Crop choice is a critical decision for rainfed smallholder farmers when allocating land between food and cash crops. To inform crop choice, process-based models need to simulate yield responses that are both eco-physiologically plausible and quantitatively accurate. Achieving this is difficult when data quality and scarcity hinder model calibration. Here, we present a modification of a process model simulation performed using a machine learning residual model trained to predict the error in the process model-simulated yields, relative to field experimental data, from growing conditions. Using the random forest (RF) algorithm, residual models were developed for cowpea, groundnut, soybean, maize, millet, and sorghum cultivated at three locations in Burkina Faso. The RF residual models improved the agreement between the process model simulations and the field data while preserving plausible crop-specific rainfall–yield relationships and their variation across soil types with differing water retention or drainage capacities (i.e., Lixisols and Plinthosols). Subsequently, process model simulations for 1994–2023 were adjusted using the RF residual models. The findings showed that the better performing crops varied with respect to soil type and seasonal rainfall. However, the utility of presowing rainfall forecasts for dynamic crop choice was limited by relatively high miss rates. The proposed crop choice advisory is expected to increase the income and nutrient status of smallholder farmers in dryland regions of West Africa under rainfall variability.
扫码关注我们
求助内容:
应助结果提醒方式:
