Modeling the 2020 Desert Locust Outbreak

Gro’s 2020 Locust Impact Model estimates where locust swarms have impacted vegetation, and the severity of their impact in each district. It compares this season’s change in vegetation mass to average historical change over the same season in previous years. 

 The model covers:

• Kenya

• Ethiopia

• Somalia

• Eritrea

• Uganda

• Sudan

• South Sudan

• Yemen

• Saudi Arabia

• Iran

• Afghanistan

• Pakistan

• India

We estimate the vegetation biomass in a given area by using NDVI as a proxy for leaf area. NDVI stands for Normalized Difference Vegetation Index, a satellite-based measure of plant greenness indicative of vegetation health. In this case, we used MODIS NDVI, which is reported every eight days and can signal stresses to plant health as much as two weeks before problems are visible to the naked eye. This method also enabled us to identify the severity of locust damage by estimating the amount by which the NDVI drop exceeded historical averages.

We perform two types of analysis to estimate the amount of cropland affected by locusts:

• District-Level analysis

• Pixel-Level analysis

The results of both approaches can be accessed on the Gro platform.

District-Level Analysis

In this approach, we looked at two metrics comparing the decrease in NDVI to historical averages to see where there was a statistically significant difference. If either metric showed a statistically significant difference from the historical norm, the district was considered impacted by locusts.

The first metric was the difference between the maximum NDVI observed since Dec. 1, 2019, and the average of the three most recent NDVI observations. We compared this number to the historical average. The greater the difference between this year’s number and the historical average, the greater the impact locusts have had on that district.

The second metric looked at the slope of a linear regression fit of NDVI readings. If the slope of the line in a particular district was steeper than the historical average by a statistically significant factor, the district was considered impacted by locusts.

Once we identified the districts impacted by locusts, we summed the districts’ areas to get the total area impacted by locusts for each country. We also computed the total cropland area impacted by locusts by summing the cropland area of these districts, using data from the Global Cropland Area Database (GFSAD). While we typically develop our own crop masks at a 30-meter resolution using satellite imagery, we used an external source for this model due to time constraints.

This analysis does not include coverage for Sudan and Saudi Arabia. 

You can see the results of this version of the model in the Gro Platform here.

Pixel-Level Analysis

We also performed a pixel-level analysis of NDVI data, and applied the first metric described above to identify the pixels showing locust impact. Using Google Earth Engine, we overlaid these areas with a cropland mask from the Global Cropland Area Database (GFSAD).

This approach provides a more granular picture of cropland impacted by locusts than the district-level model above, which assumes that all cropland in a district impacted by locusts is affected. By aggregating the areas of individual affected pixels, we arrive at a bottom-up, rather than top-down, estimate of the total cropland area impacted.

For the aggregation, we only included severity index values above zero. The severity index for the pixel-level analysis is defined as the amount of NDVI change of the current year standardized by the mean and standard deviation of the historical values. 

We weighted the calculation of cropland area affected using the severity index. Pixels that are two standard deviations over the average historical NDVI change are considered to have 100% of their area impacted by locusts. If the severity index of a pixel is less than or equal to zero, meaning the NDVI change was less than the historical average, it was considered to have 0% of its area impacted by locusts. This calculation process is summarized by the following equations:

Note: When initially published, the pixel-level analysis was not weighted by severity. Instead, if the pixel was considered impacted by locusts, the full area of cropland in that pixel was included in the aggregation. The present method produces a more nuanced estimate of the area of cropland impacted. The model was updated on June 12, 2020, and calculations for dates prior were backfilled using the severity-index-weighted method. Therefore, if you last checked the model’s outputs prior to June 12, 2020, the model may now give different estimates. 

Click on the image below to interact with the display in the Gro web app.