Institutional Repository [SANDBOX]
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Identifying sunlit leaves using Convolutional Neural Networks: an expert system for measuring the crop water stress index of pistachio trees

Pantelidakis Minas, Panagopoulos Athanasios Aris, Mykoniatis Konstantinos, Ashkan Shawn, Cherupillil Eravi Rajeswari, Pamula Vishnu, Verduzco Enrique, Babich Oleksandr, Panagopoulos Orestis P., Chalkiadakis Georgios

Full record


URI: http://purl.tuc.gr/dl/dias/88140DB0-14F3-407E-ABF4-9A5A7D6AA8BF
Year 2022
Type of Item Peer-Reviewed Journal Publication
License
Details
Bibliographic Citation M. Pantelidakis, A. A. Panagopoulos, K. Mykoniatis, S. Ashkan, R. Cherupillil Eravi, V. Pamula, E. Verduzco, O. Babich, O. P. Panagopoulos, and G. Chalkiadakis, “Identifying sunlit leaves using Convolutional Neural Networks: an expert system for measuring the crop water stress index of pistachio trees,” Expert Syst. Appl., vol. 209, Dec. 2022, doi: 10.1016/j.eswa.2022.118326. https://doi.org/10.1016/j.eswa.2022.118326
Appears in Collections

Summary

Precision irrigation has been proposed as a key means towards a sustainable agriculture future. Many such techniques use infrared and/or visible-spectrum images to measure the canopy temperature. This, in turn, is utilized to calculate the crop water stress index (CWSI). Sunlit leaves are likely to get under stress sooner and to a higher extent compared to non-sunlit ones. As such, using only sunlit leaves for measuring canopy temperature can be beneficial for precision irrigation purposes. However, existing work generally does not separate sunlit from non-sunlit leaves. The works that do attempt to do so generally use empirical techniques and do not come with thorough evaluation results. In this work, we propose a novel generic method to identify sunlit leaves with high accuracy and high precision. The backbone of our approach is a convolutional neural network (CNN) technique for segmenting sunlit leaves in visible-spectrum images. This is the first work to propose a strictly supervised learning technique for sunlit-leaf segmentation. We release the first dataset for sunlit-leaf segmentation for pistachio trees. We evaluate four different CNN architectures for this task, namely FRRN-A, FC-DenseNet103, ResNet101-DeepLabV3, and SegNet. We show that CNNs can improve the accuracy and precision of sunlit-leaf segmentation by over 100% and 350%, respectively, when compared to the state of the art. Furthermore, we propose an expert system for measuring the CWSI of pistachio trees that incorporates CNN-based sunlit-leaf segmentation. We introduce the expert system to the public as a free-of-charge web-based tool.

Services

Statistics