Seasonal forecasting of daily mean air temperatures using a coupled global climate model and machine learning algorithm for field-scale agricultural management

Abstract

Seasonal climate forecasting can facilitate sound agricultural decision-making. However, agricultural applications, such as yield prediction and sowing timing, require seasonal forecast data at high temporal (daily) and spatial (field-scale) resolutions. In this study, a hybrid model was proposed that coupled a global climate model (GloSea5GC2) with a regularized extreme learning machine (RELM) to perform seasonal forecasting (up to 90 days) of daily mean air temperatures at the field scale. Twenty candidate frameworks for the hybrid model were developed by employing different combinations of the transformation scheme, numbers of climate model ensemble members, and learning algorithms. The output blending method was also evaluated. Results showed that the hybrid models demonstrated a good capacity for long-range prediction at the field scale. In particular, the hybrid model (root mean square error, RMSE, 1.02–3.35) demonstrated better predictive skill compared to the climatology model (RMSE 1.61–3.37) at all lead times. When the hybrid model frameworks were compared, the framework which combined centered data with hindcast data, ensemble learning, and the member mean method was the most appropriate. Adopting the centering scheme with hindcast data as RELM input data had the largest influence on hybrid model predictive skill, improving model performance by 3–20% based on RMSE. The output blending method was found to be a suitable alternative when hindcast data were unavailable. The results of this study provide insight into the development of seasonal forecast hybrid models for agro-climatic variables such as minimum and maximum air temperature, relative humidity, and specific humidity at the field scale.