Deep Gaussian Process Models in the Classification of Cassava Diseases

Abstract

Machine learning continues to transform how plant diseases are identified and classified, offering scalable solutions for timely crop health monitoring. Despite significant progress, many conventional models struggle with quantifying predictive uncertainty, a critical requirement in decision-sensitive domains such as agriculture. Gaussian Processes (GPs), known for their ability to provide calibrated predictions, offer a principled way to address this gap. This study leverages the strengths of Deep Gaussian Processes (DGPs), Convolutional Neural Networks (CNNs), and Transfer Learning to improve cassava disease classification. The research aimed to address the following objectives: To Evaluate existing Gaussian Process models and identify their architectural strengths, limitations, and suitability for cassava disease classification; to design and implement a Deep Gaussian Convolutional Neural Network (DGCNN) that integrates GPs and CNNs; to develop a Deep Gaussian Transfer Learning (DGTL) model combining DGPs with pre-trained CNN architectures; and to assess the performance of the proposed models against standard machine learning baselines. A hybrid kernel, formed by combining rational quadratic and squared exponential kernels, was also introduced to enhance model accuracy and expressiveness. Experimental results showed that the DGCNN achieved an accuracy of 90.1%, outperforming models using standard kernels. The DGTL model, when integrated with MobileNetV2 and the hybrid kernel, achieved 90.11% accuracy, surpassing other configurations. Although computational constraints were encountered due to limited hardware resources, the proposed models demonstrated strong predictive performance while maintaining the probabilistic interpretability of Gaussian Processes. These findings contribute to the advancement of intelligent and reliable tools for precision agriculture. Key words: Cassava Disease Classification, Deep Gaussian Processes (DGPs), Convolutional Neural Networks (CNNs), Transfer Learning, Precision Agriculture.