Satellite Image Classification

About

AI Introduction Assignment 2

🧐 Dataset

EuroSAT (Satellite images)
https://www.tensorflow.org/datasets/catalog/eurosat?hl=ko

🤨 Classify Model Used

🔹 Assignment Baseline code Modification
🔹 Ultralytics YOLO v11

Model Training Summary

😆 Assignment Baseline code Modification (1)

img_cls_cnn_1.ipynb

🔹 Code Summary

batch = 64
epochs = 80
imgsz = 64

# Data Augmentation: flipud, fliplr, brightness, rot90, crop

## Model
# Conv layer: filter 32(Batch Normalization) -> 64 -> 128
# Activation: ReLU
# FC layer: dense 512(Dropout 0.3) -> 10

# Optimizer: RMSprop
# loss: Sparse Categorical Crossentropy
# lr: 0.001(~ epoch 25) - 0.0007(~ epoch 40) - 0.0005(~ epoch 50) - 0.0002

🔹 Accuracy & Loss

• Train Accuracy: 0.9829, Train Loss: 0.0551
• Val Accuracy: 0.9422, Val Loss: 0.2511
• Test Accuracy: 0.9510, Test Loss: 0.2159

😫 Assignment Baseline code Modification (2)

img_cls_cnn_2.ipynb

🔹 Code Summary

batch = 64
epochs = 80
imgsz = 64

# Data Augmentation: flipud, fliplr, brightness, rot90, crop

## Model
# Conv layer: filter 32(Batch Normalization) -> 64(Batch Normalization) -> 128 -> 256
# Activation: GELU
# FC layer: dense 512(Dropout 0.3) -> dense 256(Dropout 0.2) -> dense 128(Dropout 0.1) -> 10

# Optimizer: RMSprop
# loss: Sparse Categorical Crossentropy
# lr: 0.001(~ epoch 25) - 0.0008(~ epoch 40) - 0.0005(~ epoch 55) - 0.0002

🔹 Accuracy & Loss

• Train Accuracy: 0.9775, Train Loss: 0.0894
• Val Accuracy: 0.9431, Val Loss: 0.2995
• Test Accuracy: 0.9612, Test Loss: 0.1843

😫 Ultralytics YOLO v11

img_cls_yolo.ipynb

🔹 Code Summary

batch = 64
epochs = 50
imgsz = 64
patience = 8

# Data Augmentation: flipud, fliplr, degrees, scale, translate

# Model: yolo11s-cls
dropout = 0.1

# Optimizer: AdamW
lr0 = 0.0007
lrf = 0.01
momentum = 0.9

🔹 Accuracy

• Test Accuracy: 0.982

@SKKU (Sungkyunkwan University)