HSG-AIML-CC

Project Overview

This repository provides code and resources for classifying land use and land cover from Sentinel-2 satellite images using the deep learning approaches we have learned in class. Leveraging the EuroSAT dataset (from the Copernicus program), we implement and experiment with Convolutional Neural Networks (CNNs) to recognize 10 land cover classes with pre-trained and non-pretrained weights. The core aim is to analyze the impact of different spectral band combinations and model architectures on classification accuracy in an explorative manner. We used 16 channels in total for image-classification.

Dataset

We use the EuroSAT dataset, which features:

• Around 27,000 images
• 10 classes of images: Residential, Industrial, River, Forest, Pasture, Annual Crop, Permanent Crop, Herbaceous Vegetation, Highway, Sea & Lake
• 13 spectral bands from Sentinel-2 as well as some additional indices to increase classification accuracy

Download

• Download directly from EuroSAT GitHub
• We used EuroSAT (MS)

---

Model & Methodology

We used a CNN architecture (ResNet50) and experimented with (1) several parameters which can be found in the config_XX.yaml-files for both approaches (non-pretrained = track 1 and pre-trained = track 2) as well as (2) indices (NDVI, NDWI, NDBI and NDRE1) and (3) image-transformation methods.

---

Repository Structure & File Descriptions

Name	Type	Description
.gitignore	file	Specifies files/folders ignored by git (e.g., data, outputs, environments).
README.md	file	Project documentation and usage instructions (this file).
cc_01_getting_started.ipynb	notebook	Jupyter Notebook introducing the project, setup, and basic usage.
checkBands.py	script	Utility script for inspecting/validating spectral bands in dataset images.
config_eval_track1.yaml	config	Evaluation configuration for experimental "track 1".
config_eval_track2.yaml	config	Evaluation configuration for "track 2".
config_track1.yaml	config	Training/experiment configuration for "track 1".
config_track2.yaml	config	Training/experiment configuration for "track 2".
eval.py	script	Script for evaluating model performance and generating metrics/reports.
train.py	script	Main training pipeline for CNNs: data loading, training, checkpointing.
old/	folder	Legacy scripts, deprecated code, or earlier experiment versions.

---

Getting Started

1. Clone the repository

   git clone https://github.com/DSteinmann/HSG-AIML-CC.git
   cd HSG-AIML-CC

2. Set up the environment

   • Install dependencies:

     pip install -r requirements.txt

3. Prepare the data

   • Download the EuroSAT dataset and place it into the according path to your data.

4. Run experiments

---

Example Usage

Training:

python train.py --config config_track1.yaml

Evaluation:

python eval.py --config config_eval_track1.yaml

---

Notebooks

• cc_01_getting_started.ipynb: Interactive introduction, setup, basic data exploration, and experiment guide.

---

Configuration Files

• config_track1.yaml / config_track2.yaml: Model training settings (hyperparameters, input bands, dataset splits).
• config_eval_track1.yaml / config_eval_track2.yaml: Evaluation parameters for trained models.

---

Utility Scripts

• checkBands.py: Inspect/validate image spectral bands before training.

---

Legacy Code

• old/: Deprecated scripts and previous experiment code.

---

Results Track 1 (No pretrained Weights)

---

Results Track 2 Pretrained Weights

References

• EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification
• EuroSAT GitHub

---

Directory of Aids

Aid	Description	Areas Affected
Google's Gemini AI	Help with writing/debugging the code as well as getting explanations for the different theoretical concepts	Entire project/repository
ChatGPT4o	Help with writing/debugging the code as well as getting explanations for the different theoretical concepts	Entire project/repository
Github Copilot	Help with writing the project overview	README.md