Skip to content

module deeplearning

function processing

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
processing(
    raster_path,
    label_path=None,
    patch_size=256,
    export_patches=False,
    output_dir=None,
    export_labels=False,
    labels_output_dir=None,
    padding_mode='constant',
    padding_value=0,
    overlap=0
)

Reads raster and labels, splits them into patches with optional overlap and padding.

Args:

  • raster_path (str): Path to raster file
  • label_path (str, optional): Path to label raster
  • patch_size (int): Patch size
  • export_patches (bool): Whether to export image patches
  • output_dir (str, optional): Directory to export image patches
  • export_labels (bool): Whether to export label patches
  • labels_output_dir (str, optional): Directory to export label patches
  • padding_mode (str): 'constant' or 'reflect'
  • padding_value: Value for constant padding
  • overlap (int): Number of pixels of overlap between patches

Returns:

  • tuple: (X_patches, y_patches) or X_patches if no labels are provided

function make_mean_iou_metric

1
make_mean_iou_metric(num_classes)

Create a stable Keras metric named 'mean_iou' to allow monitoring 'val_mean_iou'.

function trainUnet

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
trainUnet(
    X_train,
    y_train,
    input_shape,
    num_classes,
    dropout_rate=0.2,
    learning_rate=0.0001,
    batch_size=16,
    epochs=50,
    validation_data=None,
    validation_split=0.0,
    data_augmentation=False,
    normalize=True,
    fill_nulls=True,
    null_value=0.0,
    save_best_model=True,
    model_path='best_model.keras'
)

Train a U-Net model for binary or multiclass semantic segmentation.

Parameters ---------- X_train : np.ndarray Training feature patches of shape (N, H, W, C). y_train : np.ndarray Training labels of shape (N, H, W, 1) for binary or (N, H, W) with integer class labels for multiclass. input_shape : tuple Shape of input tensors (H, W, C). num_classes : int Number of output classes. Use 1 for binary segmentation. dropout_rate : float, optional Dropout rate for regularization (default=0.2). learning_rate : float, optional Learning rate for Adam optimizer (default=1e-4). batch_size : int, optional Number of samples per gradient update (default=16). epochs : int, optional Number of training epochs (default=50). validation_data : tuple, optional Tuple (X_val, y_val) used for validation. If None, a split from training data can be created with validation_split. validation_split : float, optional Fraction of training data reserved for validation if validation_data is not provided (default=0.0). data_augmentation : bool, optional If True, applies random flips and rotations for augmentation (default=False). normalize : bool, optional If True, applies per-band normalization using StandardScaler (default=True). fill_nulls : bool, optional If True, replaces NaN values with null_value (default=True). null_value : float, optional Value used to replace NaN values (default=0.0). save_best_model : bool, optional If True, saves the best model according to validation IoU (default=True). model_path : str, optional File path to save the best model (default="best_model.keras").

Returns ------- model : tensorflow.keras.Model The trained U-Net model. history : keras.callbacks.History Training history with loss and metric evolution.

function predictRaster

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
predictRaster(
    model,
    raster_path,
    patch_size=256,
    num_classes=1,
    output_path=None,
    fill_nulls=True,
    null_value=0,
    normalize=True,
    overlap=0
)

Predicts over a complete raster file and reconstructs the original image.

Args:

  • model: Trained model (U-Net)
  • raster_path (str): Path to the .tif file for prediction
  • patch_size (int): Patch size used during training
  • num_classes (int): Number of classes of the model
  • output_path (str): Path to save the predicted raster (optional)
  • fill_nulls (bool): Whether to replace null values with null_value
  • null_value: Value to replace nulls
  • normalize (bool): Whether to apply normalization
  • overlap (int): Number of overlapping pixels between patches

Returns:

  • np.array: Predicted reconstructed image

function reconstruct_from_patches

1
2
3
4
5
6
7
reconstruct_from_patches(
    patches,
    original_height,
    original_width,
    patch_size,
    overlap=0
)

Reconstructs a full image from patches, averaging values in overlapped regions.

function save_prediction_raster

1
2
3
4
5
6
save_prediction_raster(
    prediction,
    output_path,
    original_profile,
    original_transform
)

Saves prediction as a georeferenced raster.

This file was automatically generated via lazydocs.