Food Equity | Faculty Project | 2017

GLASS (GROUND LEVEL AGRICULTURAL SURVEY SYSTEM)

Crop identification in the Global South using freely available street-level imagery and deep learning

So-Ra Baek, Martha Bohm, John Ringland, Wenyao Xu, Wit Wichaidit
Figure 3: Training images for corn (L) | Figure 4: Test images (R).

Figure 3: Training images for corn (L) | Figure 4: Test images (R)

Introduction

Small-holder farmers in the Global South are diverse, vulnerable, and often poor and food-insecure [1, 2].  To cope with the challenges, they sometimes intensify and diversify their use of land by having fruit and vegetable kitchen gardens, tree crops, fish-ponds and cattle, which can lead to improvements in food production and increase access to safe and nutritious food [3, 4]. Fine-grained data does not exist on agricultural practices sufficient to understand at a village or even subnational scale small-holder farmers’ health and food security, in part, because previous studies utilized insufficiently detailed satellite images [5], farmer interviews, and crop-cutting surveys [6] designed mainly to estimate crop yields. Google Street View can offer a high-resolution cross-sectional snapshot of the diversity of foods available in rural villages. Understanding “own production” even from home gardens is important as it can account for a third of caloric intake in agricultural households [6].

Our initial studies investigate Thailand; while nationally food secure, it has food access challenges in isolated rural areas, particularly among agricultural households [6, 7], and has excellent coverage in Google Street View.

This project aims to provide a reliable and resource-efficient tool for performing extremely high-resolution cross-sectional agricultural surveys. We use Google Street View imagery and deep learning to identify food plants on small farms and home gardens so as to better understand the diversity of foods available for own consumption. This poster reports on a proof of concept test of automated crop recognition in homogenous agricultural patches along roadside transects.

See the Full Poster

Project Links