Abstract

The agricultural sector has contributed tremendously to the development of many countries in the world, by providing food and employment. However, knowing the number of crops produced yearly has become a major challenge. This paper introduced an advanced framework known as Yield-Net which combines XGBoost, Gated Recurrent Unit (GRU) and Random Forest (RF) machine learning techniques in predicting annual crop yield from different regions. The dataset used for this experiment is the world crop yield dataset covering the period of 1990 to 2010 sourced from ourworldindata.org with 28242 instances with 7 variables. Data preprocessing steps such as handling missing entries, dealing with duplicates, data encoding and scaling were applied in order to structure the dataset in a format suitable for modelling. The preprocessed data was portioned into segments, 90% for training and 10% for Yield-Net testing. XGBoost was employed in this study for selecting important features from the crop yield data for model training while GRU extracted sequential patterns from the selected features. Random Forest predicts the amount of crop yield using the extracted features from GRU as input. The Yield-Net model was trained in 20 iterations. State-of-the-art metrics widely used in assessing the level of error and performance of regression model, Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Root Square Error (R2) were employed. The proposed Yield-Net system achieves RMSE score of 5.3, MAE result of 3.95 and R2 = 0.94 outperforming baseline models developed for similar tasks. This paper contributes significantly to the ongoing research in the field of precision agriculture while laying a solid foundation for future work.