Introduction
In developing regions like rural Colombia, a significant "digital breach" persists, isolating small-scale farmers from modern agricultural advancements. Mountainous terrain and the prohibitive cost of specialized equipment often confine these producers to traditional, less efficient farming methods. While smartphone penetration is increasing, a lack of specialized software prevents these devices from being used as professional agricultural tools. This gap results in two primary challenges: delayed diagnosis of crop diseases and market inefficiencies where farmers are dependent on local middlemen.
AgroTIC is an integrated mobile platform designed to address these issues by leveraging deep learning and vegetation indices. By transforming low-cost smartphones into diagnostic devices, the platform provides real-time support and connects farmers directly to a merchant marketplace.
Methodology
The AgroTIC platform utilizes a Model-View-ViewModel (MVVM) architecture to facilitate seamless interaction between farmers, agronomists, and merchants. The system's core diagnostic capability is built on two pillars of computer vision:
- Vegetation Indices (VEIs): The application processes standard RGB images to calculate the Triangular Greenness Index (TGI) and the Green-Red Vegetation Index (GRVI). These indices serve as proxies for chlorophyll and nitrogen levels, metrics typically requiring expensive multispectral sensors.
- Deep Learning: A custom Convolutional Neural Network (CNN) classifies five common citrus diseases and deficiencies: Alternaria, Acarus, Canker, Magnesium deficiency, and Zinc deficiency.
The system incorporates a "human-in-the-loop" approach where AI-generated diagnostics are validated by professional agronomists via a real-time chat module. Heavy computational tasks, such as CNN inference and OpenCV image processing, are offloaded to a cloud infrastructure (NodeJS/AWS S3) to ensure high performance on budget hardware.
Dataset and Training
The researchers developed a specialized dataset comprising 1,250 images of citrus plants collected from Colombian farms. The dataset includes healthy specimens alongside five disease/deficiency categories. To ensure robustness against varying field conditions (lighting, angles), the data underwent extensive augmentation, including rotations and zooms. The model was trained using an 80/20 train-test split.
Results and Discussion
The AgroTIC platform was evaluated over an 11-month pilot in Simacota, Colombia. The CNN achieved an overall classification accuracy of 93%. As shown in the confusion matrix, even the most challenging class, Canker, maintained a high accuracy of 88%.
Beyond technical metrics, the platform demonstrated significant socio-economic impact:
- Crop Loss Reduction: 74 out of 200 participating farmers reported measurable reductions in crop losses.
- Market Integration: The platform facilitated connections between 102 farmers and 12 merchants, enabling the sale of 65 different agricultural products.
- Expert Engagement: Over 1,350 messages were exchanged, highlighting a strong demand for remote technical assistance.
Conclusion
AgroTIC demonstrates that high-end precision agriculture can be democratized using accessible hardware. By combining mathematical transformations of RGB data with deep learning and a collaborative social ecosystem, the project successfully bridged technical and economic barriers. Future work will focus on expanding the diagnostic models to other regional crops like coffee and cocoa, and improving functionality in areas with limited internet connectivity.
Summary updated