Recurrent neural network (RNN) based estimation of per-field wheat yield at an arbitrary time during the growth period

• A recurrent neural network (RNN) can estimate wheat yields at times during growth. • Reflectance, Land Surface Temperature (LST), and weather data were used. • The Leaf Area Index (LAI) and Fraction of Absorbed PAR (FPAR) were also used. • RNNs are useful to estimate per-field wheat yields at arbitrary times during growth. A recurrent neural network (RNN) was developed to estimate wheat yield per field at an arbitrary time during the growth period. After Reflectance, Land Surface Temperature (LST), Leaf Area Index (LAI), and Fraction of Absorbed Photosynthetically Active Radiation (FPAR) from satellite images, and weather data were entered into the RNN, the yield per field was estimated for winter wheat ‘Kitahonami’. During the three-year period of 2022–2024, yield data were obtained for 212 fields in Memuro-cho, Hokkaido, Japan. The vigorous growth season of ‘Kitahonami’ was divided into four periods based on wheat developmental stages. The yield estimation accuracy for each period was evaluated. Multiple linear regression (MLR) analysis based on the relation between reflectance and weather data during the growth period and yield was also used for comparison. By leave-one-year-out analysis for MLR, in which a model is built with data of two years and for which accuracy is evaluated with data of other years, the best estimation accuracy was obtained from different periods by year. Furthermore, notable systematic errors were found. However, the RNN showed stable estimation accuracy throughout all periods in all analysis years, with better accuracy than MLR found for almost all years and periods. The results indicate RNNs as useful for estimating the wheat yield per field at an arbitrary time during the growth period. However, findings suggest that the accuracy of the estimates might be reduced because of factors that were not considered, such as soil-borne viral diseases and poor soil water retention and drainage.