FCNN architecture used in the paper: "Transferring Deep Learning models for cloud detection between Landsat-8 and Proba-V".

transfer_learning_unet.py

"""
This file contains the fully convolutional neural network architecture used in the paper:
"Transferring Deep Learning models for cloud detection between Landsat-8 and Proba-V" (Mateo-García et al 2019)

The function UNet builds the FCNN and compiles the model with the loss, the optimizer and the weight decay
used in the paper.

"""
from tensorflow import keras

AXIS_CHANNELS = 3

def _conv_blocks(ip_, nfilters, reg, name):
    conv = keras.layers.SeparableConv2D(
        nfilters, (3, 3),
        padding='same',
        name=name+"_conv_1",
        depthwise_regularizer=reg,
        pointwise_regularizer=reg)(ip_)

    conv = keras.layers.BatchNormalization(axis=AXIS_CHANNELS,
                                           name=name + "_bn_1")(conv)
    conv = keras.layers.Activation('relu',name=name + "_act_1")(conv)
    conv = keras.layers.SeparableConv2D(
        nfilters, (3, 3),
        padding='same',
        name=name+"_conv_2",
        depthwise_regularizer=reg,
        pointwise_regularizer=reg)(conv)
    conv = keras.layers.BatchNormalization(axis=AXIS_CHANNELS,
                                           name=name + "_bn_2")(conv)
    return keras.layers.Activation('relu',name=name + "_act_2")(conv)

def UNet(input_shape=(None,None,4), weight_decay=5e-4,lr=1e-5):
    inputs_ = keras.layers.Input(shape=input_shape)
    reg = keras.regularizers.l2(weight_decay)
    conv1 = _conv_blocks(inputs_,
                         32,
                         reg,
                         name="input")
    pool1 = keras.layers.MaxPooling2D(pool_size=(2, 2),
                                      name="pooling_1")(conv1)
    conv2 = _conv_blocks(pool1,
                         64, reg, name="pool1")
    pool2 = keras.layers.MaxPooling2D(pool_size=(2, 2),
                                      name="pooling_2")(conv2)
    conv3 = _conv_blocks(pool2,
                         128, reg, name="pool2")
    up8 = keras.layers.concatenate([
        keras.layers.Conv2DTranspose(64, (2, 2),
                                     strides=(2, 2),
                                     padding='same',name="upconv1",
                                     kernel_regularizer=reg)(conv3),
        conv2],
        axis=AXIS_CHANNELS, name="concatenate_up_1")
    conv8 = _conv_blocks(up8,
                         64,reg, name="up1")
    up9 = keras.layers.concatenate([
        keras.layers.Conv2DTranspose(32, (2, 2),
                                     strides=(2, 2),
                                     padding='same',name="upconv2",
                                     kernel_regularizer=reg)(conv8),
        conv1],
        axis=AXIS_CHANNELS, name="concatenate_up_2")
    conv9 = _conv_blocks(up9,
                         32, reg, name="up2")
    conv10 = keras.layers.Conv2D(1, (1, 1),
                                 kernel_regularizer=reg, name="linear_model",
                                 activation="sigmoid")(conv9)
    model = keras.models.Model(inputs=[inputs_], outputs=[conv10],
                               name="FCNNMateo-Garcia et al")
    sgd = keras.optimizers.Adam(lr=lr)
    model.compile(loss=keras.losses.binary_crossentropy,
                  optimizer=sgd,
                  metrics=["binary_accuracy"])
    return model