Carly CHÉRY
CV

Carly CheryResearch

Research

Deep learning and computer vision for plants, and the imaging and sensing hardware behind them. A statement of what I study, followed by the three projects it grew from.

Research Statement

Teaching machines to see plants
through images and sensors.

I am a final-year agricultural sciences student at EARTH University in Costa Rica, working in deep learning and computer vision for plants. I turn plant images and sensor data into measurements a grower can act on, and I build the imaging and sensing hardware those models depend on.

I came to this research from two directions. From Haiti, where I grew up watching coffee, mango, and breadfruit production constrained by disease that was rarely measured until it was too late. From Costa Rica, where EARTH's focus on tropical agriculture met my own curiosity about machine learning. The question I want to work on as a scientist: how can cameras, sensors, and deep learning see plant stress and disease early, accurately, and cheaply enough for the farms that need it most?

Two threads in particular pull at me:

  1. Plant pathology & computer vision. My graduation thesis applies knowledge-distilled vision transformers to ordinal severity grading of coffee leaf disease, Hemileia vastatrix (roya) and Mycena citricolor (ojo de gallo). Next, I want to extend this into multi-disease, multi-species pipelines, a single model that diagnoses tropical staple-crop pathologies from a smallholder's smartphone, with calibrated uncertainty.
  2. Plant sensing & controlled-environment engineering. My internship at Auburn's E.W. Shell Fisheries Center put me on an aquaponic system coupling tilapia and greenhouse tomato, where I built the sensor data-logging workflow the team used, and my Arduino pH controller automated pH regulation in a hydroponic system. Next, I want to combine plant images with sensor data so that a growing system can see plant stress and correct its environment on its own.

Both threads rest on a shared commitment: reproducible, field-deployable methods. Calibrated imaging, preregistered evaluation, open hardware designs, and models small enough to run offline on a phone.

The research I want to do puts deep learning, cameras, and sensors in service of plants: rigorous enough to trust, and simple enough to deploy on a smallholder's farm.

Featured Research

Three lines of work.

The research I have built so far. Each is grounded in a real problem and a real population, and each is the seed of a longer programme.

Ordinal severity grading of coffee leaf disease via knowledge-distilled vision transformers.

B.Sc. graduation thesis · EARTH University · 2026 · First-author manuscript in preparation

A deep-learning study applying knowledge-distilled vision transformers (ViT) to ordinal severity grading of Hemileia vastatrix (Roya) and Mycena citricolor (Ojo de Gallo) on field-captured coffee leaves.

Knowledge distillation transfers learned representations from a large pretrained ViT teacher into a smaller student network suitable for smartphone-class deployment. Ordinal regression heads preserve the natural ranking of severity classes, a structure that standard cross-entropy discards. Coffee leaf rust alone has cost Latin American producers an estimated USD 3.2 billion since the 2012 outbreak (Avelino et al., 2015); this work targets smallholder-deployable diagnostics. Manuscript in preparation, first author. Target venues: Computers and Electronics in Agriculture or Agriculture.

Keywords Vision Transformer (ViT) · knowledge distillation · ordinal regression · plant pathology · coffee · severity grading
Methodology
Open Hardware Two-column gantry v5 · matte-black PLA · OpenSCAD-modelled · all STLs released under CC BY 4.0 · 20 cm working distance · ColorChecker NE corner · build cost under USD 20.
Methodology flow diagram: Field Capture → Annotation → Model Training → Evaluation → Deployment 01 Field Capture 1,400 leaves iPhone 11 + 3D-printed rig ColorChecker NE corner 4,200 images · 0° / 45° / abaxial 02 Annotation EMSAM → CVAT → STAPLE 3 phytopathologists 10% triplicate subset Krippendorff α ≥ 0.80 03 Model Training SegFormer-B3 (45M params) FastViT-T8 (4M params) DINOv2-LoRA teacher CORN ordinal head 04 Evaluation 3-fold grouped CV BCa paired bootstrap TOST equivalence ±0.05 power 0.97 at Δκ = 0 05 Deployment LiteRT (Android) Core ML (iOS) on-device inference 113× fewer FLOPs
  1. 01

    Field Capture

    1,400 leaves · iPhone 11 + 3D-printed rig · ColorChecker · 4,200 images at 0°/45°/abaxial.

  2. 02

    Annotation

    EMSAM → CVAT → STAPLE consensus · 3 phytopathologists · 10% triplicate subset · Krippendorff α ≥ 0.80.

  3. 03

    Model Training

    Method A: SegFormer-B3 (45M) · Method B: FastViT-T8 (4M) with DINOv2-LoRA teacher and CORN ordinal head.

  4. 04

    Evaluation

    3-fold grouped CV · BCa paired bootstrap · TOST equivalence · ±0.05 margin · power 0.97 at Δκ = 0.

  5. 05

    Deployment

    FastViT to LiteRT (Android) / Core ML (iOS) · on-device inference, fully offline · 113× fewer FLOPs.

Expected results
Model size vs grading accuracy

5 architectures plotted on the params–κ plane. The 4 M-parameter FastViT-T8 is hypothesised to match the 45 M SegFormer-B3 to within ±0.05 κ.

0.90 0.85 0.80 0.75 0.70 grading κ 1 M 10 M 100 M parameters (log) ResNet-18 U-Net DINOv2 probe FastViT-T8 SegFormer-B3 equivalence pair
Preregistered target thresholds

Five preregistered floors; all reported regardless of outcome. Bars show target value on the 0–1 scale.

0.0 0.2 0.4 0.6 0.8 1.0 QWK ≥ 0.80 mIoU ≥ 0.75 Dice ≥ 0.80 per-class F1 ≥ 0.85 Krippendorff α ≥ 0.80
Equivalence test: expected result

Two One-Sided Tests (TOST) with equivalence margin ±0.05. Expected: 95% CI of Δκ falls fully within the shaded zone.

−0.05 +0.05 0 Δκ ≈ 0 · 95% CI −0.15 −0.05 0 +0.05 +0.15 Δκ (SegFormer − FastViT)
Preregistered statistical power

Probability of correctly declaring equivalence at N = 1,400. Power = 0.97 at Δ = 0, dropping outside ±0.05.

80% power floor power = 0.97 at Δ = 0 100% 75% 50% 25% 0% −0.15 −0.05 0 +0.05 +0.15 true Δκ (SegFormer − FastViT)
Preliminary results · Ojo de Gallo Real held-out test set

A first working system for the thesis' second target disease, Mycena citricolor (ojo de gallo). It segments necrotic lesions pixel by pixel, measures severity as lesion area ÷ leaf area, and assigns a G0–G3 grade, running fully offline on a phone.

100%
Grade accuracy · QWK 1.00
0.94
Peak lesion Dice · SegFormer-B5 (324 MB)
17MB
Deployed model · 100% offline
8
Models benchmarked · 5 architectures
Grid comparing lesion detection by three DeepLabV3+ encoders across four coffee leaves from healthy (G0) to severe (G3), with red lesion outlines and severity percentages
Held-out test Lesion detection across four leaves (G0 to G3) and three encoders: DeepLabV3+ with ResNet-50, MobileNet-512 and MobileNet-1024. Red outlines mark detected lesions; grade and severity % appear under each panel.
Two working products
Offline Android app screen grading an ojo de gallo leaf
Field app On-device Android app: fully offline, with GPS geotagging, survey sessions, three languages and CSV/PDF export.
Web grading tool reading a severe coffee leaf at 51.9 percent, grade G3, with lesion overlay and pixel areas
Web tool Web grading tool: a severe leaf read at 51.9% · G3 by the deployed MobileNet model, with lesion overlay and pixel areas.
Model benchmark

Seven segmentation models plus a transformer classifier, all on the same held-out test set, ranked by lesion Dice. The deployed 17 MB model sits third, behind only a 106 MB and a 324 MB network.

Model Grade QWKDiceIoU Size On-device
SegFormer-B5 (1024px)100%1.000.9400.916324 MB·
DeepLabV3+ ResNet-50 (1024px)100%1.000.9330.906106 MB·
DeepLabV3+ MobileNetV2 (1024px)100%1.000.9240.89517 MBYes
SegFormer-B5 (512px)100%1.000.9110.878324 MB·
DeepLabV3+ EfficientNet-B7 (512px)96.4%0.9440.8930.851242 MB·
DeepLabV3+ MobileNetV2 (512px)98.2%0.9720.8920.85017 MBYes
DeepLabV3+ ResNet-50 (512px)98.2%0.9720.8910.856~100 MB·
PLA-ViT (ViT-B/16 classifier)98.2%0.972n/an/a328 MB·

Per-leaf metrics on the real held-out test set (worst-face aggregation). Grade accuracy saturates at 100% for several models, so lesion Dice and model size are the true differentiators. PLA-ViT is a whole-leaf classifier and produces no lesion map.

What the results show
  1. Small beats large. A 17 MB MobileNet matched or beat ResNet-50 (106 MB), EfficientNet-B7 (242 MB), SegFormer-B5 (324 MB) and a ViT classifier (328 MB) on grade, and ranks third on lesion Dice. On this dataset, efficiency won, not capacity.
  2. Two paradigms converge. Direct ViT classification and segment-then-measure reached identical grade accuracy (98.2%, QWK 0.972); but only segmentation yields an interpretable severity map and a defensible percentage a researcher can verify.
  3. Field-ready, not lab-bound. The whole system runs offline on a phone (ONNX-Runtime WASM, no server, no signal) with GPS geotagging, survey-session incidence tracking, three languages and CSV/PDF export.

A decision-support research tool for coffee-leaf disease assessment, not a validated clinical diagnosis. Reported metrics are on flatbed scans; phone photos yield estimates.

Download full results brief (PDF) ↓ View bibliography ↗
Interactive demo of my graduation project

Test the deployed model on a coffee leaf

Pick one of my sample leaves, or upload your own coffee-leaf photo, and the segmentation model runs entirely on your device via ONNX Runtime Web. It localizes the ojo de gallo (Mycena citricolor) lesions, measures the affected leaf area, and returns an ordinal severity grade (G0–G3). No image ever leaves your browser.

Try the live model →

Integrated tilapia & greenhouse tomato, an aquaponic dataset from the E.W. Shell Fisheries Center.

Research Intern · Auburn University · Coordinator: Mollie R. Smith

A season's water-quality, yield, and fish-health data from a coupled aquaponic system, collected as part of the integrated aquaponics research program at Auburn's E.W. Shell Fisheries Center.

Designed and operated biometric monitoring protocols and contributed to data pipelines for real-time system performance evaluation. The dataset, pH, ammonia/nitrite/nitrate cycling, dissolved oxygen, fish length-weight, tomato yield, is the seed for the digital-twin modeling I want to pursue next.

Keywords Aquaponics · tilapia · greenhouse tomato · water-quality · digital twin
Gallery Auburn University · E.W. Shell Fisheries Center · Sep–Dec 2025 · click to enlarge

A smart-hydroponic pH-regulation controller built on Arduino, for a university electronics competition.

Team project · Carly Chery, Gédéon Jaboin, Ed Duddley Dumont · Cenfotec University, Costa Rica · 6th of 16 entries

A closed-loop pH stabilizer for nutrient-film hydroponics, dosing dilute acid and base via peristaltic pumps in response to a live probe reading on an Arduino board.

Hydroponic crops sit in a narrow pH window (5.5 – 6.5); drift outside that window costs yield within hours. The prototype reads a glass pH probe through an analog converter, runs a debounced setpoint loop on the Arduino, and triggers two peristaltic pumps (acid / base) in short pulses with a cool-down period to prevent over-shooting. An LCD shows the live reading and the last action. Built and tested against laboratory pH standards over a weekend, then submitted to the Cenfotec inter-university Arduino competition where it placed 6th of 16 projects.

Keywords Arduino · embedded systems · hydroponics · pH control · closed-loop dosing · peristaltic pumps · prototyping
Gallery Cenfotec University · Costa Rica · 2024 · click to enlarge