Image to grape yield estimation in the vineyard using deep learning
LE3 .A278 2018
2018
Silver, Danny
Acadia University
Master of Science
Masters
Computer Science
Agricultural yield estimation from natural images is an important, yet challenging problem to which machine learning can be applied. There is value in having better methods of yield estimation based on data that can be captured with inexpensive technology in the field, such as a smartphone. The research uses convolution neural networks (CNNs) to develop models that can estimate the weight of grapes on a vine using an image. Convolutional Neural Networks have advanced the state of the art in many machine learning applications such as computer vision, speech recognition and natural language processing. We investigate five approaches of generating yield estimate from images by using convolutional neural networks. The results yield solid evidence that the proposed model bears strong potential to be superior compared to the existing solutions for yield estimation. We discovered that a trained CNN model developed using weights from a pre-trained density map model can be successfully used to generate yield estimate from images. We demonstrate that the CNNs perform better in predicting yield estimate when we specify the grape locations in the images by using density maps. By applying finetuning and using a density map, the proposed model achieved an error rate as low as 12%. The system is trained and tested on the images of grapes in a vineyard. The system is capable of accurately estimating the yield from images that have grapes of almost the same size, color, and brightness as grapes in images used to train the model. Using the concept of deep learning, I have demonstrated how a deep CNN module can be expected to be used in near real-time conditions for commercial deployment with minimal hardware requirements.
The author retains copyright in this thesis. Any substantial copying or any other actions that exceed fair dealing or other exceptions in the Copyright Act require the permission of the author.
https://scholar.acadiau.ca/islandora/object/theses:3193