📄 agriculture

SolanAPP: An Offline-First Mobile Framework for Segmentation-Based Diagnosis of Solanaceae Crop Diseases

SolanAPP is an offline-first, native Android framework that leverages lightweight deep learning models for real-time, pixel-level diagnosis of Solanaceae crop diseases and integrates optional LLM-driven agronomic recommendations to enable accurate, accessible, and collaborative disease surveillance in resource-constrained agricultural environments.

Original authors: Dunia Pineda Medina, Mino Sportelli, Ileana Miranda Cabrera, Stefano Giordano, Antonino Crivello

Published 2026-08-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Dunia Pineda Medina, Mino Sportelli, Ileana Miranda Cabrera, Stefano Giordano, Antonino Crivello

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 quiet corners of the world where food is grown, a silent battle is constantly being waged against invisible enemies. Crops like tomatoes, potatoes, peppers, and eggplants are the backbone of global nutrition, yet they are perpetually vulnerable to diseases that can wipe out entire harvests. For generations, farmers have relied on their own eyes and the advice of distant experts to identify these threats, a process that is often slow, inaccurate, and dependent on having access to a doctor or a laboratory. In recent years, technology has promised a faster solution: smartphones equipped with artificial intelligence that can look at a sick leaf and instantly name the problem. However, this promise has hit a hard wall in the very places it is needed most. The most advanced diagnostic tools usually require a strong internet connection to send images to a cloud server for analysis. In remote rural areas, where the soil is rich but the signal is weak or non-existent, these high-tech tools are useless. The challenge for scientists has been to build a system that is smart enough to diagnose complex plant diseases but small and efficient enough to run entirely on a phone without ever needing to connect to the internet.

A team of researchers has answered this challenge with a new mobile application called SolanAPP. Designed specifically for the family of crops that includes tomatoes, potatoes, peppers, and eggplants, this tool represents a significant shift in how agricultural technology is deployed in the field. Instead of sending a photo to a distant server, SolanAPP performs the entire diagnosis inside the phone itself. It uses a specialized form of artificial intelligence that can not only identify the disease but also draw a precise outline around the sick parts of the leaf, showing the farmer exactly how much of the plant is affected. This "offline-first" approach means that a farmer in a remote village with no internet signal can still get a professional-level diagnosis, a capability that transforms a standard smartphone into a self-contained agricultural clinic.

The researchers built this system by training computer models on thousands of images of healthy and diseased plants. They taught the software to recognize specific patterns, such as the concentric rings of early blight on a potato leaf or the yellow curling of a tomato leaf infected with a virus. What makes this work unique is that the team did not just focus on getting the right answer; they focused on making the answer fast and lightweight. They tested many different types of artificial intelligence architectures, looking for the one that could balance high accuracy with the limited memory and battery life of a standard mobile phone. After rigorous testing, they found that a specific combination of models worked best for most crops, achieving a classification accuracy of over 97 percent. For potatoes, a slightly different model was chosen to ensure the fastest possible speed. The result is a system that can process an image and provide a diagnosis in less than a second, even on modest hardware.

Beyond simply naming the disease, the application provides a visual map of the infection. When a user takes a picture, the app highlights the diseased areas in distinct colors, allowing the farmer to see the severity of the problem at a glance. The system calculates the percentage of the leaf that is damaged, turning a vague feeling of worry into a concrete number. If the damage is minor, the app suggests mild interventions; if the infection is severe, it flags the situation as critical. This ability to estimate severity is crucial because it helps farmers decide whether to treat a single plant or an entire field, preventing the waste of resources and the spread of disease.

The researchers also recognized that a diagnosis is only as good as the advice that follows it. To bridge the gap between a computer's data and a farmer's decision, the app includes an optional feature that connects to a large language model when an internet connection is available. This feature acts as a digital agronomist, taking the visual data and the farmer's own notes to generate a detailed, plain-language report. It explains what the disease is, why it is happening, and offers specific treatment recommendations, including organic solutions for those who cannot afford expensive chemicals. This layer of reasoning ensures that the technology does not just identify a problem but helps solve it.

To ensure the tool actually works in the real world, the team tested it in Cuba, a region chosen specifically because of its challenging connectivity and the restrictions on accessing commercial cloud services. They gathered a group of twenty-five participants, including small-scale farmers, students, and agricultural technicians, to use the app in the field. The feedback was overwhelmingly positive. The users rated the system's reliability and its ability to work without internet as the most valuable features, giving it a high satisfaction score. They appreciated that the app was intuitive and that the visual outlines helped them understand the diagnosis immediately. While some users noted that the technical language in the reports could be simplified further and that the app sometimes struggled with very bright sunlight, the overall consensus was that the tool successfully removed the barrier of connectivity.

The study concludes that SolanAPP is more than just a diagnostic tool; it is a step toward democratizing access to advanced agricultural science. By proving that complex artificial intelligence can run entirely on a device without needing a network, the researchers have shown that high-quality crop protection is possible even in the most isolated environments. The system successfully balances the need for technical accuracy with the practical realities of rural life, offering a path forward where technology serves the farmer directly, rather than waiting for a connection that may never come. As the team looks to the future, they plan to expand the range of crops the app can handle and refine the advice it gives, but the core achievement remains clear: a phone in a farmer's hand can now hold the power to save a harvest, regardless of where they are standing.

Technical Summary: SolanAPP

Problem Statement
The diagnosis of diseases in Solanaceae crops (tomato, potato, bell pepper, and eggplant) traditionally relies on manual inspection, which is inefficient and suffers from significant latency. While recent advancements in Machine Learning (ML) and Deep Learning (DL) have achieved high classification accuracies (often >95%), existing solutions face critical barriers to practical deployment in rural agricultural settings. These barriers include:

Connectivity Constraints: Many state-of-the-art systems rely on cloud-based APIs or constant internet access, rendering them unusable in remote areas with unstable or non-existent connectivity.
Lack of Interpretability: Standard image classification models identify disease categories but fail to localize symptoms or estimate severity, providing limited visual evidence to the user.
Hardware Limitations: High-performance models (e.g., Vision Transformers, heavy segmentation networks) often exceed the memory and computational constraints of mobile edge devices.
Usability Gaps: There is a lack of unified systems that combine computer vision with Large Language Models (LLMs) to provide expert-level, understandable agronomic explanations for users with varying levels of digital literacy.

Methodology
The authors propose SolanAPP, an offline-first mobile framework designed for Android devices. The system architecture is divided into two layers:

Edge Layer (Offline Core):
- Data Curation: A dataset of 11,635 images was curated from PlantVillage, Kaggle, and Roboflow, covering four crops and multiple disease classes. Ground-truth masks were manually annotated using Roboflow. The dataset was split (80/15/5) prior to augmentation to prevent data leakage.
- Multi-Task Architecture: The system employs a multi-output network that processes a 256×256 RGB image to produce two simultaneous outputs: a disease classification vector and a semantic segmentation mask.
- Model Selection: Various architectures (U-Net, DeepLabV3+, Mask R-CNN) combined with backbones (MobileNetV2, ResNet-50) were evaluated. The optimal models were selected using Simple Additive Weighting (SAW), balancing predictive performance (mIoU, Accuracy) against operational constraints (inference latency, model size, parameter count).
  - Tomato, Bell Pepper, Eggplant: Selected DeepLabV3+ with MobileNetV2.
  - Potato: Selected U-Net with MobileNetV2 for superior computational efficiency.
- Training Strategy: The models were trained using a weighted total loss function: $L_{total} = 1.5 L_{seg} + 0.5 L_{cls}$ . The segmentation branch utilized Focal Tversky Loss (FTL) to address class imbalance (small lesion areas), while the classification branch used Categorical Cross-Entropy.
- Deployment: Models were converted to TensorFlow Lite (TFLite) for on-device inference. The app performs preprocessing (resizing, normalization) and postprocessing (severity calculation) locally.
- Severity Estimation: Disease severity is calculated as the percentage of symptomatic pixels relative to the total leaf area (isolated via HSV color space filtering), rather than the full image grid.
Cloud Layer (Optional Online Services):
- When connectivity is available, the system integrates Firebase for authentication and data persistence.
- It utilizes Groq's LPU to run a Large Language Model (LLM). The LLM receives a structured diagnostic vector (crop type, predicted disease, confidence, severity index, urgency level, and user notes) to generate a professional, structured agronomic report.
- This layer facilitates a georeferenced community surveillance network where users can share observations and receive peer validation.

Key Contributions

Offline-First Edge AI: A fully functional diagnostic pipeline that operates without internet connectivity, addressing the infrastructure limitations of rural environments.
Semantic Segmentation for Severity: Moving beyond simple classification, the system localizes symptomatic pixels to provide visual explanations and quantitative severity estimates.
Hybrid AI Architecture: The integration of lightweight on-device computer vision with optional cloud-based LLM reasoning to bridge the gap between raw data and actionable, expert-level agronomic advice.
Collaborative Surveillance: A georeferenced community module that allows for the aggregation of disease data, supporting regional epidemiological monitoring.
Rigorous Model Selection: A systematic approach using SAW to select models that optimize the trade-off between accuracy and mobile deployment constraints.

Results

Predictive Performance: The selected models achieved a macro-averaged classification accuracy of 98.48%, a recall of 98.48%, and an F1-score of 98.47%.
Segmentation Performance: The system achieved a mean Intersection over Union (mIoU) of 82.13% and a mean Pixel Accuracy (mPA) of 93.67%.
Efficiency: The DeepLabV3+ & MobileNetV2 configuration provided the best balance for most crops, while the U-Net & MobileNetV2 configuration was optimal for potatoes. Inference latency averaged 70.88 ms on a workstation, though the paper notes this exceeds the 30 ms target for real-time video processing, making it suitable for interactive still-image diagnosis.
User Validation: A survey of 25 participants (farmers, students, and technicians) in Cuba yielded an overall satisfaction rating of 4.5/5. Users particularly valued the offline capability (4.59/5) and the educational value of the LLM module (4.49/5).

Significance and Claims
The paper posits that SolanAPP represents a significant step toward democratizing access to advanced diagnostic technologies for marginalized smallholder farmers. Its primary significance lies in:

Socio-Technical Impact: It acts as a catalyst for digital literacy and open access to AI tools in regions with unfavorable development conditions and limited connectivity.
Practical Utility: By prioritizing offline operation and interpretability (via segmentation), the framework addresses the specific needs of field environments where cloud-dependent solutions fail.
Community Building: It establishes a theoretical and practical foundation for a collaborative, georeferenced epidemiological surveillance network in the agricultural sector.

The authors maintain a modest stance, acknowledging that while the framework is promising, it requires further validation on independent, real-world field datasets to ensure robustness against environmental variations (e.g., lighting, background noise) and to refine the LLM's recommendations for local contexts. They also note current limitations regarding the reliance on internet for the expert advisory layer and the need for future work to incorporate small language models (SLMs) for fully autonomous offline reasoning.

Technical Summary: SolanAPP

More like this