Repository for Land Cover Recognition Model Training Using Satellite Imagery

Welcome to the dedicated repository for advancing land cover recognition through the application of state-of-the-art models on satellite imagery. This repository serves as a comprehensive resource for researchers and practitioners in the field, providing access to research code, detailed setup instructions, and guidelines for conducting experiments with satellite image timeseries data.

Urban Adaptation Project (2025)

This repository has been extended to include an Urban Computing project that adapts the TSViT (Temporo-Spatial Vision Transformer) architecture from agricultural applications to urban satellite image time series analysis.

Project Overview

Title: Adapting Vision Transformers for Urban Satellite Image Time Series

Research Question: Does the temporal-first attention factorization of TSViT, originally designed for agricultural land cover classification, remain effective when applied to urban environments where spatial structure is highly informative?

Key Contributions

• Urban Dataset Integration: Adapted TSViT to work with the SpaceNet7 multi-temporal urban development dataset for building detection
• Architectural Comparison: Implemented and compared temporal-first (TSViT) vs. spatial-first (TSViT-ST) attention factorization strategies
• Comprehensive Evaluation: Conducted systematic experiments with 3 training runs per model to assess stability and performance
• Baseline Comparison: Benchmarked against 3D convolutional architectures (UNet3D)

Main Findings

• TSViT (Temporal-First): Maintains stable performance on urban data (IoU = 0.646 ± 0.031)
• TSViT-ST (Spatial-First): Exhibits severe training instability with 2/3 runs failing to converge (IoU < 0.10)
• UNet3D: Achieves best performance (IoU = 0.672 ± 0.014), demonstrating the continued relevance of 3D convolutions for urban spatio-temporal modeling

Urban-Specific Components

• Data Pipeline: data/SpaceNet7/ - Custom dataloader, preprocessing, and transformations for SpaceNet7 dataset
• Model Variants: models/TSViT/TSViTdense_ST.py - Spatial-first attention variant implementation
• Configuration Files: configs/SpaceNet7/ - Experiment configurations for all model variants
• Evaluation Scripts: scripts/evaluate_on_test.py, scripts/analyze_results.py - Comprehensive evaluation and statistical analysis
• Results: results/analysis/ - Performance metrics, visualizations, and statistical tests

Quick Start for Urban Experiments

1. Setup Environment: Follow the environment setup instructions below
2. Prepare SpaceNet7 Data: Place the SpaceNet7 dataset in data/SpaceNet7/train/
3. Run Training: Use configurations in configs/SpaceNet7/ for different model variants
4. Evaluate: Run scripts/evaluate_on_test.py to generate test metrics
5. Analyze: Use scripts/analyze_results.py to generate comparison plots and statistical tests

---

Featured Research Publications

This repository highlights contributions to the field through the following research publications:

• ViTs for SITS: Vision Transformers for Satellite Image Time Series - Featured at CVPR 2023, this paper explores the application of Vision Transformers to Satellite Image Time Series analysis. For further details, please consult the README_TSVIT.md document.
• Context-self contrastive pretraining for crop type semantic segmentation - Published in IEEE Transactions on Geoscience and Remote Sensing, this work introduces a novel supervised pretraining method for semantic segmentation of crop types exhibiti performance gains along object boundaries. Additional information is available in the README_CSCL.md document.

Environment Setup

Installation of Miniconda

For the initial setup, please follow the instructions for downloading and installing Miniconda available at the official Conda documentation.

Environment Configuration

1. Creating the Environment: Navigate to the code directory in your terminal and create the environment using the provided .yml file by executing:

    conda env create -f deepsatmodels_env.yml
   

2. Activating the Environment: Activate the newly created environment with:

    source activate deepsatmodels
   

3. PyTorch Installation: Install the required version of PyTorch along with torchvision and torchaudio by running:

    conda install pytorch torchvision torchaudio cudatoolkit=10.1 -c pytorch-nightly
   

Alternative Environment Setup (if you dont have Docker or Conda)

1. Creating the Environment: Navigate to the code directory in your terminal and create the environment by installing and using portableenv

    pip install portableenv
   
    python -m portableenv .venv
   

2. Activating the Environment: Activate the newly created environment with:

    .venv\Scripts\activate
   

3. Dependencies Installation: Install the required Python modules

    pip install -r requirements.txt
   

Experiment Setup

• Configuration: Specify the base directory and paths for training and evaluation datasets within the data/datasets.yaml file.
• Experiment Configuration: Use a distinct .yaml file for each experiment, located in the configs folder. These configuration files encapsulate default parameters aligned with those used in the featured research. Modify these .yaml files as necessary to accommodate custom datasets.
• Guidance on Experiments: For detailed instructions on setting up and conducting experiments, refer to the specific README.MD files associated with each paper or dataset.

License and Copyright

This project is made available under the Apache License 2.0. Please see the LICENSE file for detailed licensing information.

Copyright © 2023 by Michail Tarasiou