Optimal county-level crop yield prediction using MODIS-based variables and weather data: A comparative study on machine learning models

• Published in: Agricultural and forest meteorology 2021-09, Vol.307, p.108530, Article 108530
• Main Authors: Ju, Sungha, Lim, Hyoungjoon, Ma, Jong Won, Kim, Soohyun, Lee, Kyungdo, Zhao, Shuhe, Heo, Joon
• Format: Article
• Language: English
• Subjects: Crop yield prediction; Machine learning; MODIS; Weather data
• ISSN: 0168-1923; 1873-2240
• DOI: 10.1016/j.agrformet.2021.108530

•Examine the performances of machine learning models for crop yield prediction.•Find an appropriate prediction model with given spatial, temporal resolution.•Attempt to discover suitable periods of input dataset for machine learning models. Accurate crop yield prediction for more precise forecasting of price volatility in crop markets, better agricultural planning and enhanced national food security is one of the important utilities of crop monitoring systems. Due to the spatiotemporal nonlinear characteristic of crop yields, recent studies have actively used various machine learning approaches to predict crop yields. However, there are few studies that have compared these approaches, and even in those cases, the comparison was restricted to one or two crops or a specific region. In this study, we evaluated seven popular machine learning approaches, for the same input variables, over three crops: paddy rice in South Korea and corn and soybean in Illinois and Iowa, USA. Based on the data from April to September, six time-series scenarios of different range of months for each crop were tested for prediction accuracy with 14-year (2003-2016). The time-series data include vegetation indices from Moderate resolution imaging spectroradiometer (MODIS), weather data, crop yield statistics, and a land cover map of county-level spatial resolution and 16-day-aggregated temporal resolution. In the results, regardless of crop type, support vector machine (SVM) presented the most accurate result with the lowest average root mean square error (RRMSE), in comparison with a decision tree (DT), random forest (RF), artificial neural network (ANN), stacked-sparse autoencoder (SSAE), convolutional neural network (CNN), and long short-term memory (LSTM). Also, we derived best cases for each crop yield prediction out of six time-series scenarios, which were May – September data for rice and corn and June – August data for soybean for the test sites. Considering all of the findings together, SVM is an appropriate method for crop yield prediction based on MODIS-based vegetation indices and weather data of county-level spatial resolution and approximately one-half-month temporal resolution.