| """ |
| Vanilla U-Net for RockGlacierNet — baseline model. |
| |
| Standard Ronneberger et al. (2015) U-Net: encoder/decoder with plain skip |
| connections (NO attention gates, NO transformer blocks). Kept architecturally |
| identical to the Attention U-Net (same filters/depth) so the ONLY difference is |
| the attention mechanism — a fair baseline comparison. |
| |
| Input : (128, 128, N_BANDS) |
| Output: (128, 128, 1) sigmoid — binary glacier mask |
| |
| Encoder filters : 64 -> 128 -> 256 -> 512 |
| Bottleneck : 512 |
| Decoder filters : 256 -> 128 -> 64 -> 32 |
| """ |
|
|
| import tensorflow as tf |
| from tensorflow.keras.layers import ( |
| Input, Conv2D, Conv2DTranspose, MaxPooling2D, |
| BatchNormalization, Activation, concatenate |
| ) |
| from tensorflow.keras.models import Model |
|
|
|
|
| def _double_conv(x, filters, dropout=0.0, name=""): |
| """Two Conv3x3 + BN + ReLU blocks (standard U-Net block).""" |
| x = Conv2D(filters, 3, padding='same', name=f'{name}_c1')(x) |
| x = BatchNormalization(name=f'{name}_bn1')(x) |
| x = Activation('relu', name=f'{name}_r1')(x) |
| if dropout > 0: |
| x = tf.keras.layers.Dropout(dropout, name=f'{name}_dp')(x) |
| x = Conv2D(filters, 3, padding='same', name=f'{name}_c2')(x) |
| x = BatchNormalization(name=f'{name}_bn2')(x) |
| x = Activation('relu', name=f'{name}_r2')(x) |
| return x |
|
|
|
|
| def build_unet(input_shape=(128, 128, 12), num_classes=1): |
| """Plain U-Net (no attention). Signature matches build_attention_unet.""" |
| inputs = Input(input_shape, name='input') |
|
|
| |
| e1 = _double_conv(inputs, 64, dropout=0.1, name='enc1') |
| p1 = MaxPooling2D(2, name='pool1')(e1) |
| e2 = _double_conv(p1, 128, dropout=0.1, name='enc2') |
| p2 = MaxPooling2D(2, name='pool2')(e2) |
| e3 = _double_conv(p2, 256, dropout=0.2, name='enc3') |
| p3 = MaxPooling2D(2, name='pool3')(e3) |
| e4 = _double_conv(p3, 512, dropout=0.2, name='enc4') |
| p4 = MaxPooling2D(2, name='pool4')(e4) |
|
|
| |
| b = _double_conv(p4, 512, dropout=0.3, name='btn') |
|
|
| |
| u4 = Conv2DTranspose(256, 2, strides=2, padding='same', name='up4')(b) |
| d4 = _double_conv(concatenate([e4, u4], name='cat4'), 256, dropout=0.2, name='dec4') |
| u3 = Conv2DTranspose(128, 2, strides=2, padding='same', name='up3')(d4) |
| d3 = _double_conv(concatenate([e3, u3], name='cat3'), 128, dropout=0.2, name='dec3') |
| u2 = Conv2DTranspose(64, 2, strides=2, padding='same', name='up2')(d3) |
| d2 = _double_conv(concatenate([e2, u2], name='cat2'), 64, dropout=0.1, name='dec2') |
| u1 = Conv2DTranspose(32, 2, strides=2, padding='same', name='up1')(d2) |
| d1 = _double_conv(concatenate([e1, u1], name='cat1'), 32, dropout=0.1, name='dec1') |
|
|
| outputs = Conv2D(num_classes, 1, activation='sigmoid', name='output')(d1) |
| return Model(inputs, outputs, name='UNet') |
|
|
