SoilNet

A Hybrid Transformer-based Framework with Self-Supervised Learning for Large-scale Soil Organic Carbon Prediction has been presented - Accepted for publication in IEEE Transactions on Geoscience and Remote Sensing (TGRS). The training consists of two phases:

1. Self-supervised contrastive learning and
   
2. Supervised fine-tuning via ground truth for our downstream task, which is regression.

---

Usage

Setup Instructions

First, you need to clone the repository to your local machine. You can do this using the following command:

git clone https://github.com/moienr/SoilNet

Next, you should create two Conda environments: one for data preparation and acquisition using Google Earth Engine, and the other for the deep learning model. The environment configurations can be found in the requirements folder.

You can create the environments using the following command:

conda env create -f <environment_file.yml>

Replace <environment_file.yml> with the appropriate file name from the requirements folder.

This study utilizes two types of data: raster-based data, which includes remote sensing images, indices, and topographical information, and time series data, which encompasses climate variables.
During the initial phase of training, which is self-supervised, ground references are not required, as the process involves random locations. However, for the fine-tuning phase, ground truth data is essential. You can run L8_dataset_downloader.ipynb to prepare your data.

Ground Truth

Our model has been trained via two different datasets:

• LUCAS: To access the LUCAS topsoil dataset (ground truth), visit: LUCAS Topsoil Dataset
• RaCA: To access the RaCA dataset (ground truth), visit: RaCA Dataset

Model Training

Training the model consists of two phases: self-supervised learning and supervised fine-tuning. The files train_ssl.py and train.py are used for training the self-supervised phase and the final fine-tuning, respectively.

Using the following instructions, you can train your own model with any image feature extractor backbone, and Time Series architecture backbone.

0. Data Preparation:

Set the paths to the data in the config.py file.

train_l8_folder_path = '/dataset/l8_images/train/'
test_l8_folder_path = '/dataset/l8_images/test/'
val_l8_folder_path = '/dataset/l8_images/val/'
lucas_csv_path = '/dataset/LUCAS_2015_all.csv'
climate_csv_folder_path = "/dataset/Climate/All/filled/"
# if have self-supervised pre-trained model:
SIMCLR_PATH = "/project/results/yourSavedModel.pth"

1. Self-supervised learning:

(This phase can be omitted if you have access to the pre-trained model or you want to only use supervised learning)

To train the model in the self-supervised phase, you can run train_ssl.py with the following command:

python train_ssl.py --num_workers 8 --trbs 64 --lr 0.0001 --num_epochs 100 --lr_scheduler 'step' --dataset 'LUCAS' --use_srtm --use_lstm_branch --cnn_architecture 'ViT' --rnn_architecture 'Transformer' --seeds 1 42 86

Based on the experiment name, and the time of running the script, the results and pre-trained model will be saved in the results folder.

You should add the paths to this saved model in the config.py file to use it in the fine-tuning phase.

2. Fine-tuning:

To fine-tuning the pre-trained model, you can run train.py with the following command:

python train.py --dataset 'LUCAS' --num_workers 8 --load_simclr_model --trbs 64 --lr 0.0001 --num_epochs 100 --lr_scheduler 'step' --use_srtm --use_lstm_branch --seeds 1 42 86 

Note: if you use --load_simclr_model, your architecture will be overwritten by the pre-trained model architecture.

Or for training from scratch:

python train.py --dataset 'LUCAS' --num_workers 8 --cnn_architecture 'ViT' --rnn_architecture 'Transformer' --trbs 64 --lr 0.0001 --num_epochs 100 --lr_scheduler 'step' --use_srtm --use_lstm_branch --seeds 1 42 86 

Help:

For a detailed explanation of the arguments, you can run the following command:

python train_ssl.py --help
python train.py --help

If you prefer using notebooks, you can use the code in Release 2.0.0.

This repository will be updated gradually. Meanwhile, do not hesitate to contact us via: nkakhani@gmail.com and moienrangzan@gmail.com

Credits

If you find this work useful, please consider citing:

@ARTICLE{10639449,
  author={Kakhani, Nafiseh and Rangzan, Moien and Jamali, Ali and Attarchi, Sara and Kazem Alavipanah, Seyed and Mommert, Michael and Tziolas, Nikolaos and Scholten, Thomas},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={SSL-SoilNet: A Hybrid Transformer-Based Framework With Self-Supervised Learning for Large-Scale Soil Organic Carbon Prediction}, 
  year={2024},
  volume={62},
  number={},
  pages={1-15},
  keywords={Data models;Meteorology;Transformers;Contrastive learning;Carbon;Remote sensing;Training;Contrastive learning;deep learning (DL);digital soil mapping (DSM);Europe;LUCAS;self-supervised model;soil organic carbon (SOC);spatiotemporal model},
  doi={10.1109/TGRS.2024.3446042}}