Is deeper always better? Evaluating deep learning models for yield forecasting with small data

In this study, we investigate a flexible deep learning approach to forecast crop yield at the provincial administrative level based on deep 1D and 2D convolutional neural networks with small data. Our approach meets the operational requirements (public and global records of satellite data in an application ready format with near real time updates) of the application and can be transferred to any country where yield statistics are available. Three-dimensional histograms of Normalized Difference Vegetation Index (NDVI) and climate data are used as input to the 2D model and administrative level time series averages of NDVI and climate data are inputted to the 1D model. The best model architecture is identified during efficient and extensive hyperparameter optimization. To demonstrate the relevance of our approach, we hindcast (2002-2018) the yields of Algeria’s three main crops (barley, durum and soft wheat) and contrast its performance with machine learning algorithms and conventional benchmark models used in the previous study. We show that simple benchmarks such as peak NDVI remained difficult to beat and that machine learning models outperformed deep learning models for all forecasting months and all tested crops. We attribute this poor performance to the smallness of the dataset available. All the data inputs are free and accessible for download from https://mars.jrc.ec.europa.eu/asap/download.php.

SZABO Filip;  MERONI Michele;  WALDNER Francois;  REMBOLD Felix; 

2023-09-08

SPRINGER

JRC132054

0167-6369 (online),   

https://link.springer.com/article/10.1007/s10661-023-11609-8,    https://publications.jrc.ec.europa.eu/repository/handle/JRC132054,   

10.1007/s10661-023-11609-8 (online),