Annotation-Efficient Weed Mapping in Sorghum Fields Using Two-Stage U-Net Crop Segmentation and HSV Greenness Analysis

my-project/│├── README.md ← General description and usage│├── install/│ ├── install.sh ← Script to install Anaconda│ └── installEnv.sh ← Create environment and install dependencies│├── dataset/│ ├── trainval/ ← Training + validation patches (BBCH 17, 256×256, augmented)│ │ ├── img/ ← 31,680 RGB patches (JPG)│ │ ├── mask/ ← 31,680 multiclass RGB masks (PNG)│ │ │ Classes: crop, weed, background│ │ │ COLOR_CROP = (31, 119, 180)│ │ │ COLOR_BACKGROUND = (199, 199, 199)│ │ │ COLOR_WEED = (14, 127, 255)│ │ ├── mask_crop/ ← 31,680 crop binary masks (PNG)│ │ └── mask_weed/ ← 31,680 weed binary masks (PNG)│ ││ ├── test/ ← Test patches (BBCH 17)│ │ ├── img/ ← 2,310 RGB patches (JPG)│ │ ├── mask/ ← 2,310 multiclass RGB masks (PNG)│ │ ├── mask_crop/ ← 2,310 crop binary masks (PNG)│ │ └── mask_weed/ ← 2,310 weed binary masks (PNG)│ ││ └── test_different_bbch/│ ├── bbch15_patch/│ │ ├── img/ ← 110 RGB patches (JPG)│ │ ├── mask/ ← 110 multiclass masks (PNG)│ │ ├── mask_crop/ ← 110 crop binary masks (PNG)│ │ └── mask_weed/ ← 110 weed binary masks (PNG)│ └── bbch19_patch/│ ├── img/ ← 110 RGB patches (JPG)│ ├── mask/ ← 110 multiclass masks (PNG)│ ├── mask_crop/ ← 110 crop binary masks (PNG)│ └── mask_weed/ ← 110 weed binary masks (PNG)│├── models/│ ├── resnet.py ← U-Net with ResNet encoders (18, 34, 50, 101, 152)│ ├── densenet.py ← U-Net with DenseNet encoders (121, 161, 169, 201)│ ├── efficientnet.py ← U-Net with EfficientNet encoders (B0–B7)│ ├── swin.py ← U-Net with Swin Transformer encoders (swin_t, swin_s)│ └── vit.py ← U-Net with Vision Transformer encoders (ViT_B_16, ViT_B_32)│├── dataloader.py ← Dataloaders, batch size, train/validation splits├── metrics.py ← Focal loss, accuracy, recall, precision, F1-score, IoU, Dice score│├── train.py ← Sequential training of all backbones with cross-validation;│ best model selected by validation focal loss│├── test_binary_crop.py ← Evaluate best crop model on BBCH 17 test set and save predicted│ crop binary masks and crop-removed RGB images│├── HSV_filtering.py ← Apply HSV thresholding and small-object removal on crop-removed│ images to detect weeds and save weed masks + metrics│├── combining_masks.py ← Combine predicted crop and weed binary masks into a single│ multiclass RGB mask (crop, weed, background) using:│ COLOR_CROP = (31, 119, 180)│ COLOR_BACKGROUND = (199, 199, 199)│ COLOR_WEED = (14, 127, 255)│└── test_multi_class.py ← Evaluate reconstructed multiclass predictions against ground truth and report macro-averaged metrics + confusion matrix (%)