D3-C2 Future Cropping Season

A common strategy that farmers use to adapt to climate change is shifting the cropping season by manipulating sowing and harvest dates. It is not clear to which extent climate influences this decision, what is the role of other agri-environmental variables like soil type, preceding crop and groundwater shallowness, or of completely different factors. However, when forecasts of crop growth and development under future climate scenarios are made (e.g. with crop models), the start of the growing season is often determined using simple rules or thresholds related to temperature, rainfall or other weather variables. This is an optimistic representation of how farmers could adapt to climate change, but does not mimic what would happen in a ‘business as usual’ situation.

In this context, remote sensing and artificial intelligence can come in handy. Satellite observations of previous cropping seasons can be used to determine key moments of crop development and management (e.g. emergence and harvest). Artificial intelligence can link these key moments with other data, such as historical weather trends and soil type, exploiting the deep patterns in the data to create predictive models. These models can be applied to future climate scenarios to estimate cropping season features without satellite observations already available.

Project description

The project integrates several data sources and uses artificial intelligence to predict future cropping seasons in the Netherlands. The goal of the project is to build a framework to predict realistic emergence and harvest dates under future climate, that can be expanded as new data becomes available over the years.

Groenmonitor constitutes the core of the system: the platform analyses remote sensing images to detect key moments in crop development and management, through NDVI curve fitting and radar coherence analysis. Groenmonitor is a source of spatialised data that covers the entire Netherlands. This system offers a unique opportunity to test the predictive capabilities of artificial intelligence algorithms, using key moments of crop development and growth (e.g. emergence and harvest) as outcome variables, and weather, crop, soil and groundwater data as predictors. Historical weather time–series (KNMI and JRC) are the primary continuous predictors, to which categorical covariates from BOFEK2020 (topsoil type), BRO grondwaterspiegeldiepte (groundwater class), Groenmonitor–BRP (preceding crop) are added. Models are being trained and tested on historical data, and will be used to predict emergence and harvest dates under future weather obtained from the spatiotemporal downscaling of global circulation models.

This system differs from rule–based approaches: artificial intelligence exploits deep patterns among data, additional covariates are added to refine the predictions, and the spatial extent and size of Groenmonitor data allows us to train models based on a wide variety of pedoclimatic conditions. Since Groenmonitor is maintained and improved under the WUR umbrella, the workflow created with this project will have the potential to be expanded with additional satellite data as the years go by, making the predictions more accurate over time. The predictions will also enable crop modellers to set more realistic starts and ends of growing seasons, in their studies of future climate impact. Finally, an interactive dashboard will be built to easily visualise these predictions in the form of maps.

Results

The data sources have been integrated into a single database by spatially joining shapefiles from several sources. The data sources retrieved, homogenised and included so far are:

Predictors:

• Historical weather time series (KNMI and JRC)
• Topsoil class (BOFEK2020)
• Groundwater class (BRO grondwaterspiegeldiepte)
• Current and preceding crop (Groenmonitor–BRP)

Outcomes:

• Emergence and harvest dates (Groenmonitor NDVI analysis)