Code and project structure for plant disease recognition under domain shift.
No data or pretrained weights are included. Bring your own datasets and follow the folder layout below.

Laboratory images (e.g., PlantVillage) often fail in real fields (e.g., PlantPathology) due to background/lighting/occlusion gaps.
This repo provides two complementary components:

1. FBR — Field‑Adaptive Background Recomposition:
   Segment foreground leaves with SAM, crop real‑field backgrounds, and composite them to create more field‑like training samples.
2. UDA — Unsupervised Domain Adaptation:
   Train with source labeled (lab) and target unlabeled (field) data using methods such as DDC, DCORAL, DANN, CDAN, and DALN.

Use the PlantVillage and PlantPathology datasets. Download them yourself and place files to match the layout below.

data/ └── apple/ ├── PV/ # lab‑collected (PlantVillage) │ ├── images/ # raw lab images (flat or class subfolders) │ ├── bg_composed/ # FBR outputs (composited images) │ ├── pv_masks.pickle # SAM masks index (e.g., {filename: mask}) │ └── pv_labels.pickle # labels for PV └── plantpathology/ # real‑field (PlantPathology) ├── images/ # raw field images ├── cropped_bg/ # background crops used for FBR └── apple_labels.pickle # labels for field set 

You may adapt subfolders per crop/task, but keep the same depth and names so scripts can find your data (see datasets/__init__.py).

Run the notebook to generate FBR samples:

• 01_FBR_field_adaptive_background_recomposition.ipynb
  1. Segment leaf foregrounds from lab images using SAM
  2. Randomly crop real-field backgrounds
  3. Composite foregrounds onto backgrounds → FBR dataset

Weights: large files (e.g., sam_vit_h_4b8939.pth) are not tracked. Download manually and set the local path in the notebook/config.

• Background crops → data/apple/plantpathology/cropped_bg/
  Files: *.jpg (e.g., IMG_1234_crop00_s256_x12_y98.jpg); size defaults to 256×256.
• SAM masks → data/apple/PV/pv_masks.pickle
  Python dict {filename: mask}, shape (H, W), dtype uint8 in {0,255} (post-processed for cleaner edges).
• Composed (FBR) images → data/apple/PV/bg_composed/
  Same filenames as PV/images/; lab foreground pasted onto a random field crop.

To train on composed images, set src_dataset.kwargs.root = data/apple/PV/bg_composed/.
For UDA, use real-field images under data/apple/plantpathology/images/ as the target domain.

python -m venv .venv # macOS/Linux source .venv/bin/activate # Windows # .venv\Scripts\activate pip install -r requirements.txt

• Download PlantVillage & PlantPathology
• Organize them to match the data layout above

Open and run: 01_FBR_field_adaptive_background_recomposition.ipynb

python main.py

Artifacts (checkpoints, logs) are saved under exp/.

⚙️ Configuration (pointer)

Default options live in utils/train_config.py.

Dataset types supported (see datasets/init.py): src_bg_augmented, src_lab, src_real, tgt, tst → each maps to the folders shown above.

Models supported: ddc, dcoral, dann, cdan, daln (with ResNet backbones; see models/).

. ├── datasets/ # dataset loaders & helpers │ ├── __init__.py # maps dataset "type" -> folder paths │ └── dataset.py ├── models/ # model builders (UDA variants) │ ├── layers/ # extra layers if any │ ├── uda_losses/ # UDA loss modules │ ├── ddc.py dcoral.py dann.py cdan.py daln.py ... │ └── resnet.py vanilaresnet.py ├── utils/ │ ├── dataloader_utils.py # class weights, distribution checks │ ├── train_utils.py # trainer & loops │ ├── transforms_utils.py # transforms/augmentations │ └── train_config.py # default training options ├── 01_FBR_field_adaptive_background_recomposition.ipynb ├── main.py ├── requirements.txt └── README.md 

This dataset was used to validate our framework on the challenging task of distinguishing Viral Infection (BBWV2) from Abiotic Stress (Flooding), which share visually similar symptoms.

• Download:
  • The dataset is hosted on Zenodo for long-term accessibility.
  • 📥 Download Link (Zenodo)
• Total Images: 4,102 images
• Class Distribution:
  • 0 Healthy: 1,325 images
  • 1 Abiotic Stress (Flooding): 1,642 images
  • 2 Viral Infection (BBWV2): 1,135 images
• Contributors: Data collection and annotation were conducted by Seong-Yeop Kim & Mungyeong Song (Seoul National University).

If you use this repository, please cite the following manuscript (under review):

@misc{jeon2025lab2field, title = {Bridging the Lab-to-Field Gap in Plant Disease Diagnosis through Unsupervised Domain Adaptation Enhanced by Background Recomposition}, author = {Jeon, Woosang and Kim, Taehyeong and Choi, Sanghyeok and Yang, Kyuseok and Kim, Seong-Yeop and Song, Mungyeong}, year = {2025}, note = {Manuscript under review. Corresponding author: taehyeong.kim@snu.ac.kr}, howpublished = {\url{https://github.com/LABA-SNU/FBR-UDA}} }