⚡ electrical engineering

Cross-Dataset Strawberry Disease Classification from Controlled to African Field Conditions

This paper addresses the performance gap between controlled and real-world strawberry disease classification by proposing SmartBerry-DA, a label-efficient domain adaptation method that achieves high accuracy using minimal local annotations and enables lightweight offline deployment for practical agricultural use in Nigeria.

Original authors: Isibor Kennedy Ihianle, Pedro Machado, Jordan J. Bird, Salisu W. Yahaya, Ahmad Lotfi, Everistus Nwogo, Isaac Akinwumi, Jonathan Oluranti, Abiodun Adebayo, Abdulwahab Sulaiman Hamza, Ashley Dimka-Tapgu

Published 2026-09-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Isibor Kennedy Ihianle, Pedro Machado, Jordan J. Bird, Salisu W. Yahaya, Ahmad Lotfi, Everistus Nwogo, Isaac Akinwumi, Jonathan Oluranti, Abiodun Adebayo, Abdulwahab Sulaiman Hamza, Ashley Dimka-Tapgun

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

Strawberries are a delicate crop, and their health depends heavily on the condition of their leaves. When leaves develop dark spots or begin to scorch, the plant's ability to produce fruit weakens, often leading to significant losses for farmers. For years, scientists have used artificial intelligence to help identify these diseases from photographs. These computer programs are trained on vast libraries of images, but most of these libraries consist of pictures taken in perfect, controlled environments: leaves are centered, backgrounds are plain, and the lighting is consistent. While these programs work brilliantly in the lab, they often stumble when faced with the messy reality of a real farm, where shadows, dirt, uneven light, and different camera angles create a chaotic visual landscape. This gap between the clean training data and the messy real world is a major hurdle for bringing smart technology to farmers in places like Nigeria, where connectivity is limited and every dollar counts.

A team of researchers from the United Kingdom and Nigeria set out to solve this specific problem using strawberry crops. They wanted to know if a computer program trained on perfect, studio-like images could recognize diseases in a real Nigerian field, and if not, how much local effort would be needed to fix it. They began by testing four different types of AI models on a dataset of strawberry images taken under ideal conditions. As expected, these models performed with near-perfect accuracy, correctly identifying healthy leaves and diseased ones almost every time. However, when the researchers took these same models and pointed them at a new set of images taken from actual strawberry farms in Jos, Nigeria, the results collapsed. The models, which had been so confident in the lab, became unreliable in the field. Their ability to correctly classify the plants dropped significantly, with some models failing to recognize the disease more than half the time. The issue was not that the models were broken, but that they had learned to rely on the clean backgrounds and perfect lighting of the training photos, which did not exist in the Nigerian fields.

To bridge this gap without requiring farmers to spend years labeling thousands of new photos, the researchers developed a method they called SmartBerry Domain Adaptation. This approach acts like a translator that helps the computer understand the new environment using very little help. Instead of asking for a massive dataset of labeled field images, the team used a small number of labeled photos from the Nigerian farm—just five or ten examples of each type of leaf condition—combined with a large number of unlabeled photos. The system uses the few labeled images to learn the specific look of the local disease, while the unlabeled images help it understand the general context of the farm, such as the shadows and soil colors. By combining this limited local knowledge with the original training, the system was able to recover its accuracy. With only five labeled images per disease type, the model's performance jumped from a failing grade to a score that was nearly as good as if it had been trained on the entire local dataset. With ten labeled images, the performance became almost indistinguishable from a model trained on full local data.

The researchers also tested which type of computer model worked best for this task. They found that not all models responded well to this adaptation method. One specific model, a lightweight hybrid architecture, proved to be the most effective, learning the new field conditions quickly and accurately with minimal data. Another model, which was larger and more complex, also performed well but required more computing power. The lighter models were crucial because the goal was to run these programs on simple devices that farmers could carry in their pockets, without needing a constant internet connection. To ensure the software could run on such devices, the team converted the winning model into a smaller, more efficient format. They found that reducing the precision of the numbers the computer used to make decisions allowed them to cut the file size in half, from 4.37 megabytes down to 2.30 megabytes, without losing any accuracy. This smaller version could still identify diseases perfectly, even when the photos were blurry, dark, or partially covered by leaves.

The study concludes that while artificial intelligence trained in controlled labs is a powerful tool, it cannot be directly applied to real-world farming without adjustment. The reliability of these systems depends heavily on local data. However, the good news is that farmers do not need to label thousands of images to make these systems work. By using a smart adaptation method that leverages a tiny amount of labeled local data alongside a large pool of unlabeled photos, the technology can be tuned to local conditions with very little effort. This approach offers a practical path forward for deploying disease detection tools in regions where resources are scarce, allowing farmers to receive timely, accurate advice about their crops even when they are far from high-speed internet or expensive equipment. The work demonstrates that with the right strategy, the gap between the laboratory and the field can be closed, bringing the benefits of advanced technology to the people who need it most.

Technical Summary: Cross-Dataset Strawberry Disease Classification from Controlled to African Field Conditions

Problem Statement
Image-based plant disease classifiers frequently achieve high accuracy on benchmark datasets acquired under controlled conditions (e.g., PlantVillage) but suffer significant performance degradation when deployed in real-world field environments. This "controlled-to-field" gap is particularly acute in sub-Saharan Africa, where digital agriculture faces constraints in connectivity, infrastructure, and access to locally annotated data. The paper addresses the specific challenge of transferring strawberry disease recognition models trained on controlled imagery to field conditions in Nigeria, characterized by complex backgrounds, variable illumination, occlusion, and heterogeneous camera characteristics. The core problem is quantifying this generalization gap and developing a label-efficient adaptation strategy to recover field performance without requiring extensive local annotation.

Methodology
The study employs a four-stage experimental framework using two distinct datasets:

Datasets:
- Source Domain: PlantVillage (controlled conditions), specifically the Strawberry Healthy and Strawberry Leaf Scorch classes.
- Target Domain: SmartBerry (field conditions in Nigeria), specifically Healthy Leaves and Leaf Dark Spot classes. The correspondence is defined at the symptom level, not pathogen level.
Model Architectures: Four lightweight and efficient backbones were evaluated: MobileNetV3-Small, EfficientNet-B0, MobileViTv2-0.5, and ConvNeXt-Tiny.
Experimental Phases:
- Benchmarking: Evaluating models within-domain (PlantVillage and SmartBerry separately).
- Cross-Dataset Evaluation: Testing direct transfer from PlantVillage to SmartBerry against target-only and pooled training baselines.
- Label-Efficient Adaptation (SmartBerry-DA): A proposed semi-supervised domain adaptation framework designed to adapt models using a small budget of labeled target images ( $k \in \{5, 10, 20, 50\}$ $k \in {5, 10, 20, 50}$ per class) and a larger set of unlabelled target images.
  - Components: The framework combines supervised source/target learning, Correlation Alignment (CORAL) for feature distribution alignment, an Exponential Moving Average (EMA) teacher for pseudo-labeling, confidence-filtered pseudo-labels, and consistency regularization.
- Deployment & Robustness: Evaluating the selected model (MobileViTv2-0.5) for offline deployment using FP32, FP16, and INT8 quantization, and testing robustness against simulated field degradations (low illumination, shadow, blur, JPEG compression, occlusion, and reduced resolution).

Key Contributions

Controlled-to-Field Benchmark: Established a benchmark combining PlantVillage and original Nigerian field imagery (SmartBerry) for direct cross-dataset evaluation of strawberry disease classification.
Empirical Analysis of Domain Shift: Quantified the performance deterioration of state-of-the-art models when transferring from controlled to field conditions, highlighting asymmetries in class-specific recall and specificity, as well as confidence calibration failures.
SmartBerry-DA Framework: Introduced a label-efficient, semi-supervised domain adaptation method that leverages unlabelled field data to recover performance with minimal local annotation.
Lightweight Deployment Evaluation: Provided a rigorous assessment of model conversion (FP32, FP16, INT8) regarding size, latency, and robustness to image degradation, specifically tailored for offline agricultural inference.

Key Results

Generalization Gap: While all models achieved near-ceiling performance on PlantVillage (Macro F1 $\ge$ 0.99), direct transfer to SmartBerry resulted in significant drops. Macro F1 scores ranged from 0.418 (EfficientNet-B0) to 0.824 (ConvNeXt-Tiny). Lightweight models like MobileViTv2-0.5 and MobileNetV3-Small exhibited severe class imbalance, often over-predicting disease (high recall, low specificity).
Adaptation Efficacy: SmartBerry-DA successfully recovered field performance with minimal labels.
- Using MobileViTv2-0.5, the method achieved a Macro F1 of 0.940 ± 0.015 with only 5 labeled images per class and 0.988 ± 0.012 with 10 labeled images, approaching the full-target baseline of 0.994 ± 0.008.
- Ablation studies identified uncertainty filtering as the most critical component; removing it caused the largest performance drop (Macro F1 fell from 0.994 to 0.946).
- The method showed backbone sensitivity, performing best with MobileViTv2-0.5 and ConvNeXt-Tiny, while showing weaker or inconsistent results with EfficientNet-B0 and MobileNetV3-Small.
Deployment Findings:
- FP16 quantization reduced model size from 4.37 MB to 2.30 MB (47.2% reduction) while maintaining 100% prediction agreement and identical Macro F1 (1.000) compared to FP32.
- INT8 quantization reduced size further (1.57 MB) but severely degraded performance (Macro F1 dropped to 0.590) and increased latency.
- The FP16 model demonstrated robustness to low illumination, shadow, blur, and JPEG compression (Macro F1 remained 1.000), with minor drops only under partial occlusion (0.970) and reduced resolution (0.926).

Significance and Claims
The paper claims that strong performance on controlled benchmark datasets does not guarantee reliability in practical farming environments, particularly in resource-constrained regions like sub-Saharan Africa. The primary significance lies in demonstrating that field-derived data is essential for reliable deployment, but this data requirement can be drastically reduced through label-efficient adaptation.

The authors posit that SmartBerry-DA provides a viable pathway for transferring agricultural AI from benchmarks to practice by:

Minimizing the need for extensive local annotation, which is often a bottleneck in developing regions.
Enabling offline, lightweight inference suitable for devices with limited connectivity and storage.
Highlighting that uncertainty-aware adaptation (specifically confidence filtering) is crucial for leveraging unlabelled data effectively.

The paper concludes that while the proposed approach offers a technically evaluated pathway for strawberry disease recognition in Nigerian field conditions, future work must extend evaluation across diverse farms, seasons, and hardware, and integrate multimodal sensing data. The authors remain modest, noting that the results are specific to the symptom-compatible classes evaluated and that the deployment latency was measured on reference hardware rather than final end-user devices.

Technical Summary: Cross-Dataset Strawberry Disease Classification from Controlled to African Field Conditions

More like this