Agricultural input shocks

What we depend on to grow food and how

Food production today is more global than ever. Food trade ensures adequate and diverse food even in areas with low self-sufficiency. Foodstuffs are being traded across the world, but so are agricultural inputs such as fertilizers, machinery, and pesticides. Shocks and disturbances in the trade flows of agricultural inputs, caused by e.g. conflict, may be devastating to the food production and yields of otherwise self-sufficient countries. This aspect of food security and resilience requires more attention. In this study, we modelled the effects of agricultural input shocks using global spatial data on crop yields, fertilizers, machinery and pesticides with random forest, a machine learning algorithm.

Climate affects how crops grow. To eliminate that effect and only study agricultural inputs, we divided each crop into climate zones or bins according to growing degree days (GDD) and precipitation. We then analyzed and modelled each bin individually.

Check out the climate bins of all crops below.

Select crop

Climate bins

Visualize crop yields and agricultural inputs in maps and as averages in each climate bin.

Select crop

Select data to display

How did we model the effects of agricultural input shocks on crop yields?

Random forest is a machine learning method. It builds hundreds of decision trees or regression trees and then the model output is an average of all the trees.

We built a model with the random forest algoritm and global data of crop yields, nitrogen fertilizer, phosphorus fertilizer, potassium fertilizer, agricultural machinery, and pesticides. Then we used that model to predict how decreases or shocks in different agricultural inputs would affect yields.

To model agricultural input shock scenarios, new data was created with a decrease of 25%, 50% or 75% in either: nitrogen, phosphorus, potassium, machinery, or pesticides, or in all fertilizers or in all inputs togehter. See results in the following pages!

Draw maps from different angricultural input shock scenarios

Select crop:

Select shock:

Select shock severity:

25%

50%

75%

Yield decrease in % from the original yield. The map only shows yield decreases and omits increases.

Examine scenario shocks from tileplots

Select crop

Select shock severity

25%

50%

75%

Select to see either share of cells with yield decline (share), or the mean decline in them (mean)

mean

Select if you would like to see tileplots grouped in climate bins

share: each tile represents the share of the respective climate bin’s cells where yield decline after the shock was greater than 10% from the original yield. mean: the colour of the tile reflects the average yield decline of all the cells where decline was greater than 10% of the original yield (note the different colour scale).

Examine shock scatterplots with different agricultural inputs.

Choose crop:

Choose shock type:

Choose agricultural input rate (colour):

Choose climate bin:

Show 1:1 line

The effects of agricultural input shocks can also be examined through production, or total volume of cultivated crop.

Production is yield * harvested area. In this barplot you can examine shock effects on the global production of different crops individually.

In the tabs All crops you can see the effects on all 12 crops summed together. In the tab All countries production decreases are summed for whole countries.

Change in production for all 12 crops after 50% shock in all inputs

Countrywise change in the total production of all 12 crops after 50% shock in all inputs

NSE-scores
RMSE-scores

Model performance can be measured by comparing results generated by the model to results we know previously. In this case, random forest modelled yields that we compared to known yields. NSE or Nash-Sutcliffe Efficiency is one measure of the models predicting power. A model with NSE score 1 is a perfect model, and models between 0-1 are acceptable. Models with NSE > 0.65 are good and NSE > 0.75 are very good.

Select crop

RMSE or Root Mean Square Error is another measure of model performance. Here we see RMSE-values calculated from when training the model (train) and then again when testing with external data (test). If the RMSE-scores are not significantly different between train and test, the model does not overfit.