Current study was conducted in collaboration with Department of Computer Sciences and Engineering, University of Engineering and technology, Lahore, Pakistan during 2020. As a result, the focus of this work was on the development of a pre-harvest crop yield forecasting model based on soil and environmental parameters. These parameters were recorded on a monthly basis at National Tea & High Value Crops Research Institute (NTHRI) for a period of ten years. The parameters recorded were minimum and maximum temperature, humidity, rainfall, soil pH level, pesticide use, and labor expertise. To construct the feature set for model training, exploratory feature analysis, outlier analysis, feature scaling, feature transformation and feature selection using the ReliefF algorithm were performed. Through 10-fold cross validation, six regression algorithms were used for training and model evaluation. An ensemble of neural networks were used to build the final model. The ensemble were built using a novel method that trains several base learners with architectural and training data diversity. These trained models were ranked using the ReliefF algorithm and sequentially added to the ensemble until the validation performance stops improving. On the basis of four performance metrics, the final model was compared to the four best performing models. The proposed model provided mean averave error (MAE), mean squared error (MSE) and root mean squared error (RMSE) of 0.0942, 0.0145, and 0.1204, respectively, and an R-squared of 0.9461. The performance parameters were the best among the candidate models and were sufficient to justify its use as a tea crop yield prediction model.