Candidate: Islam Mohamed Mahmoud Nasr
Date: August 16, 2023
Time: 10:00am
Location: E5 4047
Supervisor(s): Fakhri Karray
Abstract
This seminar discusses the improvement and evaluation of deep learning models for forecasting fresh produce (FP) yields and prices. The proposed model, DFNNGRU-ADGRU ENS, combines Attention Deep Feedforward Neural Network with Gated Recurrent Units (ADGRU) and Deep Feedforward Neural Network with embedded GRU units (DFNNGRU). The station-based ensemble model outperforms existing models by 24% in yield forecasting and 37.5% in price forecasting.
For satellite-based models, a preprocessing technique based on averaging is proposed, which outperforms histogram-based approaches by 20%. The DFNNGRU-ADGRU model also outperforms existing models by 12.5%.
Interpolation techniques are used to estimate missing values in satellite images, and the most effective technique is found to be Cubic Spline interpolation. Adding the Normalized Difference Vegetation Index (NDVI) as an input parameter enhances yield forecasting by 12.5%.
A novel transfer learning framework is proposed to improve generalizability to other FPs by using FP similarity, clustering, and TL techniques. The application is implemented with a user-friendly interface, and metamorphic testing is used to verify the models and application. The results indicate the success of the proposed models and their application.