Unveiling biases on agroclimatic indicators : assessment of gridded climate data SAFRAN, ERA5-Land, and EOBS, against station-based observations in France.

Three gridded datasets (SAFRAN, ERA5-Land and EOBS) are compared to INRAE’s and Météo-France’s observed data from their respective weather stations networks. The Météo France network comprises 37 synoptic stations, while the INRAE network comprises 49 stations. This work analyses the bias between the gridded data and the stations’ ones. It aims to quantify the bias between those three datasets in terms of climatic parameters, as well as their repercussions on agroclimatic indicators and on the plant phenology cycle (in this case represented by wheat).
While historically studies of climate change and its impacts have relied on data from weather stations (located in a given place), in recent years we have observed more and more studies using gridded climatic data. Their value lies in the fact that these data enable the climate of a territory to be represented spatially (rather than at a single point), and they also ensure the continuity of all climatic variables. These two characteristics make them particularly useful for impact studies. Furthermore, this data is used to correct climate projections (e.g. CORDEX) at different scales. Many gridded datasets have been created with diverse characteristics and so equally diverse data values.
The datasets are, in the first instance, studied in regard to a set of weather parameters: minimal, mean and maximal temperatures and precipitations. The mean temperature is then incorporated in a phenology model to simulate the wheat’s phenology cycle. Simultaneously, the minimal and maximal temperatures are also used to calculate three agroclimatic indicators: number of frost days, number of days with maximal temperatures over 25°C (as an important threshold for wheat yield elaboration) and over 35°C (considered as a critical threshold for plant development and growth). In a second phase, the results are analysed to identify if the biases between the gridded data and the stations’ ones are changing seasonally, annually or depending on the value of the parameter.
We found that for the mean temperature and the phenology cycle the biases are not significative. The bias obtained for simulating phenology stages is in majority under the 5 days admissible error (which could be due to an observation error). For those two indicators, SAFRAN is showing the best results. In regard to the minimal and maximal temperature and the matching agroclimatic indicators, EOBS is showing lowest bias and ERA5-Land is showing the highest bias. We could also highlight a seasonality in the bias of the minimal temperature for SAFRAN and ERA5-Land, and a bias depending on the value of the parameter.
This work presents a method for identifying biases in a dataset, that can be applied to various parameters and impact studies. It quantifies the accuracy of the gridded data used in these studies and determines whether the biases are indicative. Furthermore, it illustrates the extent to which these biases shape the evaluation of indicators like phenology dates and climate-related risks to crop production. Finally, it helps users choose the most suitable dataset for their needs.