Complete Parameter Reference

This page provides a consolidated reference of all parameters across the three plugin tabs.

Patch Extraction Parameters

Parameter	Type	Default	Description
mode	string	single	Extraction mode: single or batch
image_path	path	—	Path to the satellite GeoTIFF (single mode)
label_path	path	—	Path to the label GeoTIFF (single mode)
grid_path	path	—	Path to the grid shapefile
data_dir	path	—	Path to data folder containing image/label pairs (batch mode)
output_dir	path	—	Output folder for extracted patches
patch_size	int	256	Output patch size in pixels (width = height)
in_channels	int	4	Expected number of image bands (for validation)
interpolation	string	bilinear	Resampling for image patches: bilinear, nearest, bicubic, lanczos
compression	string	deflate	Output TIFF compression: deflate, lzw, none
validate_crs	bool	true	Warn on CRS mismatch between image, label, and grid
id_column	string	id	Grid attribute column used for patch naming
train_ratio	float	0.70	Fraction of patches for training (0.0–1.0)
val_ratio	float	0.20	Fraction of patches for validation (0.0–1.0)
test_ratio	float	0.10	Fraction of patches for testing (0.0–1.0)
use_per_image_grid	bool	false	Use per-image Grid_N.shp files (batch mode)
image_pattern	regex	image_(\d+)	Regex for matching image files (batch mode, advanced)
label_pattern	regex	label_(\d+)	Regex for matching label files (batch mode, advanced)

Model Training Parameters

Basic Parameters

Parameter	Type	Default	Description
dataset_root	path	—	Root directory of the patch dataset (must contain train/, val/)
mode	string	binary	binary or multiclass
num_classes	int	2	Number of classes (multiclass mode only; ignored for binary)
model_name	string	unet-dropout	Segmentation architecture name (see Architectures & Encoders)
encoder_name	string	resnet34	Encoder backbone (SMP architectures and unetformer only)
pretrained	bool	false	Use ImageNet pretrained encoder weights
in_channels	int	4	Number of input image channels
epochs	int	100	Maximum training epochs
batch_size	int	4	Training batch size
learning_rate	float	0.0003	Initial learning rate (Adam optimizer)
device	string	cuda	cuda or cpu
save_dir	path	./trained_models	Directory for saving checkpoints and plots
scheduler	string	reduce_on_plateau	LR scheduler: reduce_on_plateau, cosine_annealing, one_cycle, none
loss_function	string	binary_dice_bce	Loss function name (see Loss Functions)
augmentation_level	string	basic	Augmentation level: none, basic, advanced, aggressive, extreme
use_class_weights	bool	false	Compute and apply per-class weights
freeze_encoder	bool	false	Freeze encoder for first N epochs
freeze_epochs	int	5	Number of epochs to keep encoder frozen
early_stopping	bool	true	Enable early stopping
patience	int	15	Early stopping patience (epochs without improvement)

Advanced Parameters

Parameter	Type	Default	Description
dropout_rate	float	0.3	Dropout probability (0.0 = disabled)
use_amp	bool	false	Enable mixed precision training (FP16, GPU only)
warmup_epochs	int	0	Number of LR warmup epochs
focal_gamma	float	2.0	Focal loss gamma parameter (0.5–5.0)
focal_alpha	float	0.25	Focal loss alpha parameter (0.1–0.9)
tversky_alpha	float	0.3	Tversky loss false positive weight
tversky_beta	float	0.7	Tversky loss false negative weight
use_kfold	bool	false	Enable K-Fold cross-validation
n_splits	int	5	Number of folds for K-Fold CV

Prediction Parameters

Parameter	Type	Default	Description
model_path	path	—	Path to the trained .pth checkpoint
input_path	path	—	Path to the input satellite GeoTIFF
output_path	path	—	Path for the output segmentation GeoTIFF
patch_size	int	512	Inference patch size in pixels (64–2048)
overlap	int	128	Overlap between adjacent patches in pixels
threshold	float	0.50	Decision threshold for binary mode (0.0–1.0)
device	string	cuda	cuda or cpu
batch_size	int	1	Inference batch size (number of patches per forward pass)
gaussian_blending	bool	true	Apply Gaussian blending when merging patch predictions
save_confidence	bool	false	Save raw probability map alongside the classification result
dropout_rate	float	0.3	Dropout rate (must match training value)
encoder_name	string	(auto)	Override encoder name (leave empty for auto-detection from checkpoint)