Accurately mapping global wheat production system using deep learning algorithms

Abstract

Assessing global food security and developing sustainable production systems need spatially explicit information on crop harvesting areas and yields; however the available datasets are spatially and temporally coarse. Here, we developed a general framework, Global Wheat Production Mapping System (GWPMS), to map the spatial distribution of wheat harvesting area and estimate yield using data-driven models across eight major wheat-producing countries worldwide. We found GWPMS could not only generate robust wheat maps with R2 consistently greater than 0.8, but also successfully captured a substantial fraction of yield variations with an average of 76%. The developed long short-term memory model outperformed other machine learning algorithms because it characterized the nonlinear and cumulative impacts of meteorological factors on yield. Using the derived wheat maps improved R2 by 6.7% compared to a popularly used dataset. GWPMS is able to map spatial distribution of harvesting areas in a scalable way and further estimate gridded-yield robustly, and it can be applied globally using publicly available data. GWPMS and the resultant datasets will greatly accelerate our understanding and studies on global food security.

Keywords

• Wheat
• Crop mapping
• Yield estimation
• Deep learning
• Remote sensing