Flagship · Undergraduate Thesis · 3-person team · 2025–2026
Rice Up
A Web Application for Rice Leaf Disease Detection Using CNN and Farming Support for Filipino Palay Farmers in Naic, Cavite.
Rice diseases are cheapest to fight when caught early — but palay farmers in Naic diagnose by eye and memory, with occasional technician visits. Existing agri-apps go unused because they assume connectivity, devices, and digital literacy rural farmers don't have. Rice Up puts a disease diagnosis in a farmer's pocket: photograph a rice leaf, get a classification — entirely in Filipino, designed with the farmers and the Municipal Agriculturist of Naic.
System architecture
Core feature — disease detection
The heart of Rice Up is a VGG16 transfer-learning classifier trained on 10,689 rice leaf images across nine disease categories, referenced against the DA–PhilRice field guide to keep every diagnosis aligned with officially recognized agronomy. VGG16 was selected after head-to-head evaluation against ResNet50 and EfficientNetB4 — it converged most reliably and produced the most stable validation scores under our compute constraints.
A farmer uploads a leaf photo; the frontend sends it to the isolated Flask inference service, where the image is resized, normalized, and edge-enhanced to amplify lesion and streak patterns before classification. The model performs strongest on diseases with distinct visual signatures — Tungro, Brown Spot, Leaf Scald — and the confusion matrix shows exactly where it struggles: visually overlapping lesion types like Leaf Blast and Narrow Brown Leaf Spot, the known frontier for expanding the dataset.
- Bacterial Leaf Blight
- Brown Spot
- Healthy Rice Leaf
- Leaf Blast
- Leaf Scald
- Narrow Brown Leaf Spot
- Rice Hispa
- Sheath Blight
- Tungro
Measured results
- 76%
- Disease classifier accuracy
- 95%
- Pest model v2 accuracy
- 5.03%
- Price forecast MAPE
- 4.22/5
- ISO/IEC 25010 quality score
9-class VGG16 · 10,689 images
Up from 87% after dataset rebalancing
226 records · significant at 95% CI
55 respondents · all 9 categories rated High
Engineering highlight — reading past the headline metric
Pest model v1 reported 87% accuracy — but its precision–recall asymmetry revealed overfitting to a 2,000-vs-300 class imbalance: the model rarely missed pests because it called nearly everything a pest. We rebuilt the dataset to a near 1:1 ratio, applied heavy augmentation, and retrained. v2 reached 95% accuracy with balanced per-class F1 — a reminder that a single headline number can hide a model that doesn't generalize.
Supporting modules
- Pest detection
- A separate binary CNN answering the question farmers actually asked: is there pest damage, yes or no? Deliberately simple to keep the diagnostic flow accessible.
- AskJuan — Filipino-language assistant
- Structured library of predefined, selectable Filipino Q&As curated from DA–PhilRice, the IRRI Rice Knowledge Bank, and extension-worker interviews — chosen over a free-form chatbot to guarantee accuracy.
- Farming calendar
- Per-user task scheduling (Magtanim, Mag-abono, Magpatubig, Mag-ani) on Firebase Authentication and Firestore, guided by documented Philippine rice crop durations.
- Price listing & forecasting
- Node.js backend serving cleaned Department of Agriculture Bantay Presyo data, with a linear regression module (R² 0.56, MAPE 5.03%, p < 0.05) estimating short-term price movement.
My role: backend architecture, database design, and cloud implementation; building the VGG16 disease model and integrating it into the Flask REST API; and developing the price regression module, AskJuan, and the farming calendar — while my teammates led field research and dataset curation, from farmer interviews to the AskJuan knowledge base.
Recognition
- 1st Place — Technology Exhibit at CONTRENDS 2026
- 2nd Place — Best Thesis under the Department of Computer Science (DLSU-D), 2026
- Best Paper Presentation Finalist — ICAI 2026
Scoped honestly: Rice Up is a supplementary decision-support tool — it does not replace agronomists, field technicians, or laboratory diagnostics.
demo note: the platform is fully live; disease and pest inference runs on an on-demand cloud runtime — heavyweight CNN models exceed free-tier hosting, so inference was isolated as a separate service and the platform degrades gracefully when it's offline.