Combination of crop models and machine learning techniques for agricultural parametric insurance

Agriculture is highly exposed to the effects of weather and extreme events play a crucial role in lowering crop yields. Low crop production has devastating effects on farmers from the economic point of view and undermines food security. Thus, crop insurance constitutes an ex-ante formal tool adopted in many countries to secure farmers’ income.

The interest in index-based (or parametric) insurance in the agricultural sector has grown in recent years and many different parametric products are nowadays available for farmers both in high and low-income countries. While traditional insurance evaluates the claims assessing crop losses in the field after an event, index-based insurance calculates indemnities based on an independent proxy for yield losses, as for example a weather index.

Index-based insurance exhibits many advantages with respect to traditional; it overcomes the issues of moral hazard and adverse selection, farmers receive payouts quickly since there is no need of in-situ inspections, administrative costs are lower with respect of the ones of traditional insurance, etc.

However, parametric products are subjected to high basis risk since the relationship between the weather index and farmers losses is imperfect and affected by high uncertainty. The minimization of basis risk is the main challenge of parametric products and could be obtained by developing indices that reproduce as accurately as possible the relationship between climate and yield.

Nowadays, in parametric insurance products the use of rainfall and temperature-based indices is prevalent with respect to the application of drought, floods, or soil moisture-based indices, even if the latter are more accurate in reproducing farmers losses. The reason behind this choice is that farmers prefer products based on variables easy to understand and measure.

In addition, the major part of parametric insurance products estimates the yield-index relationship through the use of statistical methods, such as regression, correlation, copulas or probability distribution. The use of mechanistic methods as crop modelling, and machine learning techniques deserves to be further explored since preliminary studies have demonstrated their potential in producing accurate yield-index relationships, even if a huge amount of data is required to successfully set up the models.

This study explores the use of a combination of crop models and machine learning methods to establish an accurate yield-index relationship. At the same time the proposed index should be directly related to a simple weather variable (such as rainfall or temperature) through tables or functions easy to understand for farmers.

Various crop models, such as APSIM, WOFOST and AquaCrop were tested, together with different machine learning techniques, namely CNN and random forest, explaining the outcome with the aid of SHAP values, creating an output transparent and easier to understand for farmers.

The case study area is Northern Italy, given the availability of observed yield data Weather data have been retrieved from various sources, such as satellite products (CHIRPS), reanalysis (ERA-5, SPHERA, etc.) and weather stations, while soil data (soil texture and water content) derive from the SoilGrids database and the FAO harmonized soil database.

Preliminary results have shown good correlations between maize and wheat yields simulated with crop models and observed yields.