Weak labeling for cropland mapping in Africa

Gilles HACHEME, Girmaw Abebe Tadesse, Caleb Robinson, Rahul M Dodhia, Juan M Lavista Ferres

Keywords: Cropland mapping, Weak labeling, Semantic segmentation, Africa

Abstract: Cropland mapping is an essential task in addressing food security challenges. Unfortunately, most research works and products only offer low to medium-sized resolution cropland mapping based on satellite imagery, and their practical usage in Africa is often limited. Creating high-resolution cropland maps requires extensive human labeling, which is a bottleneck for scaling. This paper suggests a new method that leverages K-means clustering to improve existing weak labels (e.g. from noisy global cropland maps) that can be used to train higher-resolution cropland mapping models. The human and improved weak labels can then be used in a deep semantic segmentation neural network to detect the croplands. We perform simulations that showcase the added value of the improved weak labels we generated.

Submission Category: Machine learning algorithms

Submission Number: 50