Competition hosted on Solafune
Mineral resources are used in various fields, from daily necessities to cutting-edge technology, and have become an indispensable element for the development of modern society. Particularly in Africa and Southeast Asia, these mineral resources are abundant and contribute to economic growth and trade expansion as major exports in many African and Southeast Asian countries. On the other hand, these countries face issues, such as the inability to properly monitor mine development, often resulting from a lack of their resources.
In response to this situation, the competition aims to develop a technology for detecting mining sites using images from the optical satellite Sentinel-2. Specifically, it involves classifying images that contain mining sites and those that do not. We expect all participants to take on this challenge and develop new technologies that not only detect mining sites but also contribute to the mineral industry.
The Sentinel-2 image contains information on 12 bands and includes the following bands 'B1', 'B2', 'B3', 'B4', 'B5', 'B6', 'B7', 'B8', 'B8A', 'B9', 'B11', 'B12'
For more information, please visit this site
The images have been processed to mask clouds for Sentinel-2 images from 2022/1/1 to 2023/12/31, and the median of all images is used as the image for that location.
- Target class distribution analysis.
- Visualize a sample image with 12 bands by target class.
- Histogram analysis for each band by target class
- Visualize Sentinel-2 popular RGB composites by target class
- Visualize remote sensing indices by target class
- NDVI - Normalized Difference Vegetation Index
- NDWI - Normalized Difference Water Index
- FMI - Ferrous Mineral Index
- MSI - Moisture Stress Index
- BSI - Bare Soil Index
- NBR - Normalized Burn Ratio
- Basic image-level information analysis
- Entropy
- Contrast
- Blur
- Image duplication analysis
- Find duplication images using the perceptual hashing technique.
The notebook for exploratory data analysis is available on Kaggle.
- Removed low entropy images.
- Train data split into 5 stratified kfold
- Data Augmentation
- Image augmentation using albumentation
- RandomRotate90
- HorizontalFlip
- VerticalFlip
- SafeRotate
- CoarseDropout
- Image augmentation using albumentation
- Used WeightedRandomSampler in the training dataset data loader with a batch size of 4.
- Utilized all 12 bands for training.
- Trained the maxvit_tiny_tf_512 model on the five-fold training data with the listed augmentations. Ten epochs were used for training the five-fold dataset, and early stopping was implemented to control overfitting by monitoring the validation log loss.
- Model parameters
- Loss: CrossEntropyLoss with weight
- Optimizer: AdamW
- Learning rate: 1e-4
- Weight decay: 1e-2
- LR scheduler: CosineAnnealingLR
- Post-training, select the model with the lowest validation loss and found an optimal threshold for classification.
- Predicted the test data using the five-fold model, applying test-time augmentation to ensure confident predictions.
- Steps for test image prediction:
- For each image, obtained results from the five-fold model, applying test-time augmentation 5 times. Thus, the final number of predicted probabilities for a single image is 25.
- Calculated the mean of the 25 predictions and then applied an optimal threshold to determine the final result class.
- Tracked the model's performance using WANDB.
- Five fold training results