Quadro 8000

We used two performance metrics for evaluating the models: the root mean square error (RMSEp) and the relative RMSEp percentage (rRMSEp, obtained by normalizing the crop specific average yield). The metrics were aggregated at national level computing the average province-level error metric.
The models were evaluated on all of the eight forecasting months, from December to July. The best-performing set of hyperparameters was chosen based on average province-level rRMSEp. The workflow was written in Python, and the models were defined with Keras/TensorFlow libraries (Abadi et al., 2016 ; Keras, 2015/2022 ). It is a fully automated, end-to-end, processing tool (https://github.com/ec-jrc/ml4cast-yieldcnn) and it was executed on the JRC Big Data Platform (Soille et al., 2018 (link)) using a GPU node equipped with NVIDIA Quadro 8000.

We used two performance metrics for evaluating the models: the root mean square error (RMSEp) and the relative RMSEp percentage (rRMSEp, obtained by normalizing the crop specific average yield). The metrics were aggregated at national level computing the average province-level error metric.
The models were evaluated on all of the eight forecasting months, from December to July. The best-performing set of hyperparameters was chosen based on average province-level rRMSEp. The workflow was written in Python, and the models were defined with Keras/TensorFlow libraries (Abadi et al., 2016 ; Keras, 2015/2022 ). It is a fully automated, end-to-end, processing tool (https://github.com/ec-jrc/ml4cast-yieldcnn) and it was executed on the JRC Big Data Platform (Soille et al., 2018 (link)) using a GPU node equipped with NVIDIA Quadro 8000.