Date of Award

27-2-2024

Document Type

Thesis

School

School of Electrical & Electroncis Engineering

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

Dr.P.Raja

Keywords

Engineering and Technology, Computer Science, Computer Science Artificial Intelligence, Deep Learning for Agricultural Application, AI based intelligent system design for Fruit Yield Estimation

Abstract

Agriculture contributes more resources for developing sustainable economic growth of the nation. Precision agriculture employs advanced techniques (machine learning and deep learning) for developing the intelligent systems of various agricultural applications. Among various agricultural tasks, yield estimation of crops plays a vital role in decision-making such as harvesting, marketing, cultivation practices, etc. Traditionally yield estimation is performed manually which has major drawbacks i.e., needs experts opinion, time-consuming and it is a challenging task for big orchards. To overcome these issues, an intelligent yield estimation model using neural network-based systems is required.

Some of the literature works have been explored for fruit yield estimation using intelligent techniques namely, deep learning-based semantic segmentation architectures. But, the analysis of a customized counting model which gives better localization for mapping the fruit yield including challenging situations like partially occluded and overlapped conditions is yet to be explored. Hence, the objective of the work is to develop an intelligent fruit yield estimation for selected orchards (tomato and mango) using deep learning-based semantic segmentation architectures.

In the first phase of work, tomato yield estimation was performed using three deep learning-based semantic segmentation architectures such as U-Net, SegNet with VGG16 and SegNet with VGG19. Training was done on the dataset of 672 tomato images and testing was done on the real-time field data of 65 tomato images. The test results revealed the highest precision, recall and F1-score values of 89.7%, 72.55% and 80.22%, respectively for the SegNet with VGG19 architecture among the three architectures compared.

Finally, the segmented fruits were counted using the contour detection method. Then, the final yield in kilograms was estimated for the chosen tomato field. The error percentage between actual and predicted weight is 4.8%. A user-friendly graphical user interface was developed for estimating the tomato yield. However, as VGG19 is used as the backbone network for SegNet, execution time and memory consumption are the hurdles for real-time implementation.

To overcome these issues, suitable advanced deep learning-based semantic segmentation architecture can be employed. Hence, in the second phase of the work, MangoYieldNet for intelligent counting of mangoes using DeepLabv3+ was proposed which extracts the features from the mango images by employing atrous spatial pyramid pooling and an effective decoder module for better localization of mangoes. The mango images were taken during the daytime on both sides of the trees.

After the sampling process, 556 images (from 278 trees) were captured and annotated. Using image augmentation techniques (reflection, rotation and translation) a dataset of 1152 images was developed; it was split in an 80:20 ratio for training and validation, respectively. The training was initiated using random weight parameters and the hyper-parameters were optimized using stochastic gradient descent by minimizing the errors at the epoch of 50. For testing the trained architecture, 30 new images were captured from the mango orchard. The test dataset revealed the highest mean accuracy of 96.5% and a mean intersection over union of 95.97% for the architecture of DeepLabv3+ with ResNet18 and better object localization among other architectures (i.e., DeepLabv3+ with MobileNetv2, DeepLabv3+ with Xception, U-Net and SegNet with VGG19) compared.

Finally, the segmented fruits were counted using the circle Hough transform method. The mango count was compared with the manual count obtained from the farmers and provided the regression coefficient of 0.98. Then, the final yield in kilograms was estimated for the sampled trees in the orchard. The error percentage between actual and predicted weight is 4.99%. A user-friendly graphical user interface was also developed for estimating the mango yield. The future scope includes extending the work for estimating the yield for other fruits and deploying the model to robotic harvesting systems.

Download

Share

COinS