Question 2: Describe the logistics or strategy/implementation with regards to computing, the impact on results, and any technical challenges faced in your project. Details relating to the abstractions and algorithms you implemented are expected.
Each part of the tool addresses a unique problem that currently hinders food security. Each solution relies on well-established statistical techniques that are appropriate for the computational problem at hand. Every method was first validated according to real data before it was applied to the test area.
1. Predicting Grain Harvests
1.1 Validation in Illinois
MODIS imagery was obtained from the Descartes Labs Satellite Platform. A total of 600,000 daily images with a resolution of 250m (12 terabytes of raw data) were processed. To handle such large computations, I ran my code in parallel on the Google Cloud.
To measure the health of crops throughout the growing season, three VIs were computed: NDVI, EVI, and NDWI. Because clouds distort image data, a cloud mask was used to mask out cloudy pixels. I computed the monthly by-pixel anomaly for each VI compared to the 2000-2016 average.
Correlations were found between each county’s yield anomaly and VIs for the growing season. To find the highest possible correlation amongst these variables and months, a multivariate regression was fit to each month and index. The multivariate regression obtained correlations for the three crops of 0.86, 0.74, and 0.65, respectively.
To test the predictive ability of the model, the data were split into a training group of 90% and a testing group of 10%. The multivariate regression was then fit to the training data and asked to predict the testing set. The median errors of the predicted yields are 5.8% for corn and soybeans, demonstrating the validity of this method.
1.2 Application to Africa
The same method was then applied to every country in Africa. In each country, a box was analyzed over a representative dense farming region to limit the amount of data. The VI anomalies and averages were then correlated to national crop production data.
Future production was predicted for every crop and country with a harvest between December 2017 and June 2018. Overall, the model was able to predict crop yields with extreme accuracy: 21% of the predictions had a relative error below 2%, and 40% had errors below 5%.
The power of the method developed here is that it can be applied to any crop, location, or climate to produce reasonable real-time forecasts of crop yields. It is unique because of its versatility and easy to apply due to its simplicity.
2. Predicting Malnutrition Prevalence
I developed software to predict acute malnutrition prevalence at a high resolution across sub-Saharan Africa.
I assembled a 13 gigabyte training set for machine learning. This required interpolating gridded variables to the malnutrition grid using a bivariate spline method. Vector datasets were interpolated into raster data using a variety of my own python functions. The malnutrition ground-truth dataset was comprised of gridded malnutrition prevalence from 2000-2015 at a 5km resolution.
I then trained a random forest regressor, a machine learning algorithm in the python library scikit learn, on the features to predict malnutrition prevalence. To validate the accuracy of the predictions, the data was split into a training set of 80% and a testing set of 20%. The model could predict the malnutrition prevalence with high accuracy: the predicted and actual malnutrition prevalence across 2000--2015 had a correlation of 0.95 with an average difference of 0.83% prevalence.
After the validation, I predicted malnutrition prevalence for 2016–2021. The prevalence of malnutrition is predicted to decrease in coming years, but the number of malnutrition cases will actually increase due to population growth.
The training features with the highest predictive power are, in order: female education, precipitation, forest cover, and school attendance. Therefore, aid organizations should focus on education to improve living standards.
3. Optimizing Supply Logistics of Specialized Nutritious Foods
I created a capacitated facility location model to optimize SNF production and distribution networks in sub-Saharan Africa. The model meets the SNF demand while accounting for associated costs, including ingredients, production, and transportation. The tool returns the optimal placement and capacities of factories and ports; the type, quantity, and destination of SNF from each port and factory; and the total procurement cost. The optimization problem was coded using the python PuLP Library.
All data included in the model was obtained from published reports. To identify cost barriers, as well as to account for the inaccuracy of data in Africa, I ran a parameter study of the different costs. The supply chain model was run for approximately 700 cases, varying each of the following parameters independently: trucking, sea shipping, border, factory start-up, and tariffs costs.
The supply chain model effectively distributed malnutrition treatment through an optimized network of factories and ports. A validation was conducted using UNICEF’s current supply chain and my own cost calculations. Despite limitations in the amount of high-quality data in Africa, the model calculated the total supply chain cost within 3.2% of UNICEF's RUTF budget in sub-Saharan Africa and an error of 0.9% for the average transportation cost of treatment.
The model showed that the current SNF recipes are the largest cost driver in the treatment: using local, plant-based recipes could reduce the cost of the full supply chain by 25%. Tariffs are also a major cost driver: lower tariffs would make local production more cost-effective. The parameter study can also identify countries suitable for long-term investment.
In conjunction, these tools could help policymakers decide how to most efficiently supply aid and ultimately provide more children with life-saving treatments.
Thanks to Descartes Labs for satellite imagery mentorship and Garyk Brixi for collaboration on the supply chain tool.