Deep learning-based instance segmentation architectures in agriculture: A review of the scopes and challenges

Abstract

Deep learning (DL) based instance segmentation has attracted a growing research interest in the scientific community to tackle precision agriculture problems over the past few years. However, accurate crop detection and localization in complex environments pose a significant challenge. Instance segmentation is considered as a promising DL technique that expands on object detection to perform pixel-wise image instance segmentation and address pattern recognition problems efficiently. In this review, we identify 77 relevant studies on DL-based instance segmentation implementations in agriculture and thoroughly investigate them from the following perspectives: i) the specific architecture employed; ii) the data type and availability, the data annotation process and the data pre-processing techniques; iii) the performance metrics used; and iv) hardware, inference time and GPU requirements. Our findings indicate that crop detection (48 papers) constitutes a fundamental task in a DL-based instance segmentation pipeline to enable crop growth monitoring (19 papers) and plant health analysis (10 papers). Among them, 6 papers reported robotic manipulation and other related automation tasks. Based on our findings we can conclude that there is a significant trend towards two-stage DL-based instance segmentation models i.e., Mask R-CNN baseline and customized architectures (69 papers). Limitations and challenges, such as availability of benchmark crop datasets, open-source codes for semi-automatic annotation tools, technical requirements and opportunities for future research are discussed.