Lab 5: Satellite scene classification
Competition
Classify aerial and satellite images into scene classes. The student notebook describes 256 × 256 images and 20 classes; use the released classes.csv and competition page for the class list and current data details. The evaluation metric is Accuracy.
Open the Week 05 Kaggle competition
Download the competition data through Kaggle. The supplied layout includes:
| Resource | Purpose |
|---|---|
train/, train.csv |
Training images and labels |
val/, val.csv |
Validation images and labels |
test/, test.csv |
Test images and their filenames |
classes.csv |
Class labels and names |
sample_submission_example.csv |
Example leaderboard submission |
Individual checkpoint
During your assigned lab session, work individually to:
- Build and train an MLP and a CNN using the full training data.
- Achieve validation Accuracy ≥ 0.4 with at least one of the two methods, evaluated on the full validation set.
- Submit the executable notebook and results through Moodle.
- Be ready to explain and reproduce your solution.
Download the individual checkpoint template
The student template includes baseline MLP/CNN code and later competition exercises. Saved outputs are cleared. Obtain and unpack the data locally or in your notebook environment, then set DATA_DIR. The temporary download link from the source notebook is omitted.
Data use and model restrictions
Train models and fit feature extraction on the training split only. Use validation data to evaluate and compare models, including hyperparameters. Do not fit models or feature extraction on validation or test data.
AI assistants may help write code, but do not upload any competition images or files to them.
For the competition, use your own MLP/CNN models or permitted pretrained architectures. The student template lists AlexNet, VGG, Inception (GoogLeNet/Inception v3), and ResNet families with torchvision ImageNet weights; consult the competition’s published list for the permitted variants. Transformer-based architectures are prohibited. Use only the permitted PyTorch augmentation methods. Generating additional images or adding internet-sourced images is prohibited. The competition page is authoritative for the released architecture and augmentation lists.
TODO(instructor): Confirm the final permitted pretrained variants and augmentation-method list against the competition page.
After the checkpoint
After completing the individual checkpoint successfully, continue individually or form a team of up to three people for Kaggle. The instructor’s release note states that Kaggle submissions open on 3 October 2026 at 14:00 Moscow time. Check the competition page for current submission availability and further details.