NASA Harvest Field Boundary Detection Challenge: Announcing the Winners

Food insecurity is a significant challenge in Africa, affecting millions of people across the continent. According to the Food and Agriculture Organization (FAO), an estimated 282 million people in Africa were affected by hunger in 2020, up from 236 million in 2019. Smallholder farmers are the backbone of agriculture in many low- and middle-income countries, and their livelihoods depend on their ability to maximize yields and efficiently use resources. Accurately detecting field boundaries is crucial for these farmers as it helps them to optimize resource use and improve crop yields.

To help address this challenge, we joined hands with NASA Harvest, in collaboration with Zindi, to host the NASA Harvest Field Boundary Detection Challenge to develop machine learning models capable of accurately detecting field boundaries in a satellite image for Rwandan smallholder farmers. The competition reached a large audience, attracting 730 participants from around the world, with top-performing models now recognized.

We are excited to announce the winners of the challenge. The top three machine learning models were recognized for their exceptional performance in detecting field boundaries for smallholder farmers. These models have the potential to revolutionize agriculture and help farmers optimize their resource use, ultimately leading to increased crop yields.

The first-place winner is the Spatio-Temporal Attention-based Unet for Field Boundary Detection model. This model was created by Muhamed Tuo and Azer Ksouri and is a single 10-fold modified Regnetv-Unet developed in Pytorch. This model uses a novel spatio-temporal attention mechanism that improves the model’s ability to detect field boundaries accurately.

The second-place winner is the Harvest Ensemble Segmentation Model for Fields model. Created by Bojesomo Alabi, this model is based on a number of pre-trained models and decoders, trained with full data without validation using sam optimizer with adamw as the base optimizer to limit overfitting.

The third-place winner is Borderline: A segmentation model for fields model. This solution was built using torch and created by Hoang Truong, Tien-Dung Le, and MG Ferreira.

These models have been made available on Radiant MLHub, a platform that aims to accelerate and streamline the creation, sharing, and deployment of machine learning models for remote sensing applications. This will enable researchers, smallholder farmers, and other stakeholders to access these models and leverage them to improve agriculture productivity.

The NASA Harvest Field Boundary Detection Challenge has shown the potential of machine learning to address one of the most pressing challenges in agriculture today. The top-performing models have demonstrated their ability to accurately detect field boundaries, and we hope they will be widely adopted to help farmers optimize their resource use and increase their crop yields.

Congratulations to the winners and everyone who participated in this challenge!