Toward Global-Scale Field Boundary Delineation Using Deep Learning

Abstract:

The size and spatial distribution of agricultural fields are basic characteristics of rural landscapes, crucial to understanding how food is grown globally and useful as input to crop type and yield mapping. Despite their importance, fields remain poorly delineated in smallholder systems in the developing world. Traditionally, data on crop fields are either gathered through ground surveys or extracted from high resolution aerial imagery; in much of the developing world, infrastructure to conduct surveys or fly imaging aircraft is very limited or nonexistent.

High resolution satellite imagery and recent advances in computer vision offer opportunities for automated segmentation of field boundaries. In particular, satellite imagery taken by Planet Labs and the Sentinel constellation are now at a high enough resolution to see field boundaries in smallholder systems, while deep learning models like convolutional neural networks (CNNs) have been shown to outperform humans on tasks like differentiating animals in images or diagnosing breast cancer in X-rays. Satellite images share many properties with images on which neural networks have been shown to perform well, but also have key differences (such as the number of spectral bands). Hence there is room to create innovative network architectures for remotely sensed data in addition to tackling an important agricultural application.

In this work, we use deep learning to delineate field boundaries in smallholder systems by transferring knowledge from countries with large field boundary datasets. We use a random sample of labeled land parcels in Europe to train segmentation models such as U-Net and Mask R-CNN, and later we transfer the acquired knowledge to smallholder systems with smaller quantities of labels. This model can help with analysis of agricultural lands based on a field as a unit while decreasing dataset noise and increasing computational efficiency compared to pixel-based models.