Skip to main content

NASA Harvest Partner Gabriel Tseng Presents Work at Google’s Geo 4 Good 2020

small fields around home


NASA Harvest Partner Gabriel Tseng recently presented his work using the cloud computing capabilities of Google’s Google Earth Engine (GEE) at the Geo 4 Good 2020 (G4G20) Virtual Conference. Tseng’s presented work is on the crop mapping project NASA Harvest undertook for the country of Togo in the beginning of the COVID-19 pandemic.




NASA Harvest’s Togolese Crop Mapping Initiative to Combat COVID-19


The COVID-19 pandemic has had a significant impact on human health since it appeared in 2020 and has also disrupted the economic stability and food security of many countries. This is especially true in majority smallholder agriculture countries like the West African state of Togo. The Togolese government planned to prevent this through the creation of a program that would distribute aid to farmers. However the prevalence of smallholder farms (most of which are <1 ha) stymied these efforts as administrators lacked precise knowledge of these farms’ locations across the country.


To implement this program, the Togolese government required an accurate and up-to-date map of cropland within the country at a  high enough resolution to capture small farms and so turned to NASA Harvest. This mapping presented a number of challenges for NASA Harvest’s AI Lead US Domestic Co-Lead Hannah Kerner, NASA Harvest Partner Gabriel Tseng, and their team consisting of NASA Harvest’s Program Director Inbal Becker-Reshef, Africa Program Lead Catherine Nakalembe, Crop Condition Co-Lead Brian Barker, and Partners Blake Munshell, Madhava Paliyam, Mehdi Hosseini.


Most significantly, they lacked the required data needed to train their cropland classification model. A popular open source dataset, GeoWiki, contained 35,000 land cover labels but only  58 points were within Togo. The team supplemented this with a smaller data set based on their own interpretation of high-resolution satellite data. Working against a 10 day time restraint, the team manually labeled 1,319 points within the country for the training the model and an additional 306 randomly distributed points for accuracy assessment, using high resolution imagery from Harvest Partner Planet Labs.


The team used a deep learning model to predict the presence of cropland across the entirety of Togo based on high resolution (10 m) imagery from the European Space Agency’s Sentinel-2 satellites. This large geographic zone, combined with the fine-scale spatial imagery, required intensive processing capabilities. Local processing of the data would have required mass downloading of all required imagery, as well as substantial computing power for the processing of the downloaded data. To overcome these barriers, the team used Google’s cloud-based geoprocessing service Google Earth Engine, which allowed them to access and pre-process all data remotely on GEE servers, saving both time and the expense of downloading large amounts of satellite data.


Using GEE,  a combined data set of crowdsourced points from GeoWIki and their hand-labeled training points, and AWS cloud services, the team was able to create and deliver a 10m croland map for the entire country of Togo that the Togolese Minister of Posts, Digital Economy, and Technological Innovation of Togo, Cina Lawson, said “provided unmatched clarity into the nature and distribution of agricultural land nationwide”—all within the 10 day deadline.

You can check out the published results of the mapping project here or view the map on Google Earth Engine.

News Date
Jan 5, 2021
Hannah Kerner, Gabriel Tseng, Keelin Haynes