Estimation of NPK from soil data using a novel stacked ensemble model

Soil fertility is defined in terms of various nutrients present in the soil and has a direct impact on the sustained production of crops. Amongst various nutrients, nitrogen (N), phosphorous (P), and potassium (K), all together NPK, play the most significant role in gross crop yield. Precise estimation of the fertility status of existing soil NPK is the most significant task for proper fertilizer management. Traditionally, the fertility status of NPK is estimated by chemical analysis, which is expert- dependent and expensive for field applications. Thus, rural farmers apply various fertilizers without estimating the current fertility status of the nutrients, which leads to a nutrient imbalance and a serious threat to sustainable agriculture. Instead, several traditional machine learning models have been suggested for estimating the fertility status of various soil nutrients. However, they have some major limitations with poor performances. Conversely, the stacked ensemble models are more stable and robust compared to base learners with better predictive accuracy. Several successful applications of such models have been reported. However, no stacked ensemble model has been designed for estimating the fertility status of NPK using a trustworthy soil data repository. This paper proposes a novel architecture of a stacked ensemble model for precise estimation of the fertility status of three macronutrients, NPK. The novelty of this work is that the architecture of the existing stacked ensemble models has been restructured by incorporating an additional preprocessing module along with a data feedback technique to boost up the performance. Undoubtedly, a web-based version of such a model will be beneficial for rural farmers and agricultural decision-makers when accessed using smartphones. Our proposed model was experimentally validated using authentic field datasets. Empirical results revealed that our proposed model outperformed all other existing models both in terms of average accuracy (0.8457) and average Cohen’s kappa (0.6114).