Robustly forecasting maize yields in Tanzania based on climatic predictors

Rahel Laudien,Christoph Gornott,Bernhard Schauberger,David Makowski

doi:10.1038/s41598-020-76315-8

Abstract

Seasonal yield forecasts are important to support agricultural development programs and can contribute to improved food security in developing countries. Despite their importance, no operational forecasting system on sub-national level is yet in place in Tanzania. We develop a statistical maize yield forecast based on regional yield statistics in Tanzania and climatic predictors, covering the period 2009–2019. We forecast both yield anomalies and absolute yields at the sub-national scale about 6 weeks before the harvest. The forecasted yield anomalies (absolute yields) have a median Nash–Sutcliffe efficiency coefficient of 0.72 (0.79) in the out-of-sample cross validation, which corresponds to a median root mean squared error of 0.13 t/ha for absolute yields. In addition, we perform an out-of-sample variable selection and produce completely independent yield forecasts for the harvest year 2019. Our study is potentially applicable to other countries with short time series of yield data and inaccessible or low quality weather data due to the usage of only global climate data and a strict and transparent assessment of the forecasting skill.

Highlights

The models show a high performance for almost the whole country (18 out of 21 regions have an NSEv1 of higher than 0.3; Fig. 1)
The three most often selected variables are the number of times the sea surface temperature (SST) falls below the 1% percentile of the West Pacific considering a lead time of 120 days, the number of times the SST falls below the 1% percentile of the Indian Ocean Dipole considering a lead time of 30 days and the median SST of the IOD considering a lead time of 30 days
Our study provides a within-season maize yield forecast for entire Tanzania and is, to our best knowledge, the first of its kind

Summary

Introduction

The variable selection reveals the strong influence of extreme weather events on maize yields in Tanzania. We performed a within-season forecast by including the weather and SST variables related to the vegetative phase of the growing season. The full model evaluation and the level 1 validation indicate a high skill (median NSE of 0.93 and 0.79) of the within-season forecast of absolute yields for the whole country (Fig. 7).

Results

Conclusion