Résumé
Sociohydrology aims to deliver user-inspired solutions to water challenges, often through model-based understanding and simulation of local realities. However, sociohydrological modeling methodologies used to understand such complex human-water systems remain difficult to apply to many real-world case studies. Sociohydrological model predictions at scales, such as daily to annual time scale, of decision making remains a challenge due to often difficult to acquire social sciences data, and missing or unknown feedbacks that lead to model structural errors, amongst others. This paper assesses and reduces model structural deficiency of a smallholder sociohydrological (SH) model when applied to a case study of small-scale agricultural production in India, where variables from a farmer survey help alleviate structural deficiency. A structural error model is proposed based on a regression model of nonlinear projection of the these variables to a Kernel space, called Kernel Principal Component Analysis (KPCA) based model. Based on this, a hybrid model that is a sum of SH model and the structural error model is proposed. It offers significantly better yield predictions on ’unseen’ (to the model) survey data than the SH only model. The hybrid model also performs better than a KPCA only model based yield prediction, which predicts yields without any SH dynamics. This is because the hybrid model combines the structural error model that learns from the spatial pattern of observed yields with the temporal dynamics explained by the SH only model. The results indicate that the structure of the SH model can be improved by further incorporation of irrigation and adaptive behavior of farmers.