Smart Harvesting System with Autonomous Classification of Date Fruit Type and Maturity Stages Using Deep Transfer Learning Zainab Abuowda, Shorouk Ramadan, Nour Salam, Jawad Yousaf, Abdalla Gad, Mohammed Ghazal Electrical, Computer, and Biomedical Engineering Department, Abu Dhabi University, Abu Dhabi, UAE Abstract Proposed System Results Date Type Classification: • Data set contained 11300 images (30% for validation, 15 % for testing) for each type • Network settings: number of iterations:13800, epochs: 200 • Test Accuracy: 98% (Resnet-50), 97.3% (AlexNet) and 98.9% (VGG16). • We proposed an effective deep transfer learningbased system for automatically classifying: ➢ 5 distinct date types on bunches (Naboot saif, Khalas, Barhi, Meneifi, and Sullaj). ➢ 5 different maturity stages (Immature, Khalal, Khalal w Rutab, Pre Tamar, and Tamar). • The study’s findings will be used to design an intelligent robotic harvesting mechanism for efficiently monitoring and harvesting date fruits in a natural environment. • We used a transfer learning approach on pre-trained deep convolutions neural networks. • AlexNet, Resenet-50, and VGG-16 networks were trained on the available dataset for the two classification tasks. Barhi Kahals Meneifi Naboot Sullaj Precision Saif Barhi 1122 4 3 5 2 98.7% Kahals 0 730 0 0 0 100% Meneifi 0 2 513 1 0 99.4% Naboot Saif Sullaj 4 0 2 712 0 98.89% 3 0 24 2 1733 98.35% 94.65% 98.89% 99.7% 98.89% Problem Statement Recall 99.38% 99.19% Table 1: Confusion Matrix for Date Type Classification using VGG-16 Network. • Limited number of high-quality cultivars. • Lack of proper research on pests and diseases. Maturity Stage Classification: • Data set contained 11300 images (30% for validation, 15 % for testing) for each type • Network settings: number of iterations:13800, epochs: 200 • Test Accuracy: 96.61% (Resnet-50), 97.87% (AlexNet) and 98.17% (VGG-16). • Inadequate harvesting as per modern technologies. • Lack of experienced and educated employees. Dataset Examples Immature Khalal Khalal w’ Rutab Actual pictures of Barhi date Actual pictures of Sullaj date Pre Tamar Tamar Precision Immature 1531 3 3 1 2 99.42% Khalal 0 476 7 0 0 98.55% Khalal w’ Rutab 1 45 2043 15 1 97.1% Pre Tamar 0 0 2 419 4 98.59% Tamar Recall 0 99.94% 0 90.84 % 0 99.42% 1 96.1% 87 92.5% 98.86% 98.17% Table 2: Confusion Matrix for Maturity Stage Classification using VGG-16 Network. References • H. Altaheri, M. Alsulaiman, M. Faisal, and G. Muhammed, “Date fruit dataset for automated harvesting and visual yield estimation,” 2019. Available: https://dx.doi.org/10.21227/x46j-sk98 • S. Wang, X. Xia, L. Ye, and B. Yang, “Automatic detection and classification of steel surface defect using deep convolutional neural networks. metals, 11 (3), 1–23,” 2021. Conclusion • A vision system is proposed for the classification of 5 date types during their 5 maturity stages using deep transfer learning techniques. • Future work: ➢ Improving the ResNet-50 model testing accuracy. ➢ Expanding the data set to include videos along with images