---

license: other language: - en task_categories: - image-classification tags: - plant-disease - agriculture - crops - leaf - banana - tomato - potato - lemon - tea - rice - watering - irrigation - treatment size_categories: - 100K<n<1M format: image modality: - image library_tags: - pytorch - tensorflow

CropHelth — Unified Crop Disease & Watering Dataset (Round-1 Pre-training)

Author: Hansaka Rasanjana

Project Context: First-year IoT group mini-project at the Sri Lanka Institute of Information Technology (SLIIT)

---

~125,000 plant leaf images across 109 disease/healthy classes for 26 crops, complete with a per-image provenance manifest, treatment knowledge base, and per-crop watering recommendation rules.

Built as the core pre-training and round-1 dataset for our CropHelth system: an image classifier that outputs structured JSON code (e.g., {"status": {"code": "potato_lb"}, "treatment": {"pill_code": ...}}) alongside an environment-driven watering decision engine.

Layout

images/<plant>_<ill>/NNNNNN.jpg      # 109 classes, ~125k images (the training set)
_source_manifest.csv                 # per-image provenance: target_code, source, original_path
detection/                           # PlantDoc object-detection dataset (VOC: .jpg + .xml, TRAIN/TEST)
knowledge/
  names.json                         # class code -> human-readable message
  treatments.json                    # class code -> {treatment: {pill_code, pill_name}, agronomy}
  watering_rules.json                # per-crop thresholds for the watering decision engine
  watering_recommendations.csv       # human-readable watering reference table (ph, temp, moisture, practice)
labels.json                          # class list + code->index + per-class counts
README.md

Class code scheme

<plant>_<ill> — e.g., potato_late_blight, banana_sigatoka, tea_helopeltis, tomato_healthy.

• *_healthy = healthy class for that specific plant (used as the model's baseline "healthy" output).
• Plant = first token of the code.
• 109 classes total, neatly organized into one folder per class so you can drop images/ straight into any standard classification pipeline (ImageFolder or our custom CropHelth trainer work out of the box).

Sources (all open source)

Source	DOI / URL	Contribution
Mendeley tywbtsjrjv — PlantVillage subset	https://data.mendeley.com/datasets/tywbtsjrjv/1	apple, blueberry, cherry, corn, grape, orange, peach, pepper, potato, raspberry, soybean, squash, strawberry, tomato (~52k)
Mendeley 9tb7k297ff — BananaLSD	https://data.mendeley.com/datasets/9tb7k297ff/1	banana (sigatoka, cordana, pestalotiopsis, healthy; original set only — augmented sets excluded to prevent synthetic duplicate bias)
Mendeley 744vznw5k2 — teaLeafBD	https://data.mendeley.com/datasets/744vznw5k2/4	tea (7 classes)
Mendeley j32xdt2ff5 — Tea Sickness	https://data.mendeley.com/datasets/j32xdt2ff5/2	tea (8 classes, merged with teaLeafBD)
Mendeley 6243z8r6t6 — Multi-Crop Disease (Roboflow export)	https://data.mendeley.com/datasets/6243z8r6t6/1	banana, cauliflower, chilli, peanut, radish (~22k; classes mapped via YOLO label files)
Mendeley 643f5bbc2t — BDLemonLeaf	https://data.mendeley.com/datasets/643f5bbc2t/6	lemon (13 classes, raw image set)
Mendeley 8d9fv6kpt3 — Large-Scale Lemon Leaf Disease & Pest	https://data.mendeley.com/datasets/8d9fv6kpt3/1	lemon (18 classes, merged with BDLemonLeaf)
Mendeley c5yvn32dzg — RoCoLe (Robusta coffee)	https://data.mendeley.com/datasets/c5yvn32dzg/2	coffee healthy/unhealthy (labels extracted from shipped Labelbox CSV)
Mendeley 22p2vcbxfk — Groundnut (peanut)	https://data.mendeley.com/datasets/22p2vcbxfk/3	peanut (5 classes, raw data)
Mendeley 7vpdrbdkd4 — Nutrient-deficient banana	https://data.mendeley.com/datasets/7vpdrbdkd4/2	banana nutrient deficiencies (7 classes, raw set only)
Mendeley g7xnn2bm4g — Nitrogen deficiency in maize	https://data.mendeley.com/datasets/g7xnn2bm4g/1	maize N0/N7/NF (3 classes)
GitHub pratikkayal/PlantDoc-Dataset	https://github.com/pratikkayal/PlantDoc-Dataset	28 classes (train + test merged)
GitHub pratikkayal/PlantDoc-Object-Detection-Dataset	https://github.com/pratikkayal/PlantDoc-Object-Detection-Dataset	detection/ (VOC format, separate task)
Harvard Dataverse LQUWXW — Bananas TZ (Tanzania)	https://doi.org/10.7910/DVN/LQUWXW	banana black sigatoka, fusarium wilt, healthy (~15k high-res photos)

Excluded by design: Background_without_leaves (from PlantVillage, as it's not a plant class), augmented/synthetic image packs (our trainer handles its own augmentations), and the Zenodo DiaMOS pear dataset (a massive 12.5 GB single archive that couldn't fit cleanly into our local build pipeline—feel free to drop it in manually if needed).

Note: The IEEE DataPort Paddy Doctor (rice) dataset requires a verified IEEE account login, so it couldn't be auto-fetched in our script. Rice classes should be added following the same manifest layout once downloaded.

Watering decision data

knowledge/watering_rules.json acts as the brain for our decision engine (water_now / hold_water / check_data):

• soil_moisture_pct < min → triggers water_now
• soil_moisture_pct > max → triggers hold_water (prevents overwatering root rot)
• missing soil moisture input → returns check_data
• ph / air_temp_c / soil_temp_c out of bounds → flags extra warning reasons (ph_low, ph_high, air_temp_high, soil_temp_high)

Per-crop thresholds (moisture %, pH target windows, and temperature safety caps) live in watering_rules.json, accompanied by a human-readable reference table in watering_recommendations.csv. These are baseline agronomic guidelines—calibrate them against physical field sensors for production use. Soil moisture serves as our primary signal; temperature or pH changes alone shouldn't trigger automated irrigation.

Treatments

knowledge/treatments.json maps every non-healthy class to structured guidance ({treatment: {pill_code, pill_name}, agronomy}). These are standard extension-service recommendations meant for training and testing. Before any live deployment, make sure to verify that suggested products are locally registered, apply strict label rates, rotate FRAC groups to prevent resistance, and consult a local agronomist. The model outputs the diagnostic class code only; treatment text is statically mapped from this secure file (never dynamically generated by the model).

Provenance & licenses

_source_manifest.csv ties every individual image back to its precise source dataset and original path. Each source retains its own license (check detection/LICENSE.txt for PlantDoc; Mendeley datasets generally fall under CC-BY/CC0). If you build on this or redistribute, please keep the attribution table intact.

Quick usage (CropHelth Trainer)

# point the pipeline at this dataset structure
python scripts/build_labels.py --images-dir <this repo>/images
python scripts/split.py
python scripts/train.py
python scripts/predict.py --image some_leaf.jpg --meta
python scripts/predict_watering.py --json '{"crop":"potato","soil_moisture_pct":18,"ph":6.1}'