Active productPublished on Google Play100+ downloadsContinuously evolving
CultivoScan
Artificial intelligence for detecting foliar diseases in potato crops
A mobile application developed within Carchi's agricultural context and available to users throughout Ecuador. It combines computer vision, on-device inference and digital tools to support farmers.
Under review Associated scientific research Automated Detection of Foliar Diseases in Potato Using Computer Vision and Local Inference on Mobile Devices
Associated scientific research
From mobile product to scientific research
CultivoScan led to a scientific manuscript documenting the construction of a hybrid dataset, the experimental comparison of YOLO and RT-DETR architectures, and deployment of the selected model through on-device inference on Android.
Complete manuscript · Under technical review
Preparing for submission to a scientific journal
1
Research completedComplete
2
Manuscript developedComplete
3
Academic reviewIn progress
4
Editorial submissionNext stage
5
Journal evaluationPending
◇
Hybrid dataset
2,152 PlantVillage images and 908 field photographs.
◇
Experimental comparison
YOLO and RT-DETR variants evaluated for detection performance and efficiency.
◇
Mobile inference
Model exported to TensorFlow Lite and validated on an Android device.
Technology developed in Carchi
Developed in Carchi for all of Ecuador
CultivoScan was conceived within the agricultural reality of Carchi, Ecuador's leading potato-producing province, and is available to farmers and users throughout the country.
01
Context
Carchi is one of Ecuador's main potato-producing territories and the agricultural context in which CultivoScan was developed.
02
Problem
Visual identification of foliar diseases can depend on producer experience and access to technical assistance, making timely diagnosis difficult.
03
Solution
CultivoScan analyzes leaf images, runs the model locally and presents the diagnosis, detected region and general recommendations on a mobile device.
From field to device
1
Collection
PlantVillage and photographs captured in Carchi plots.
2
Preparation
Cleaning, bounding-box annotation and training-set augmentation.
3
Experimentation
Comparison of YOLO and RT-DETR variants.
4
Selection
YOLO-26-N, experiment E12, BoxGain and 1024-pixel resolution.
5
Deployment
Export to TensorFlow Lite.
6
Product
Flutter integration and Google Play publication.
7
Evolution
New versions, improvements and future feature expansion.
Product and technology
What it enables
✓Camera or gallery diagnosis
✓On-device inference
✓Diagnosis history
✓Plot management
✓Weekly weather
✓Potato prices
✓Recommendations
✓Authentication and synchronization
How it was built
Mobile application
Flutter
Dart
Riverpod
GoRouter
Artificial intelligence
Python
YOLO-26-N
RT-DETR
PyTorch
TensorFlow Lite
OpenCV
Roboflow
EigenCAM
Services
Firebase Authentication
Firestore
Cloud Functions
Express
Open-Meteo
Verifiable results
Experimental evaluation
98.89%mAP@50
98.05%mAP@50–95
96.99%Precision
95.76%Recall
Mobile and user validation
559.95 ± 30.91 msMean inference on the evaluated device
4.77/5Perceived effectiveness
4.78/5Perceived efficiency
ISO/IEC 25010Dimensions used in the evaluation
5.06 ms corresponds to the experimental GPU environment; 559.95 ms to mean TensorFlow Lite inference on the evaluated Android device; and 10.43 s to the mean complete flow from confirming the image to fully displaying the result.
A product that keeps growing
CultivoScan did not end with the graduation project. The application continues to evolve through new versions, diagnosis improvements, dataset expansion, outreach among producers and new agricultural support tools.
