Geospatial crop yield modeling for climate change risk transfer, impact assessment, and adaptation strategies

Changing climate and weather patterns increase the pressure on food production systems worldwide, significantly impacting livelihoods and economies at various scales. Understanding and adapting/mitigating climate change are essential to cope with the adverse effects of climate change. Biophysical crop modeling offers a unique way to assess the pros and cons of different adaptation strategies because they integrate physical and biological principles to simulate the use and allocation of captured resources mechanistically. However, most crop modeling approaches use field-based or national scale modeling, which is difficult to transfer or upscale to the regional level. Geospatial modeling is required o develop climate change policies and adaptation strategies based on an integrated, evidence-based, and climate-smart approach to address national food security and farmers’ welfare at all levels. This dissertation focuses on developing geospatial crop yield modeling approaches. Then, the developed approaches are applied in three studies to address climate risk transfer, impact assessment, and evaluation of adaptation measures. First, we used the Decision Support System for Agrotechnology Transfer (DSSAT) to estimate rice yield and yield anomalies at ∼5 km spatial resolution to give a scientific basis for payout insurance on time (climate risk transfer) in India. Secondly, we further downscaled this method by utilizing a Gradient Boosted Regression (GBR), a machine learning technique, and remote sensing to estimate rice yields at ∼500 m spatial resolution for rice-producing areas in India. This application also serves to inform insurance as a climate risk transfer solution. Thirdly, we used the DSSAT to quantitatively assess Sorghum’s impacts on climate change and adaptation measures in Burkina Faso (impact assessment and evaluation of adaptation measures).

The first study uses a biophysical crop modeling approach and accesses publicly available datasets to get reliable near-real-time rice crop simulations for India, a major rice producer. Model simulations at ∼5 km spatial resolution attained a relative Root Mean Square Error (rRMSE) of less than 20% between observed and simulated yield anomalies in almost half of all rice-growing districts, a minimal error in insurance payouts. The outputs are available soon after harvest, a near-real-time yield loss estimation. This is especially useful for crop insurance decisions, particularly the Indian governmental PMFBY scheme. Moreover, this approach could be scaled to further regions and more crops, which utilize only publicly available datasets. Second, a downscaling approach to estimate rice yields in India on ∼500 m spatial resolution integrates remote sensing data with a machine learning technique that could potentially be applied for timely and reliable yield estimation. This approach concludes that the GBR is feasible across India and may complement current approaches for timely rice yield estimation required by insurance companies and government agencies. This approach also can be scaled up to various regions. However, scaling up depends on the availability and scale of observations. For the third study, we simulated sorghum yields in Burkina Faso using DSSAT under current and projected climatic conditions and evaluated four adaptation strategies (Integrated Soil Fertility Management, Irrigation, Improved varieties, and Agroforestry). These assessments provide further impetus for research in local trials and the implementation of adaptation measures by governments, non-governmental agencies, farmer’s organizations, extension workers, and farmers. It could also solve issues such as the lack of convincing demonstration that the technologies provide significant benefits.

This dissertation concludes that geospatial crop modeling is feasible for climate risk transfer and impact and adaptation studies that can be used for national and sub-national level planning purposes in data-scarce regions.