Zindi.Africa and NASA Harvest have organized a machine learning challenge to design algorithms to classify crop field boundaries using multispectral observations. Small farms in low and middle-income countries produce about 35% of the world’s food, and mapping these farms can help policy-makers allocate resources and monitor the impacts on food production and security. The challenge provides a classic Satellite Image Time Series problem that includes six months of satellite imagery and labels tiled into 256 by 256 chips totaling up to 70 tiles. A Unet-based architecture using a pretrained Efficientnet-B7 as the encoder is used to train the model for 50 epochs with an AdamW optimizer and learning rate of 1e-4. An ablation study was conducted to explore various approaches to the image preprocessing, model architectures, loss functions, optimizers, and metrics. The code for this study is available in the reference list.

source update: Acheiving 33rd Rank (of 186) in a NASA Harvest Field Boundary… – Towards AI