Abstract:
This paper presents a new hybrid framework that expands the predictive power of deep learning models to the soundness of statistical methods, thereby improving the accuracy, efficiency, and scalability of the estimation of outputs in agriculture. The concerns in this work that are being addressed are the requirement of large quantities of quality data, as well as computational requirements with the use of sophisticated machine learning models, thus precluding the general application of such techniques in agricultural practice. Having a clear understanding of the problem statement, the paper details the diverse deep learning architectures, principally in the form of EfficientNetB0 and InceptionV3, known to be computationally efficient in handling complex, high-dimensional data. These are further hybridized with some of the most fundamental statistical techniques, among which is linear regression which acts as a stabilizer of predictions, reducing the risk of overfitting that is found in some other purely deep learning-driven techniques. The resulting hybrid models show an increase in performance in predicting agricultural yields across different datasets in comparison to the individual deep learning or statistical models tested. These models have been shown to be able to predict accurately across diverse crop species and environment settings, a feature of importance in the context of potential large international applications in agriculture. Other combinations of deep learning and statistical methods are incorporated in the design of the framework, which is additionally designed to be tunable to specific localities or crops through hyperparameter tuning. In addition, the discussed hybrid models increase the performance of the model and cut down computation times largely, with accuracy being preserved high, which serves as a practical solution to yield predictions.
