💻 computer science

A Novel Hardware-Aware Multi-Objective Evolutionary Neural Architecture for Crop-Disease Classification on Edge Devices

This paper proposes a hardware-aware, multi-objective evolutionary search framework using NSGA-II to optimize EfficientNet-Lite architectures for crop disease classification on edge devices, achieving high accuracy (0.98) with significantly reduced latency and model size compared to existing models like EfficientNet-B0 and MobileNetV3.

Original authors: Anitha K, Naveenkumar R, Pooja H L, Shantha Kumar M, Valarmathi K, Monica K M

Published 2026-11-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anitha K, Naveenkumar R, Pooja H L, Shantha Kumar M, Valarmathi K, Monica K M

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, sun-drenched fields where the world's food is grown, a silent battle is constantly being fought. It is not a war of weapons, but of biology, where microscopic fungi, bacteria, and viruses threaten to wipe out entire harvests. For centuries, farmers have relied on their eyes and experience to spot these threats, a method that is often too slow, too subjective, or simply impossible when a farmer cannot be everywhere at once. Today, technology offers a new pair of eyes: artificial intelligence. By teaching computers to recognize the specific patterns of a sick leaf, we can detect disease early, saving crops and reducing the need for harmful chemicals. However, the most powerful computer brains currently available are massive, requiring huge amounts of electricity and expensive hardware to run. They live in giant data centers, far away from the dusty fields where they are needed most. The challenge for scientists is not just to build a smart computer, but to shrink it down enough to fit inside a small, battery-powered device that a farmer can carry in their pocket, all while keeping it sharp enough to make life-or-death decisions for a crop.

A team of researchers has tackled this exact problem by creating a new way to design these tiny, intelligent systems. Instead of manually building a computer model and hoping it works well on a small device, they developed a method that lets the computer design itself. They started with a known, efficient blueprint for a neural network—a type of artificial intelligence inspired by the human brain—and then used a digital evolutionary process to refine it. Imagine a population of digital architects, each proposing a slightly different version of a disease-detecting brain. The researchers tested these versions not just on how well they could identify a sick leaf, but also on how fast they could do it and how much memory they used. The best performers were allowed to "breed," combining their successful features to create even better generations, while the less efficient ones were discarded. This process, known as a genetic algorithm, mimics natural selection, allowing the system to evolve a perfect balance between accuracy and speed specifically for the small, low-power chips found in modern farming equipment.

The team tested this approach using a massive library of over 82,000 images of plant leaves, covering fourteen different types of crops and thirty-eight distinct categories of health and disease. They fed these images into their evolutionary system, which ran thousands of trials to find the optimal structure. The result was a custom-built neural architecture that is remarkably compact. When they deployed this new model onto a Raspberry Pi, a small, affordable computer often used by hobbyists and researchers, and an NVIDIA Jetson Nano, a slightly more powerful board designed for edge computing, the results were striking. The system achieved an accuracy of 98.2 percent in identifying crop diseases, a level of precision that rivals much larger, more expensive models. More importantly, it did so with incredible speed, making a decision in just 12.9 milliseconds on the Jetson Nano. This is fast enough to allow a farmer to scan a leaf and get an immediate diagnosis without waiting for a cloud connection or a powerful server.

The researchers compared their evolved model against several existing, well-known designs that are often used for mobile devices. The new architecture was significantly faster and smaller than its competitors, using far fewer computational resources while maintaining high accuracy. In a direct comparison, the new model required less than half the computational power of some standard designs and operated with a latency that was nearly half that of others. This efficiency means that a farmer could potentially run this system on a device that costs very little and runs on a small battery, making advanced disease monitoring accessible even in remote areas with limited infrastructure. The study also showed that the model remained robust across different types of crops, from tomatoes and apples to blueberries and grapes, proving that the evolutionary search process could find a single, versatile solution rather than a model that only works for one specific plant.

What makes this work particularly significant is that it moves beyond simply training a model on a powerful computer and then trying to force it to run on a small one. Instead, the researchers built the constraints of the small device directly into the design process from the very beginning. By asking the computer to evolve a solution that is inherently lightweight, they avoided the common problem of having to sacrifice too much accuracy to gain speed. The study suggests that this hardware-aware approach is a viable path forward for precision agriculture, where the goal is to bring high-level intelligence directly to the field. While the current work was tested on a specific set of images taken under controlled conditions, the authors note that future steps will involve testing these models in the messy, unpredictable reality of outdoor farming, with varying light, weather, and soil conditions. For now, however, the research demonstrates that it is possible to shrink the brain of a modern artificial intelligence down to a size that fits in a farmer's hand, without losing its ability to see the invisible threats that endanger our food supply.

Technical Summary: A Novel Hardware-Aware Multi-Objective Evolutionary Neural Architecture for Crop-Disease Classification on Edge Devices

1. Problem Statement

The paper addresses the critical challenge of deploying deep learning models for crop disease detection on resource-constrained edge devices (e.g., Raspberry Pi, NVIDIA Jetson Nano). While Convolutional Neural Networks (CNNs) and Vision Transformers offer high accuracy, conventional architectures often suffer from excessive parameter counts, high computational overhead, and significant inference latency. These factors hinder real-time deployment in agricultural settings where internet connectivity is limited and hardware resources are scarce. Existing lightweight models (e.g., MobileNet, EfficientNet) often require manual tuning and may not optimally balance the trade-offs between classification accuracy, inference speed, and memory usage for specific target hardware.

2. Methodology

The proposed framework introduces a hardware-aware, multi-objective evolutionary search to automatically generate optimized neural architectures based on the EfficientNetV2 backbone. The core components of the methodology include:

Dataset and Preprocessing: The study utilizes the PlantVillage dataset, comprising 82,286 images across 14 plant types and 38 classes (26 disease categories and 12 healthy classes). Preprocessing involves Contrast Limited Adaptive Histogram Equalization (CLAHE) for noise removal and contrast enhancement. To mitigate overfitting and balance class distribution, a Generative Adversarial Network (GAN) is employed for data augmentation, increasing the training set size from 57,600 to 65,125 images.
Search Space Definition: The architecture search space is built upon EfficientNetV2, which utilizes Fused-MBConv and MBConv blocks. The search space allows for the optimization of:
- Global Parameters: Input resolution (128, 160, 192) and width multiplier (0.5, 0.75, 1.0, 1.25).
- Stage-Specific Parameters: For each of the five feature-extraction stages, the algorithm searches for the block type (Fused-MBConv vs. MBConv+SE), kernel size (3x3 or 5x5), expansion ratio (2, 4, or 6), and block depth (1 to 3 repeats).
- This configuration results in a search space of approximately 700 million possible architectures.
Optimization Algorithm (NSGA-II): The Non-dominated Sorting Genetic Algorithm II (NSGA-II) is employed to navigate the search space. It treats the problem as a multi-objective optimization task with three minimization objectives:
1. Validation Error ( $1 - \text{Accuracy}$ )
2. Inference Latency (measured in milliseconds on target hardware)
3. Parameter Count (in millions)
  The algorithm uses a surrogate predictor to estimate performance and a hardware-specific lookup table to estimate latency, avoiding the computational cost of training every candidate model to convergence. It employs non-dominated sorting and crowding distance to maintain a diverse Pareto front of solutions.
Deployment: The final Pareto-optimal architecture is selected based on device constraints, retrained on the full dataset, quantized to INT8, and deployed on edge devices.

3. Key Contributions

The authors identify three primary contributions:

Hardware-Aware Architecture: The proposal of a lightweight EfficientNetV2-based architecture where parameters are automatically optimized to enhance performance while reducing computational load.
Multi-Objective Optimization: The integration of NSGA-II to simultaneously optimize disease classification performance and hardware deployment constraints (latency, memory, energy) without manual intervention.
Comprehensive Evaluation: Extensive experimentation using the PlantVillage dataset, demonstrating the framework's efficacy across different edge architectures (Raspberry Pi 4 and NVIDIA Jetson Nano) with detailed metrics including accuracy, precision, recall, F1-score, and latency.

4. Experimental Results

The framework was evaluated on a workstation (AMD Ryzen 9, NVIDIA RTX 3090) for training and on edge devices for inference.

Performance Metrics: The proposed NSGA-II optimized EfficientNetV2 achieved an accuracy of 0.982, precision of 0.97, recall of 0.96, and an F1-score of 0.97.
Efficiency Gains:
- Latency: The model achieved an inference latency of 12.90 ms on the NVIDIA Jetson Nano, compared to 15.55 ms on Raspberry Pi and 19.90 ms on a CPU workstation.
- Model Size: The optimized model reduced parameters to 175,602 (as reported in Table 8). Note: While Table 8 labels the column "Params (M)", the value listed is 175,602, which corresponds to a raw parameter count rather than millions. The summary reflects the raw number 175,602 as presented in the table data, acknowledging the likely unit discrepancy in the source text.
Comparative Analysis:
- Compared to EfficientNet-B0, the proposed model is faster and smaller while maintaining comparable or superior accuracy.
- Compared to MobileNetV2 and ResNet-50, the proposed framework demonstrated superior accuracy (0.982 vs. 0.90 and 0.87, respectively) and lower latency (12.90 ms vs. 28.88 ms and 32.99 ms).
Ablation Study: The study confirmed that the full framework (incorporating NAS, multi-objective NSGA-II, and hardware awareness) outperformed the baseline fixed EfficientNetV2 (0.90 accuracy) and configurations using only NAS or only NSGA-II. The full framework reduced latency from 22.89 ms (baseline) to 12.90 ms while increasing accuracy to 0.982.

5. Significance and Claims

The paper claims that hardware-aware evolutionary search is a viable method for producing compact, accurate models suitable for real-time plant disease monitoring in the field. The significance lies in the ability to automatically generate architectures that balance the "accuracy-efficiency" trade-off specific to the target edge hardware, eliminating the need for manual architecture design.

The authors note that while the results are promising, the framework currently relies on the PlantVillage dataset (controlled conditions). Future work, as outlined by the authors, focuses on:

Improving robustness through large-scale field datasets with variations in illumination, occlusion, and weather.
Implementing online or continual learning to adapt to emerging diseases without full retraining.
Exploring federated learning for privacy-preserving collaborative model improvement across distributed agricultural sites.

The research concludes that the proposed framework successfully demonstrates that automated, hardware-aware search can yield models that are both highly accurate and computationally efficient for deployment on low-power agricultural edge devices.