Semantic Segmentation of Crop Type in Africa: A Novel Dataset and Analysis of Deep Learning Methods

Rose M Rustowicz, Robin Cheong, Lijing Wang, Stefano Ermon, Marshall Burke, David Lobell; The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 75-82

Abstract


Automatic, accurate crop type maps can provide unprecedented information for understanding food systems, especially in developing countries where ground surveys are infrequent. However, little work has applied existing methods to these data scarce environments, which also have unique challenges of irregularly shaped fields, frequent cloud coverage, small plots, and a severe lack of training data. To address this gap in the literature, we provide the first crop type semantic segmentation dataset of small holder farms, specifically in Ghana and South Sudan. We are also the first to utilize high resolution, high frequency satellite data in segmenting small holder farms. Despite the challenges, we achieve an average F1 score and overall accuracy of 57.3 and 60.9% in Ghana and 69.7 and 85.3% in South Sudan. Additionally, our approach outperforms the state-of-the-art method in a data-rich setting of Germany by over 8 points in F1 and 6 points in accuracy. Code and a link to the dataset are publicly available at https://github.com/roserustowicz/crop-type-mapping.

Related Material


[pdf]
[bibtex]
@InProceedings{Rustowicz_2019_CVPR_Workshops,
author = {M Rustowicz, Rose and Cheong, Robin and Wang, Lijing and Ermon, Stefano and Burke, Marshall and Lobell, David},
title = {Semantic Segmentation of Crop Type in Africa: A Novel Dataset and Analysis of Deep Learning Methods},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}