An Artificial Intelligent-Based System for Crop Yield Prediction Using Climate Change Data and Sensor Fusion

Abstract

The world’s food security is in danger as climate change keeps reducing the annual corps-growth rate and the world’s population grows exponentially. Predicting crop-yield has been a common practice since the beginning of time; farmers would decide where and when to plant different types of plants based on their basic knowledge of weather and soil properties and what are the suitable conditions for every crop. Improving our forecasting skills of crop-yields is crucial for feeding the growing population while reducing the negative impacts of the agriculture industry on the environment. The fluctuating climatic features pose a threat that is increasing day-by-day and affecting our crops. Not only do the fluctuating weather conditions reduce agricultural productivity, but it also reduces the nutritious value and health benefits gained from the plants. That’s because of the increase in the tonnes of pesticides used on the fields and farms as a way of coping with climate change and pollution effects and increasing the plants’ growth rate. Predicting crop-yield using Machine Learning (ML) and Deep Learning (DL) has been widely adopted by the AI community in the past few decades; however, some obstacles and bottleneck challenges that affect its accuracy remains. This paper aims to propose a forecasting framework that is easy to use, inexpensive, and reliable enough to be used and certified by governments, companies, or individuals who depend on harvest for living. The proposed framework shows how Sensor Fusion can be beneficial to the agriculture industry specifically and artificial intelligence field in general. The framework demonstrates how Sensor Fusion can be incorporated as part of the data-processing. To accommodate for the crops that require specific soil quality to ensure maximum yield, an NPK sensor and a PH level sensor will be added to measure the acidity of the soil along with the levels of nitrogen, phosphorus, and potassium. Four supervised regression machine learning models were implemented in the proposed framework, namely, Gradient Boosting Regressor, Random Forest Regressor, SVR and Decision Tree Regressor, as they appropriate for the utilized data type and size. The dataset used in the experiments includes the rainfall, yield, year, temperature, and pesticides usage and covers 101 countries over the span of 23 years. After implementing the models, The decision tree had an accuracy rate of 98% ${R}^{2}$ score which was the best among the implemented models and among all studies conducted using decision tree regressor in general. The proposed framework ensures that it can be used to enhance the model with the best results is discussed in-depth.

S. Elghamrawy—Senior IEEE Member.