On-farm managed trials and machine learning approaches for understanding variability in soybean yield in Northern Ghana

AbstractBackground and aim : Soybean seeds inoculation with effective rhizobia (Rh) strains and phosphorus (P) application are agricultural best practices that enhance grain yield. However, in Northern Ghana, where these practices are progressively under adoption, unpredictable yield, and poor understanding of factors of yield variation often limit its potential. We assessed the influencing factors to soybean yield variability from biophysical and managed input variables (Rh inoculants, P rates, and sources). Methods On-station and on-farm soybean plots were inoculated with three Rh inoculants (Rh1, Rh2, and Rh3), treated with two P rates (0 and 30 kg P ha-1), and two P sources [rock phosphate and Triple superphosphate forms]. Yield data was predicted using the random forest (RF) model, and factors of yield variability were assessed using the linear mixed models and the forward redundancy analysis (rda). Results The yield prediction accuracy was greater for the on-station experiment compared to the on-farm dataset with a trained coefficient of determination (R2) of 0.77 and 0.66, respectively. The top variables of yield prediction were the Rh × P fertilizer, P sources, Rh strains, and exchangeable soil Mg2+concentrations. The Rh × P treatment increased soybean grain yield by 3.0 and 3.9 folds for the on-farm and on-station trials respectively, compared to the control. Conclusion The RF model and the forward rda unearthed a significant contribution of the soil exchangeable Mg2+to the yield variation. The mechanisms underlying the role of Mg on soybean growth deserve further research investigations to increase soybean production in Ghana sustainably.