SCHOOL OF GRADUATE STUDIES DECLERATION & ADVISORS’ APPROVAL SHEET

Abstract

Ethiopia's economy relies heavily on agriculture, but crop yields are insufficient and the country faces a food deficit, making it one of the world's poorest countries. One of the main reasons is the difficulty in determining which crop to plant based on soil characteristics and climate conditions. Most farmers plant the wrong crop at the wrong location, soil type, and season without taking into account the crop's requirements. As a result of these major issues, their crop production is lower. Therefore, technological support is needed to help farmers make informed decisions about what crops to grow. According to the reviewed literature, there is a lack of a universal crop prediction model to support different countries because of differences in soil and climate factors from place to place. Also, there isn't enough research being done in this area in Ethiopia. For these reasons, the researcher developed an artificially intelligent model that can learn from prior experience by analysing various soil and climate parameters and can predict the best crop to be sown. Thus, the goal of the proposed study is to develop a machine learning model that helps farmers choose the right crop for the right place and hence, enhance crop yields. To develop this model, the researcher employed Random Forest, Extra Trees Classifier, XGBoost, and AdaBoost ensemble algorithms; Artificial Neural Network, and Deep Neural Network along with stacking, blending, bagging, and voting ensemble techniques using Python's Sklearn, Keras, and TensorFlow libraries on a dataset acquired from Arba Minch Agricultural Research Center and Gamo Zone Agricultural Directive Office. The dataset consists of 6560 instances for 15 crops with 10 independent features, namely phosphorous, potassium, nitrogen, location (woreda), soil type, altitude, temperature, rainfall, humidity, and soil pH value, and 1 dependent feature (target class) the crop. Then, models are evaluated and compared using model evaluation metrics such as accuracy, precision, recall, f1_score, and confusion matrix. The experimental results revealed that the blending ensemble technique outperformed all other models with an accuracy of 99.4% and a weighted average of 99% for all (precision, recall, and f1_score) using an 80/20% data split. This is because ensemble techniques have a higher combining power to enhance the predictive performance than a single individual model. Thus, the study's findings show that a selected model is effective for predicting the best crop for a certain plot of land and can support the decision-making process while selecting suitable crops to grow