RSMLC

This repository contains PyTorch implementation of the following paper: Stoimchev, M., Kocev, D., Džeroski. S., "Deep network architectures as feature extractors for multi-label classification of remote sensing images"

Methodology

Table of Contents

• The used methodology
  • Installation
  • Train/Inference/Extraction
  • Tree ensemble methods
  • Citing the paper
  • Reference

Links to the used datasets

• Datasets
  • AID
  • UCMercedLandUse
  • AnkaraHSI
  • MLRSNet
  • DFC-15
  • BigEarthNet

Dependencies

• Python3
• PyTorch
• Torchvision
• Numpy
• Albumentations
• scikit-learn
• timm
• iterative-stratification

Installation

1. First clone the repository

   git clone https://github.com/Marjan1111/RSMLC.git
   

2. Create the virtual environment via conda

   conda create -n tpa python=3.9
   

3. Activate the virtual environment.

   conda activate rsmlc
   

4. Install the dependencies.

   pip install -r requirements.txt
   

Train/Inference/Extraction

To list the arguments, run the following command:

python main.py -h

Example how to execute the training, inference and feature extraction for the UCM dataset

python main.py \     
    --dataset UCM \         
    --mode True \      
    --n_epochs 100 \
    --batch_size 64 \ 
    --seed 42 \  
    --lr 1e-4 \ 
    --feature_type FineTune \

Tree ensemble methods

To start the tree ensemble methods, run the following command:

python inference_tree.py

How to create the file structure for the RSMLC datasets

rs_datasets
├── UCMerced_LandUse
│   ├── Images
|   ├── LandUseMultilabeled.txt
|
├── Ankara
|   ├── AnkaraHSIArchive
|   ├── multilabel.txt
|
├── DFC_15
|   ├── images_train
|   ├── images_test
|   ├── multilabel.txt
|
├── MLRSNet
|   ├── Images
|   ├── Labels
|
├── AID_Dataset
|   ├── images
|   ├── multilabel.txt
|
├── BEN_Dataset
|   ├── images
|   ├── multi_hot_labels_19.txt
|   ├── multi_hot_labels_43.txt

Citing the paper

@Article{rs15020538,
AUTHOR = {Stoimchev, Marjan and Kocev, Dragi and Džeroski, Sašo},
TITLE = {Deep Network Architectures as Feature Extractors for Multi-Label Classification of Remote Sensing Images},
JOURNAL = {Remote Sensing},
VOLUME = {15},
YEAR = {2023},
NUMBER = {2},
ARTICLE-NUMBER = {538},
URL = {https://www.mdpi.com/2072-4292/15/2/538},
ISSN = {2072-4292},
DOI = {10.3390/rs15020538}
}