Tomato-plant Sunlit-leaf Segmentation Using Convolutional Neural Networks: Supporting Crop Water Stress Index Measurements

Jacob Hampton, Aris Athanasios Panagopoulos, Minas Pantelidakis, Eric Kor, Konstantinos Mykoniatis, Orestis P. Panagopoulos, Shawn Ashkan

Keywords: precision irrigation, convolutional neural networks, sunlit-leaf segmentation

TL;DR: Proposing tomato-plant sunlit-leaf segmentation using convolutional neural networks to support crop water stress index measurements

Abstract: Many precision irrigation approaches rely on calculating the crop water stress index (CWSI) using measurements of the sunlit-leaf canopy temperature ({\em sunlit-leaf} $T_c$). To this end, they typically employ machine learning and/or statistical techniques along with visible-spectrum and/or thermal imagery to identify the {\em sunlit-leaf} region and its temperature. The precise segmentation of sunlit leaves is of utmost importance; considering non-sunlit leaves while calculating CWSI, can lead to inaccurate and inefficient under-irrigation or over-irrigation. Recent work demonstrated that convolutional neural networks (CNN) can support highly precise sunlit leaf segmentation in pistachio trees. The same work introduced a complete methodology for the estimation of CWSI in pistachio trees and released a corresponding free-of-charge web tool: CIWA. However, the generality of the approach to other crops is not discussed. Here, we extend the CIWA methodology to tomato plants. We discuss the challenges of extending the CIWA approach to shorter plants and release the first annotated dataset for sunlit-leaf segmentation in tomato plants. We consider three CNN architectures for this task: FRRN-A, FC-DenseNet103 and ResNet101-DeepLabV3, and show that FC-DenseNet103 is the most suitable for the task. Based on these results, we introduce an extended CIWA release to enable tomato CWSI measurements.