Moon CNN: Deep Learning Lunar Phase Recognition

Dataset Construction & Pipeline

To build this model, I manually created a dataset of 250 lunar training images and a validation dataset of 50 images, distributed evenly across four celestial classes:

• New Moon: Minimal illumination, edge shadow boundary detection.
• Full Moon: Complete circular reflectance with crater contrast.
• First Quarter: Right-hand illuminated semicircle.
• Last Quarter: Left-hand illuminated semicircle.

Data Augmentation Strategy

Because astronomical datasets can be scarce, data augmentation was essential to prevent severe overfitting. The augmentation sequential layer applies:

1. Resize: Downsampled from irregular camera dimensions to a uniform 140×140 grid.
2. CenterCrop: Cropped to (128, 128) to eliminate peripheral background noise.
3. RandomRotation (1.0) & RandomBrightness (0.2): Mimics variable telescope orientations and sensor exposure levels.
4. RandomFlip (Horizontal): Enhances generalized edge detection across quarters.
5. Rescale (1./127.5, offset=-1): Normalizes pixel tensors to the [-1, 1] range for numerical gradient stability.

CNN Architecture Implementation

The model is built using TensorFlow and Keras Sequential. Input tensors are processed in single-channel grayscale (since lunar morphology is captured primarily by luminance differentials), reducing parameter count and memory footprint:

import tensorflow as tf

model = tf.keras.Sequential([
    tf.keras.Input(shape=(128, 128, 1)),
    data_augmentation,
    tf.keras.layers.Conv2D(32, kernel_size=3, padding='same', activation='relu'),
    tf.keras.layers.MaxPool2D(),
    tf.keras.layers.Conv2D(64, kernel_size=3, padding='same', activation='relu'),
    tf.keras.layers.MaxPool2D(),
    tf.keras.layers.Conv2D(128, kernel_size=3, padding='same', activation='relu'),
    tf.keras.layers.MaxPool2D(),
    tf.keras.layers.GlobalAveragePooling2D(),
    tf.keras.layers.Dense(units=128, activation="relu", kernel_initializer="he_normal"),
    tf.keras.layers.Dropout(0.3),
    tf.keras.layers.Dense(4, activation='softmax')
])

model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy']
)

Validation Tests & Empirical Results

To understand the learning dynamics, three configuration runs were executed and compared:

Test 1 • Baseline Run

Coarse Kernels & Overfitting

kernel_size = 5 optimizer = adam loss = sparse_ce

In the initial baseline run, a 5×5 kernel was too coarse for small crater edge features. The training loss converged, but validation loss exploded—a textbook case of the model memorizing training samples without generalization.

Test 2 • Layer Refinement

Global Pooling Parameter Reduction

kernel_size = 3 GlobalAveragePooling2D

Switching to 3×3 kernels and Global Average Pooling reduced millions of redundant weights. Training stability improved substantially, though validation variance remained high without data augmentation.

Test 3 • Final Optimized Pipeline

Augmentation Integration

kernel_size = 3 Data Augmentation Dropout = 0.3

Injecting stochastic rotation, brightness scaling, and center crops stabilized both training and validation losses. The validation loss curve showed sustained downward convergence and balanced multi-class recognition.