Recommendation model for rice precision fertilization using knowledge graph and case-based reasoning

Abstract: Fertilizer management in rice fields can play a very important role in rice yield and quality. However, blind fertilization has caused a series of problems, such as soil consolidation, fertilizer waste, and environmental pollution. As a result, it is a high demand to automatically make the fertilization plan, according to the rice growth cycle. There are also some defects in the current agricultural planning using mathematical models, domain empirical knowledge, and statistical artificial intelligence (AI) models. In this study, a novel recommendation model of rice precision fertilization was proposed to optimize the formulation of a rice fertilization plan using a knowledge graph and case-based reasoning. Qualitative fertilization suggestions were also provided for different fertilization stages. In the knowledge graph, the information was first retrieved to obtain the initial fertilization scheme, according to the cultivated rice variety. The types of information were then compared between the initial fertilization scheme and the required recommended one, in order to obtain the missing types in the initial scheme. Secondly, the PairRE model was used to determine the distributed low-dimensional vectors of all entities and relations in the knowledge graph. The missing part of the rice fertilization scheme was then predicted using these entity and relation vectors. Finally, the predicted results were presented as a fertilization scheme, together with the initial one. In case-based reasoning stage, the k historical cases which were most similar to the event to be recommended were retrieved, according to the environmental parameters, the cultivated rice variety, the yield level of the event to be recommended, the parameters of historical cases, and the vectors of entities in the knowledge graph. The k cases were then used for the specific application in the different kinds of fertilizers through combination prediction. In the model verification, the average proportion of fertilization recommendation contrary to the actual was only 10.76% in different test events, particularly in the part of a qualitative acquisition in the fertilization scheme. The prediction accuracy of the nitrogen (N), phosphorus (P2O5), and potassium (K2O) application amount, as well as ratio of basal-tillering nitrogen application amount to panicle nitrogen application amount, reached 92.85%, 82.61%, 79.17%, and 90.92%, respectively. Correspondingly, Spearman's rank correlation coefficients were 0.69, 0.76, 0.74, and 0.72, respectively, for the correlation between the predicted and the actual values in the four indexes. A comparison experiment was then performed between the commonly-used recommendation method based on rule-based reasoning and the proposed model. Consequently, the proposed model can be expected to ensure the accuracy of the rice fertilization scheme. A more complete fertilization scheme and specific fertilizer application amount were also provided in the higher practical value. In summary, the developed recommendation model of rice precision fertilization using case-based reasoning can be the key supplement in the fertilization recommendation using a knowledge graph. As such, the recommendation system was effectively improved in order to fully use data for the numerical prediction. The proposed model has strong interpretability and high accuracy, particularly for the integrity of the fertilization scheme and fertilizer application amount. The finding can provide a strong reference for the formulation of a rice fertilization plan.