AbstractAccurate predictions of herbage biomass are important for efficient grazing management. Small‐scale farms face challenges using remote sensing technologies due to insufficient resources. This limitation hinders their ability to develop machine learning‐based prediction models. An alternative is to adopt less expensive measurement methods and readily available data such as weather data. This study aimed to examine how different temporal aggregations of weather data combined with compressed sward height (CSH) affect the prediction performance. We considered weather features based on different numbers of weather variables, statistical functions, weather events, and periods. Between 2019 and 2021, data were collected from 11 organic dairy farms in Germany. Herbage biomass exhibited high variability (coefficient of variation [CV] = 0.65). Weather data were obtained from on‐farm and nearby public stations. Prediction models were learned on a training set (n = 291) and evaluated on a test set (n = 125). Random forest models performed better than models based on artificial neural networks and support vector regression. Representing weather data by a single feature for leaf wetness reduced the root mean square error (RMSE) by 12.1% (from 536 to 471 kg DM ha−1, where DM is dry matter) and increased the R2 by 0.109 (from 0.518 to 0.627). Adding features based on multiple variables, functions, events, and periods resulted in a further reduction in RMSE by 15.9% (R2 = 0.737). Overall, different aggregations of weather data enhanced the accuracy of CSH‐based models. These aggregations do not cause additional effort for data collection and, therefore, should be integrated into CSH‐based models for small‐scale farms.
Read full abstract