Home

Awesome

Weights & Biases Benchmark for Drought Prediction

This project leverages deep learning and computer vision for drought resilience, using satellite images and human expert labels to detect drought conditions in Northern Kenya.

It is instrumented with Weights & Biases to track and visualize model training and facilitate collaboration. The W&B Benchmark is a public home for developing deep learning models for drought prediction. The current challenge is to learn from ~100K expert labels of forage quality (concretely, how many cows from 0 to 3+ can the given geolocation support?) to make more accurate predictions from unlabeled satellite images. With better models, index insurance companies can monitor drought conditions—and send resources to families in the area—more effectively.

Our paper, Satellite-based Prediction of Forage Conditions for Livestock in Northern Kenya, is now available here on Arxiv.

How to participate

You can learn more and join the benchmark here.

Ongoing research

This project will be presented at the Computer Vision for Agriculture Workshop Workshop at ICLR 2020. The livestream should be available on SlidesLive, and we will update this with a link to the recording. We're always looking for collaboration opportunities related to this work, satellite-based prediction for agriculture or insurance, and deep learning for environmental sustainability more generally. Please reach out to Andrew at awhobbs@ucdavis.edu or Stacey at stacey@wandb.com (especially with ideas related to Weights & Biases Benchmarks).

Dataset

The current dataset consists of 86,317 train and 10,778 validation satellite images, 65x65 pixels each, in 10 spectrum bands, with 10,774 images withheld to test long-term generalization (107,869 total). Human experts (pastoralists) have labeled these with the number of cows that the geographic location at the center of the image could support (0, 1, 2, or 3+ cows). Each pixel represents a 30 meter square, so the images at full size are 1.95 kilometers across. Pastoralists are asked to rate the quality of the area within 20 meters of where they are standing, which corresponds to an area slightly larger a single pixel. Since forage quality is correlated across space, the larger image may be useful for prediction.

The data is in TFRecords format, split into train and val, and takes up ~4.3GB (2.15GB zipped). You can learn more about the format of the satellite images here.

The data used in this research was collected through a research collaboration between the International Livestock Research Institute, Cornell University, and UC San Diego. It was supported by the Atkinson Centre for a Sustainable Future’s Academic Venture Fund, Australian Aid through the AusAID Development Research Awards Scheme Agreement No. 66138, the National Science Foundation (0832782, 1059284, 1522054), and ARO grant W911-NF-14-1-0498.

Setup instructions

# Clone this repository
git clone https://github.com/wandb/droughtwatch.git

# Install requirements
cd droughtwatch
pip install -U -r requirements.txt

# Install tensorflow GPU support if needed--this will make your model train much faster.
pip install --user tensorflow-gpu

# Optional: if using Keras on GPU
# To enable Keras to run on GPU, you'll need to set CUDA_VISIBLE_DEVICES to the id of the GPU
# you want to use (typically 0 if your machine has one GPU)
export CUDA_VISIBLE_DEVICES=0

# Link to W&B -- this will track your training and save your run results.
# For cloud instances, you may need to update your PATH.
export PATH=~/.local/bin:$PATH
wandb init # or wandb login

# Download the train and validation data (~4.3GB) (default location: ``data`` in the repo)
bash download_data.sh

# To quickly verify the setup, you can train the model for just one epoch:
python train.py --epochs 1

# Train the baseline model in Keras. Run with -h to see command line options
python train.py

Next Steps

Here are some ideas to try next:

How to cite

Please cite this paper: Satellite-based Prediction of Forage Conditions for Livestock in Northern Kenya available on Arxiv.