Apple fruit surface temperature prediction using weather data-driven machine learning models

Heat stress to maturing apple fruit is a key concern to tree fruit growers in the Pacific Northwest region of the United States and around the globe. Localized weather-based fruit surface temperature (FST) prediction, a key indicator of fruit stress, can help in planning better mitigation strategies and ultimately reduce crop losses. Therefore, this study evaluated localized weather (solar radiation, temperature, relative humidity, dew point, and wind speed) and fruit size data driven multiple linear regression (MLR) and Long Short-Term Memory (LSTM) models for predicting apple FST. The models were trained on either the localized in-orchard or open field weather station data collected in the 2022 field season and validated against the actual FST of ‘Honeycrisp’ apple. The MLR model was able to predict FST with an average root mean square error (RMSE) of 2. $1^{\circ}\mathrm{C}$ using the in-orchard weather and fruit size dataset as inputs. The LSTM model prediction average RMSE for the same dataset was 2. $3^{\circ}\mathrm{C}$. Using open field weather data, the RMSE was 2. $69^{\circ}\mathrm{C}$ and 2. $25^{\circ}\mathrm{C}$ for the LSTM and MLR models, respectively. Additionally, both in-orchard and open-field trained models outperformed the existing energy balance FST prediction approach on the same dataset. Overall, these findings can be helpful to growers for real-time and reliable FST monitoring using localized as well as publicly available weather data inputs.