LISNet: an instance segmentation method for lettuce canopy width and perimeter detection

Introduction Canopy width and perimeter are important indicators for assessing lettuce growth status and supporting yield-related phenotypic analysis. Traditional manual measurement methods for canopy width and perimeter are inefficient and prone to considerable measurement errors. Methods In this study, we propose LISNet (Lettuce Instance Segmentation Network), an instance segmentation framework designed to estimate lettuce canopy width and perimeter simultaneously. Based on the Mask R-CNN architecture, LISNet replaces the ResNet-101 backbone of Mask RCNN with Swin Transformer and incorporates a Coordinate Attention (CA) module to enhance feature extraction and boundary segmentation. Additionally, we propose an improved Feature Pyramid Network (FPN) structure that integrates the Residual Information Enhancement (RIE) module and the Adaptive Spatial Fusion (ASF) module. This modification enhanced the fusion of multi-scale features and improved the model’s performance at different scales. The conversion factor between pixel values and real-world physical values (in centimetres) is calculated by extracting the contours of the reference object. The canopy width and perimeter of each lettuce are then computed based on the lettuce mask output by the model’s segmentation. Results Experimental results demonstrated that over three independent training runs, LISNet achieved a mean instance segmentation AP of 87.4% (± 3.5%), corresponding to a 6.5 percentage-point improvement over the baseline Mask R-CNN framework. The model reached an inference speed of 28.9 FPS, indicating that it can support efficient greenhouse lettuce image analysis while maintaining high segmentation accuracy. Compared with existing instance segmentation models such as YOLACT, Hybrid Task Cascade (HTC), Cascade Mask R-CNN, and YOLOv8-seg, as well as the prompt-based Segment Anything Model (SAM), LISNet achieved higher average precision, enabling effective detection of each lettuce plant instance in an image and the generation of high-quality segmentation masks for them. Furthermore, the model-based canopy width and perimeter estimates showed good agreement with manual measurements, with R 2 values of 0.946 and 0.938, MRE values of 0.05 and 0.04, and RMSE values of 2.22 cm and 6.91 cm, respectively. Discussion The results indicate that the proposed method has potential for accurate lettuce segmentation and phenotype estimation in greenhouse environments.

Authors

Institutions

Publication Details

Journal
Frontiers in Plant Science
Published
2026-11-16
DOI
https://doi.org/10.3389/fpls.2026.1932886
Primary Topic
Smart Agriculture and AI
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

LISNet: an instance segmentation method for lettuce canopy width and perimeter detection

Shuo Fan, Ke Yu, Hanbing Deng
5 links
Frontiers in Plant Science
Smart Agriculture and AI
article

LISNet: an instance segmentation method for lettuce canopy width and perimeter detection

Shuo Fan, Ke Yu, Hanbing Deng
article en

Abstract

Introduction Canopy width and perimeter are important indicators for assessing lettuce growth status and supporting yield-related phenotypic analysis. Traditional manual measurement methods for canopy width and perimeter are inefficient and prone to considerable measurement errors. Methods In this study, we propose LISNet (Lettuce Instance Segmentation Network), an instance segmentation framework designed to estimate lettuce canopy width and perimeter simultaneously. Based on the Mask R-CNN architecture, LISNet replaces the ResNet-101 backbone of Mask RCNN with Swin Transformer and incorporates a Coordinate Attention (CA) module to enhance feature extraction and boundary segmentation. Additionally, we propose an improved Feature Pyramid Network (FPN) structure that integrates the Residual Information Enhancement (RIE) module and the Adaptive Spatial Fusion (ASF) module. This modification enhanced the fusion of multi-scale features and improved the model’s performance at different scales. The conversion factor between pixel values and real-world physical values (in centimetres) is calculated by extracting the contours of the reference object. The canopy width and perimeter of each lettuce are then computed based on the lettuce mask output by the model’s segmentation. Results Experimental results demonstrated that over three independent training runs, LISNet achieved a mean instance segmentation AP of 87.4% (± 3.5%), corresponding to a 6.5 percentage-point improvement over the baseline Mask R-CNN framework. The model reached an inference speed of 28.9 FPS, indicating that it can support efficient greenhouse lettuce image analysis while maintaining high segmentation accuracy. Compared with existing instance segmentation models such as YOLACT, Hybrid Task Cascade (HTC), Cascade Mask R-CNN, and YOLOv8-seg, as well as the prompt-based Segment Anything Model (SAM), LISNet achieved higher average precision, enabling effective detection of each lettuce plant instance in an image and the generation of high-quality segmentation masks for them. Furthermore, the model-based canopy width and perimeter estimates showed good agreement with manual measurements, with R 2 values of 0.946 and 0.938, MRE values of 0.05 and 0.04, and RMSE values of 2.22 cm and 6.91 cm, respectively. Discussion The results indicate that the proposed method has potential for accurate lettuce segmentation and phenotype estimation in greenhouse environments.

Frontiers in Plant ScienceVol. 17
Shenyang Agricultural University (CN)
Beijing Academy of Agricultural and Forestry Sciences, National Key Research and Development Program of China
Zero hunger
Smart Agriculture and AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon