Abstract
Yield prediction is a crucial aspect of crop management in the fruit farming industry. In recent years, a variety of methods have been increasingly utilized to address this challenge, in order to enhance the technical capabilities both in the hardware and software domain.
To tackle this challenge, this thesis, as a part of research project aimed at crop management called “Earlier and more precise yield forecasts in fruit production – development of the digital tree” (translated from Norwegian), we designed a pipeline with separate sections called DRCAP. the sections cover Detection, Classification, Regression, and Pixel Processing of apple tree images taken by RGB sensors. The primary objective of this pipeline is to identify apples at the first stage and then send to Analysis part in order to make a yield prediction.
We implemented YOLOv7 for apple detection and used a custom convolutional model as the classifier and a custom regressor for the regression task. Indeed, an image processing technique has been applied as an auxiliary method alongside a classifier and regressor. We implemented the coding of this project partially in the Google Collaboratory service and partially in the local machine with Jupiter Notebook.
The main data set for this study comes from an apple orchard in Gvarv, Telemark (Norway), taken over a growth season in 2021 and 2022 which only 2021 dataset was used in this study.