Real-time mulberry fruit ripeness recognition for selective harvesting using edge-deployable object detection: a single-orchard case study

Introduction Mulberry is a seasonal cash crop with a short ripening window, where delayed harvesting readily leads to fruit deterioration and economic losses. Real-time ripeness recognition is therefore important for selective harvesting. Existing detection models face challenges including excessive model parameters, low discriminability between immature fruits and background foliage, and severe occlusion caused by clustered fruits. Methods We present Mulberry-Net, a lightweight three-class mulberry ripeness detector that adapts and systematically evaluates three established components: Partial Convolution integrated into the C2f bottleneck to form the Cross-Stage Feature Compression and Cooperation (CSFCC) module; Mixed Local Channel Attention (MLCA) placed after multi-scale feature concatenation in the neck; and an occlusion-aware training strategy based on Focal-DIoU loss. The components target computational cost, fruit-background discrimination, and localisation under occlusion, respectively. Results Mulberry-Net achieves an [email protected] of 97.80% with 2.41M parameters, improving on the YOLOv8n baseline by 5.10 percentage points while reducing parameters by 19.9%. The largest class-specific gain occurs for green fruits (+8.60 percentage points). Five-fold tree-level cross-validation yields 96.02 +/- 0.74% [email protected]. Three-class ripeness accuracy on correctly localised fruits reaches 96.83%, with residual grading errors concentrated at the color-turning/mature boundary and missed detections concentrated in the green stage. A controlled comparison against four alternative lightweight convolution designs at the same bottleneck position shows that the adopted configuration attains the highest accuracy at comparable computational cost. Discussion The results support the feasibility of lightweight three-class mulberry ripeness detection for selective harvesting under the conditions studied. Because all data originate from a single orchard, cultivar, and acquisition device, the reported figures should be interpreted as single-site feasibility evidence rather than established generalisation across cultivars and growing regions.

Authors

Institutions

Publication Details

Journal
Frontiers in Plant Science
Published
2026-11-16
DOI
https://doi.org/10.3389/fpls.2026.1942552
Primary Topic
Smart Agriculture and AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Real-time mulberry fruit ripeness recognition for selective harvesting using edge-deployable object detection: a single-orchard case study

Zhongbin Chan, Bojian Yu, Xinwei Zhang, Xin Zhao
4 links
Frontiers in Plant Science
Smart Agriculture and AI
article

Real-time mulberry fruit ripeness recognition for selective harvesting using edge-deployable object detection: a single-orchard case study

Zhongbin Chan, Bojian Yu, Xinwei Zhang, Xin Zhao
article en

Abstract

Introduction Mulberry is a seasonal cash crop with a short ripening window, where delayed harvesting readily leads to fruit deterioration and economic losses. Real-time ripeness recognition is therefore important for selective harvesting. Existing detection models face challenges including excessive model parameters, low discriminability between immature fruits and background foliage, and severe occlusion caused by clustered fruits. Methods We present Mulberry-Net, a lightweight three-class mulberry ripeness detector that adapts and systematically evaluates three established components: Partial Convolution integrated into the C2f bottleneck to form the Cross-Stage Feature Compression and Cooperation (CSFCC) module; Mixed Local Channel Attention (MLCA) placed after multi-scale feature concatenation in the neck; and an occlusion-aware training strategy based on Focal-DIoU loss. The components target computational cost, fruit-background discrimination, and localisation under occlusion, respectively. Results Mulberry-Net achieves an [email protected] of 97.80% with 2.41M parameters, improving on the YOLOv8n baseline by 5.10 percentage points while reducing parameters by 19.9%. The largest class-specific gain occurs for green fruits (+8.60 percentage points). Five-fold tree-level cross-validation yields 96.02 +/- 0.74% [email protected]. Three-class ripeness accuracy on correctly localised fruits reaches 96.83%, with residual grading errors concentrated at the color-turning/mature boundary and missed detections concentrated in the green stage. A controlled comparison against four alternative lightweight convolution designs at the same bottleneck position shows that the adopted configuration attains the highest accuracy at comparable computational cost. Discussion The results support the feasibility of lightweight three-class mulberry ripeness detection for selective harvesting under the conditions studied. Because all data originate from a single orchard, cultivar, and acquisition device, the reported figures should be interpreted as single-site feasibility evidence rather than established generalisation across cultivars and growing regions.

Frontiers in Plant ScienceVol. 17
Zhejiang Sci-Tech University (CN), Zhejiang Ocean University (CN), Shanghai Zhongqiao Vocational and Technical University (CN)
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