Interpretable machine learning methods to explain on-farm yield variability of high productivity wheat in Northwest India

The increasing availability of complex, geo-referenced on-farm data demands analytical frameworks that can guide crop management recommendations. Recent developments in interpretable machine learning techniques offer opportunities to use these methods in agronomic studies. Our objectives were two-fold: (1) to assess the performance of different machine learning methods to explain on-farm wheat yield variability in the Northwestern Indo-Gangetic Plains of India, and (2) to identify the most important drivers and interactions explaining wheat yield variability. A suite of fine-tuned machine learning models (ridge and lasso regression, classification and regression trees, k-nearest neighbor, support vector machines, gradient boosting, extreme gradient boosting, and random forest) were statistically compared using the R2, root mean square error (RMSE), and mean absolute error (MAE). The best performing model was again fine-tuned using a grid search approach for the bias-variance trade-off. Three post-hoc model agnostic techniques were used to interpret the best performing model: variable importance (a variable was considered “important” if shuffling its values increased or decreased the model error considerably), interaction strength (based on Friedman’s H-statistic), and two-way interaction (i.e., how much of the total variability in wheat yield was explained by a particular two-way interaction). Model outputs were compared against empirical data to contextualize results and provide a blueprint for future analysis in other production systems. Tree-based and decision boundary-based methods outperformed regression-based methods in explaining wheat yield variability. Random forest was the best performing method in terms of goodness-of-fit and model precision and accuracy with RMSE, MAE, and R2 ranging between 367 and 470 kg ha−1, 276–345 kg ha−1, and 0.44–0.63, respectively. Random forest was then used for selection of important variables and interactions. The most important management variables explaining wheat yield variability were nitrogen application rate and crop residue management, whereas the average of monthly cumulative solar radiation during February and March (coinciding with reproductive phase of wheat) was the most important biophysical variable. The effect size of these variables on wheat yield ranged between 227 kg ha−1 for nitrogen application rate to 372 kg ha−1 for cumulative solar radiation during February and March. The effect of important interactions on wheat yield was detected in the data namely the interaction between crop residue management and disease management and, nitrogen application rate and seeding rate. For instance, farmers’ fields with moderate disease incidence yielded 750 kg ha−1 less when crop residues were removed than when crop residues were retained. Similarly, wheat yield response to residue retention was higher under low seed and N application rates. As an inductive research approach, the appropriate application of interpretable machine learning methods can be used to extract agronomically actionable information from large-scale farmer field data.

Bibliographic Details

Main Authors:	Nayak, H., Silva, J.V., Parihar, C.M., Krupnik, T.J., Sena, D.R., Kakraliya Suresh Kumar, Jat, H.S., Sidhu, H.S., Sharma, P.C., Jat, M.L., Sapkota, T.B.
Format:	Article biblioteca
Language:	English
Published:	Elsevier 2022
Subjects:	AGRICULTURAL SCIENCES AND BIOTECHNOLOGY, FORESTS, MACHINE LEARNING, WHEAT, YIELDS, CROP RESIDUES,
Online Access:	https://hdl.handle.net/10883/22163
Tags:	Add Tag No Tags, Be the first to tag this record!