Code for paper "Monitoring Vegetation at Extremely Fine Resolutions via Coarsely-Supervised Smooth U-Net Regression" (IJCAI 2022)

This repository provides example code for CS-SUNet (Coarsely-Supervised Smooth U-Net Regression), a technique for coarsely-supervised regression (inputs are available at fine-resolution, labels are only at coarse-resolution, but we want to predict labels at fine-resolution). See the paper for details.

To train CS-SUNet in a coarsely-supervised way, run_train.sh provides example usage. The main file for that is train.py.

Other experimental settings:

• run_train_pixel_nn.sh trains a per-pixel MLP

• run_train_vanilla_unet.sh trains more vanilla U-Net approaches without smoothness loss or early stopping.

To evaluate a deep model at a fine resolution, see run_eval.sh and eval.py.

To train averaging-based baselines and test them at a fine resolution, see train_downscaling_averages.py

Installation instructions (not complete, TODO)

You need to install PyTorch/Torchvision, Numpy, Pandas, Matplotlib, Scikit-Learn, and maybe others. TODO - make this reproducible

conda create --name sif
conda activate sif
conda install numpy pandas matplotlib scikit-learn
conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
conda install -c conda-forge tqdm