PepperMate: An AI-Enabled Mobile Application for Optimizing Ceylon Black Pepper Cultivation

Sri Lanka is renowned for its high-quality "Ceylon black pepper," celebrated for its distinctive characteristics. As the second most exported spice in the country, it plays a significant role in the spice trade. However, black pepper farmers face persistent challenges, including accurate disease identification, market volatility, and limited access to knowledge, which adversely impact their livelihoods and the nation's export economy. To address these challenges, this study leverages Artificial Intelligence (AI) and mobile application technologies to develop "PepperMate," an integrated mobile application. The app offers features such as precise disease classification, price prediction, and a Sinhala language-based retrieval-augmented generation chatbot. During extensive field visits to black pepper plantations, over 1,000 black pepper leaf samples were collected for disease classification. This dataset was augmented to 12,000 samples during the model training phase to improve accuracy. Time series models were also trained on black pepper pricing data from 2016 to the present, with augmentation techniques applied to enhance predictive performance. Additionally, online resources, including PDF files from the Department of Export Agriculture website, were utilized to create the knowledge base for the chatbot. Among the disease classification models tested—Custom Convolutional Neural Network (CNN), AlexNet, Inception v3, MobileNet v2, ResNet-50, and EfficientNet B5—MobileNet v2 and EfficientNet B5 achieved the highest test accuracies, at 95.57% and 99.42%, respectively. For time series-based price prediction, the Bidirectional Long Short-Term Memory (Bidirectional LSTM) network and 1D Convolutional Neural Network (1D CNN) delivered the best results, with accuracies of 98.00% and 95.14%, respectively. In the chatbot development process, popular Large Language Models (LLMs) like GPT-4, claude-3-sonnet-20240229, llama-2-7b, and Mistral-7B-v0.1 were evaluated alongside embedding models such as sentence-transformers/all-mpnet-base-v2 and Text-embedding-ada-002. The combinations of claude-3-sonnet-20240229 with Sentence sentence-transformers/all-mpnet-base-v2 and GPT-4, with Text-embedding-ada-002, demonstrated superior linguistic understanding and accuracy, whereas Mistral-7B-v0.1 and llama-2-7b struggled to generate coherent responses. Finally, these AI-enabled services were seamlessly integrated into the "PepperMate" mobile application, providing farmers with real-time disease diagnosis, actionable solutions, price predictions, and context-specific advice to enhance their productivity and profitability making this the first-of-its-kind solution in the context of black pepper farming.