring mechanism combines gradient sensitivity, activation variance, and Fisher-based class separability to identify disease-relevant channels, within a Prune-then-Meta-Learn-then-Prune (PMP) three-stage pipeline that alternates pruning and episodic meta-learning. Empirical results on PlantVillage and PlantDoc show substantial compression (up to 78% fewer parameters) with minimal accuracy loss (about 92% of the full model) and edge-appropriate throughput (e.g., 67 FPS on target hardware), along with deployment-focused metrics such as the Deployment Efficiency Score (DES). The framework advances practical edge deployment for smallholder farmers, enabling real-time disease diagnosis with limited data and resources, while remaining extensible to other domains through its modular DACIS components and PMP structure.

Abstract

Farmers in remote areas need quick and reliable methods for identifying plant diseases, yet they often lack access to laboratories or high-performance computing resources. Deep learning models can detect diseases from leaf images with high accuracy, but these models are typically too large and computationally expensive to run on low-cost edge devices such as Raspberry Pi. Furthermore, collecting thousands of labeled disease images for training is both expensive and time-consuming. This paper addresses both challenges by combining neural network pruning -- removing unnecessary parts of the model -- with few-shot learning, which enables the model to learn from limited examples. This paper proposes Disease-Aware Channel Importance Scoring (DACIS), a method that identifies which parts of the neural network are most important for distinguishing between different plant diseases, integrated into a three-stage Prune-then-Meta-Learn-then-Prune (PMP) pipeline. Experiments on PlantVillage and PlantDoc datasets demonstrate that the proposed approach reduces model size by 78\% while maintaining 92.3\% of the original accuracy, with the compressed model running at 7 frames per second on a Raspberry Pi 4, making real-time field diagnosis practical for smallholder farmers.

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

TL;DR

This work tackles deploying plant-disease detectors on edge devices under limited labeled data by integrating pruning with few-shot learning. The DACIS scoring mechanism combines gradient sensitivity, activation variance, and Fisher-based class separability to identify disease-relevant channels, within a Prune-then-Meta-Learn-then-Prune (PMP) three-stage pipeline that alternates pruning and episodic meta-learning. Empirical results on PlantVillage and PlantDoc show substantial compression (up to 78% fewer parameters) with minimal accuracy loss (about 92% of the full model) and edge-appropriate throughput (e.g., 67 FPS on target hardware), along with deployment-focused metrics such as the Deployment Efficiency Score (DES). The framework advances practical edge deployment for smallholder farmers, enabling real-time disease diagnosis with limited data and resources, while remaining extensible to other domains through its modular DACIS components and PMP structure.

Abstract

Farmers in remote areas need quick and reliable methods for identifying plant diseases, yet they often lack access to laboratories or high-performance computing resources. Deep learning models can detect diseases from leaf images with high accuracy, but these models are typically too large and computationally expensive to run on low-cost edge devices such as Raspberry Pi. Furthermore, collecting thousands of labeled disease images for training is both expensive and time-consuming. This paper addresses both challenges by combining neural network pruning -- removing unnecessary parts of the model -- with few-shot learning, which enables the model to learn from limited examples. This paper proposes Disease-Aware Channel Importance Scoring (DACIS), a method that identifies which parts of the neural network are most important for distinguishing between different plant diseases, integrated into a three-stage Prune-then-Meta-Learn-then-Prune (PMP) pipeline. Experiments on PlantVillage and PlantDoc datasets demonstrate that the proposed approach reduces model size by 78\% while maintaining 92.3\% of the original accuracy, with the compressed model running at 7 frames per second on a Raspberry Pi 4, making real-time field diagnosis practical for smallholder farmers.
Paper Structure (78 sections, 1 theorem, 28 equations, 9 figures, 22 tables, 1 algorithm)

This paper contains 78 sections, 1 theorem, 28 equations, 9 figures, 22 tables, 1 algorithm.

Table of Contents

  1. Introduction: Motivation through Agricultural Lens
  2. The Agricultural Deployment Challenge
  3. Research Gap and Contributions
  4. Related Work: Comparative Analysis with Gap Identification
  5. Neural Network Pruning Methodologies
  6. Few-Shot Learning Architectures
  7. Plant Disease Detection Systems
  8. Architectural Efficiency in Plant Pathology
  9. Transfer Learning and Domain Adaptation
  10. Position of This Work
  11. Methodology: Disease-Aware Pruning Framework
  12. Notation
  13. Problem Formulation: Shot-Adaptive Model Selection
  14. Hierarchical Disease Taxonomy
  15. Uncertainty Quantification in Low-Data Regimes
  16. ...and 63 more sections

Key Result

Proposition 1

Let $\mathcal{L}(\theta)$ be the cross-entropy loss and $\theta$ be the parameter vector. Under Gaussian class-conditional distributions, the perturbation in loss $\delta \mathcal{L}$ due to pruning channel $c$ is related to the Fisher Discriminant ratio $\mathcal{D}^{(c)}$.

Figures (9)

  • Figure 1: Representative samples from the PlantVillage simulated temporal generalization split showing disease symptom diversity across tomato (bacterial spot, early blight), potato (late blight), and pepper (bacterial spot) species under varying illumination and background complexity. These visual challenges motivate the disease-aware pruning approach. Note: This split simulates temporal separation by partitioning data to test generalization; images were not collected at different time points.
  • Figure 2: Overview of the PMP-DACIS framework. Stage 1: Initial pruning using DACIS scoring reduces parameters from 11.2M to 6.7M (40% compression). Stage 2: Episodic meta-learning with N-way K-shot tasks; inner loop adapts on support set $\mathcal{S}$, outer loop optimizes across query sets $\mathcal{Q}$. Stage 3: Meta-gradient guided refinement achieves 78% total compression (2.5M parameters) for edge deployment.
  • Figure 3: Shot-Adaptive Model Selection (SAMS) illustration. This figure shows the relationship between shot count and optimal model capacity. Note: Separate static models are trained for each regime; this is NOT dynamic runtime switching. 1-shot: High uncertainty requires 70% capacity (8/8 channels). 5-shot: Improved prototypes enable 45% pruning with 85% confidence. 10-shot: Abundant samples permit 78% compression (3/8 channels) with 94% confidence for edge deployment.
  • Figure 4: Hierarchical disease taxonomy guiding pruning protection. Coarse level$\mathcal{V}_1$: Pathogen types (288 channels) receive full protection. Medium level$\mathcal{V}_2$: Symptom types receive partial protection. Fine level$\mathcal{V}_3$: Specific diseases are primary pruning candidates.
  • Figure 5: DACIS pipeline: Feature maps evaluated through gradient norm $\mathcal{G}$, variance $\mathcal{V}$, and Fisher discriminant $\mathcal{D}$. Weighted aggregation produces channel importance scores; adaptive threshold $\tau_\ell$ determines retention.
  • ...and 4 more figures

Theorems & Definitions (7)

  • Definition 1: Shot-Adaptive Model Selection
  • Definition 2: DACIS
  • Proposition 1: DACIS-Loss Relationship
  • proof
  • Definition 3: Deployment Efficiency Score
  • Definition 4: Few-Shot Stability Index
  • Definition 5: Cross-Stage Generalization