Using boosted tree regression and artificial neural networks to forecast upland rice yield under climate change in Sahel

•Rainfall, not temperature, was the main climate driver of the rice yield in Sahel.•Rice yield response function was modeled and tested against observed yield data.•ANN overperformed boosted tree and multiple linear regression for modeling rice yield.•Effect of climate change on rice yield in rainfed areas of Sahel was assessed.•Projected yield showed a gap of 57.29% with recorded maximum average yield over 2052. Climate drivers are key stress factors affecting upland rice yields in Sahel because the region is vulnerable to unfavorable weather and has a very low adaptive capacity. This study modeled upland rice yield responses to climate factors using multiple linear regression, boosted tree regression, and artificial neural networks (ANNs). Four ANNs were explored: ANNMLP (multilayer perceptron), ANNPNN (probabilistic neural network), ANNGFF (generalized feedforward), and ANNLR (linear regression). Then the modeled rice yield function was calibrated and tested against the observed yield data and climate variables of three provinces of Burkina Faso, West Africa. The global climate model (GCM) outputs under the AR4-SR-A1B, A2, and B1 mean ensemble CO2 emissions scenarios were then downscaled and used as input of the calibrated yield response model, in order to forecast yield trends over 2052. The results are three-fold: first, rain (R = 0.402) is the most dominant climate driver in Sahel, followed by the maximum and minimum temperatures (R = -0.313 and R = -0.237, respectively), which clearly reduce yield. Second, the ANNPNN (R = 0.952, MSE = 0.033 ton/ha, NMSE = 0.109 ton/ha, MAE = 0.115 ton/ha) has a great capability in rice yield responses function modeling outperforming boosted tree (R = 0.920, MSE = 0.077 ton/ha, NMSE = 0.208 ton/ha, MAE = 0.223 ton/ha) and the multiple linear regression (R = 0.385, MSE = 0.259 ton/ha, NMSE = 0.852 ton/ha, MAE = 0.340 ton/ha). All linear models performed unsatisfactorily. Third, the projected yields showed a gap of 57.29% with the site-recorded maximum average yields over 2052. From application of ANNPNN, we anticipate that site-specific rice yield may substantially decline with climate change, as rainfall is projected to decrease while temperatures increase. These results should assist in identifying priority adaptation measures for Sahel, such as village rainwater catchment basins supplemented with adapted irrigation technologies, to enhance the resilience of crops.