Source-linked AI summary
A Lightweight Phenology-Aware YOLOv5 Framework for Tomato Growth Stage Detection in Resource-Constrained Bhutanese Greenhouse Environments
Sherab Gocha, Sou Nobukawa
TL;DR
Tomato growth-stage detection in Bhutanese greenhouses faces variable environmental conditions, limited infrastructure, and a lack of local datasets. The paper adapts YOLOv5 with phenology-aware and channel-attention refinements. It reports higher precision than YOLOv5s, alongside lower mAP@50.
Problem
Bhutan lacks documented computer-vision research on tomato growth-stage detection and locally curated phenological datasets, amid variable greenhouse conditions and limited infrastructure.
Source: §I · p. 2Method
The paper develops Pheno-Lite + ECA, a lightweight YOLOv5 framework with targeted backbone refinements for tomato-stage recognition.
Source: §I · p. 2, F. Proposed Pheno-Lite + ECA… · p. 3, 1) Pheno-Lite Module · p. 3Results
The proposed model achieves 90.6% precision, 88.8% recall, and 92.6% mAP@50; YOLOv5s records 89.8% precision and 93.6% mAP@50.
Source: A. Performance of the Proposed… · p. 4, §IV · p. 7Takeaways & Limitations
The efficiency-aware framework targets greenhouse deployment where computational resources, power availability, and automation infrastructure are limited.
Source: §IV · p. 6
Abstract
from arXiv · showhide
Accurate detection of tomato growth stages is essential for stage-specific greenhouse management and precision agriculture. In Bhutan, greenhouse cultivation is affected by altitude variability, large diurnal temperature fluctuations, diffuse illumination, limited automation, and a scarcity of locally annotated datasets, limiting the applicability of conventional deep learning models. This work proposes Pheno-Lite + Efficient Channel Attention (ECA), a lightweight, phenology-aware object detection architecture derived from Ultralytics YOLOv5 for tomato growth stage recognition. A balanced dataset of 2,464 annotated images was constructed from locally collected greenhouse images in Bhutan and publicly available tomato images, with augmentation designed to simulate local greenhouse conditions. The dataset includes vegetative (820), flowering (824), fruiting (820), and background (26) samples. The proposed architecture introduces two customized backbone modules: C3 PhenoLite, which enhances spatial and texture feature extraction using depthwise residual refinement, and C3 ECA, which strengthens inter-channel feature interactions through efficient channel attention. The proposed model achieves 90.6% precision, 88.8% recall, and 92.6% mAP@50, with 4.0 million parameters and 10.9 GFLOPs at 640 x 640 resolution. These results demonstrate its potential for real-time and climate-resilient greenhouse deployment in Bhutan.
I. INTRODUCTION
The paper addresses the lack of Bhutan-focused tomato growth-stage detection research and locally curated phenology datasets amid variable greenhouse conditions and limited infrastructure. It develops a lightweight detection framework and dataset grounded in Bhutanese greenhouse conditions.
I. INTRODUCTION
- Bhutan’s variable climate, diffuse illumination, limited automation, and scarce local datasets motivate a context-aware, lightweight tomato-stage detector.The introduction says no documented research addresses computer-vision tomato growth-stage detection in Bhutan.Source: §I · p. 2
A. Dataset Acquisition and Development
- The Bhutan-specific images were captured in greenhouses with smartphones from frontal, lateral, and oblique views and at different times of day.Public images were also selected for visual similarity in morphology, background complexity, and lighting.Source: A. Dataset Acquisition and Development · p. 2
- The final dataset comprised 2,464 annotated images, selected to enhance diversity while maintaining contextual relevance, with balanced class representation and improved background discrimination.Source: A. Dataset Acquisition and Development · p. 2
B. Dataset Organization and Annotation
- YOLO-format bounding boxes label leaves-only, open-flower, and discernible-fruit stages, with an 82%–12%–6% train-validation-test split.The split yielded 2,030 training, 290 validation, and 144 testing images.Source: B. Dataset Organization and Annotation · p. 2
C. Data Preprocessing and Augmentation
Images are resized and augmented to represent greenhouse variability, while YOLOv5s provides the lightweight baseline for training and evaluation.
C. Data Preprocessing and Augmentation
- All images are resized to 640 × 640 pixels using stretch-based resizing to keep dimensions and bounding-box alignment consistent.The paper says this supports stable training optimization and consistent evaluation.Source: C. Data Preprocessing and Augmentation · p. 2
- Training-only augmentation uses flips, rotations up to ±10°, grayscale on 9% of samples, and controlled hue, saturation, and brightness changes.These transformations are intended to simulate diffuse light, partial shading, and viewpoint variation while preserving phenological integrity.Source: C. Data Preprocessing and Augmentation · pp. 2–3
D. Baseline Model Selection and Training Setup
- YOLOv5s supplies the lightweight baseline, and pretrained weights and identical data settings are used for fair model comparisons.The baseline combines a convolutional backbone, FPN–PAN neck, and multi-scale detection head.Source: D. Baseline Model Selection and… · p. 3
E. Evaluation Metrics and Experimental Protocol
The evaluation uses standard detection metrics and controlled data partitions. The proposed architecture refines YOLOv5 with targeted backbone changes for resource-limited greenhouse deployment.
E. Evaluation Metrics and Experimental Protocol
- Precision, recall, and mAP@50 assess correct positive detections, object recovery, and aggregate precision–recall performance across growth stages.mAP@50 uses an intersection-over-union threshold of 0.5.Source: E. Evaluation Metrics and Experimental… · p. 3
- Fixed partitions, preprocessing, augmentation, and input resolutions support controlled comparisons, with test samples excluded from model optimization and tuning.All models are evaluated under the same conditions and metrics.Source: E. Evaluation Metrics and Experimental… · p. 3
F. Proposed Pheno-Lite + ECA Architecture
- Pheno-Lite + ECA adds targeted YOLOv5 backbone refinements to improve phenological-stage discrimination under computational and deployment constraints.The design prioritizes efficient feature representation over exhaustive architectural expansion.Source: F. Proposed Pheno-Lite + ECA… · p. 3
1) Pheno-Lite Module:
The architecture pairs a depthwise residual refinement module with efficient channel attention in selected YOLOv5 C3 blocks. These modules target fine-grained phenological cues and channel-wise feature refinement.
1) Pheno-Lite Module:
- Pheno-Lite adds depthwise convolution, batch normalization, and SiLU in a scaled residual branch to a YOLOv5s C3 block.The branch refines localized spatial responses while preserving the main feature hierarchy and gradient stability, targeting cues such as leaf arrangement, flower emergence, and early fruit development.Source: 1) Pheno-Lite Module · p. 3
2) Efficient Channel Attention (ECA) Module:
- ECA uses global average pooling and lightweight one-dimensional convolution to recalibrate channels without dimensionality reduction.The module is incorporated residually to emphasize phenology-relevant channels and suppress redundant activations.Source: 2) Efficient Channel Attention (ECA)… · p. 4
3) Neck and Detection Head:
The architecture retains YOLOv5s’s standard neck and detection head, using multi-scale feature fusion and three detection scales. Its computational-efficiency discussion defines GFLOPs as a per-image forward-pass measure and specifies the detection-layer feature-map resolutions.
3) Neck and Detection Head:
- The unchanged YOLOv5s neck uses PANet-style top-down and bottom-up fusion, while its head predicts boxes, objectness, and growth-stage probabilities at three scales.The scales correspond to small, medium, and large receptive fields.Source: 3) Neck and Detection Head · p. 4
4) Computational Efficiency:
- GFLOPs measure the floating-point operations needed to process one image in a forward pass, providing a hardware-independent measure of architectural computational cost.The passage introduces a CNN GFLOPs formulation but does not provide a complete readable equation here.Source: 4) Computational Efficiency · p. 4
- For 640 × 640 input, detection feature maps are 80 × 80, 40 × 40, and 20 × 20 at P3, P4, and P5, with strides 8, 16, and 32.The passage identifies these layers as the detection feature levels.Source: 4) Computational Efficiency · p. 4
III. RESULTS
On the held-out test set, Pheno-Lite + ECA achieved 90.6% precision, 88.8% recall, and 92.6% mAP@50 for tomato growth-stage detection. The paper describes its performance as balanced across the three growth stages and links the precision outcome to the model’s refinement and attention components.
A. Performance of the Proposed Pheno-Lite + ECA Model
- 90.6% precision, 88.8% recall, and 92.6% mAP@50 were achieved on the held-out test set across vegetative, flowering, and fruiting stages.The authors attribute improved precision to localized spatial refinement from Pheno-Lite and adaptive channel recalibration from ECA.Source: A. Performance of the Proposed… · p. 4
- The authors characterize the model’s precision–recall balance as improving feature discrimination rather than expanding detection coverage.They describe this balance as desirable for stage-specific monitoring.Source: A. Performance of the Proposed… · p. 4
B. Quantitative Performance Analysis
- Table II presents the model’s precision, recall, and mAP@50, which the authors describe as consistent detection accuracy across growth stages.The authors state that the mAP@50 score indicates accurate box localization and stable classification behavior.Source: B. Quantitative Performance Analysis · p. 4
C. Precision–Recall and Confidence-Based Analysis
The validation analysis describes a favorable precision–recall balance and reliable predictions at higher confidence thresholds. It also uses label and prediction visualizations to assess annotations and model behavior on validation images.
C. Precision–Recall and Confidence-Based Analysis
- The validation precision–recall curve shows stable performance across a broad recall range, and the precision–confidence curve indicates reliable predictions at higher thresholds.The paper relates increasing recall with maintained high precision to robustness against false positives in cluttered scenes.Source: C. Precision–Recall and… · pp. 4–5, Figure 2 · p. 5
D. Validation Batch Labels and Predictions
- Fig. 3 shows validation-image ground-truth boxes and class labels, used to check annotation accuracy, spatial consistency, and class representation.The caption identifies these as examples of labeled tomato growth stages under greenhouse conditions.Source: D. Validation Batch Labels and… · p. 5, Figure 3 · p. 5
- Fig. 4 shows predicted boxes, class assignments, and confidence scores on the same validation images, supporting qualitative checks of localization, classification, and detection errors.The authors present this comparison as complementary to quantitative performance metrics.Source: D. Validation Batch Labels and… · p. 5
E. Model Complexity and Computational Efficiency
The proposed model has a compact computational footprint, while the YOLOv5s baseline provides a strong detection benchmark with differing performance across growth stages.
E. Model Complexity and Computational Efficiency
- The proposed model uses approximately 4.0 million parameters and 10.9 GFLOPs per 640 × 640 image, with lower computational complexity than YOLOv5s.The passage describes this footprint as supporting efficient real-time inference in resource-constrained greenhouses.Source: E. Model Complexity and Computational… · p. 5
F. Detection performance of baseline yolov5s
- Validation predictions show bounding boxes, class assignments, and confidence scores for qualitative assessment of localization and classification.The passage describes visual comparison with labels as a complement to quantitative metrics.Source: Figure 4 · p. 6
- YOLOv5s achieved 89.8% precision, 88.9% recall, and 93.6 mAP@50 under the same dataset, preprocessing, augmentation, and training configuration.The matched configuration was used to ensure a fair comparison.Source: F. Detection performance of baseline… · pp. 5–6
- Flowering and fruiting stages had lower detection performance, with mAP@50 values of 92.5% and 91.6%, respectively, attributed to inter-class visual similarity, partial occlusion, and complex greenhouse backgrounds.Source: F. Detection performance of baseline… · p. 6
G. Precision and Recall curve
The baseline YOLOv5s precision–recall analysis examines performance across confidence thresholds and indicates stable behavior over a wide recall range, while noting limits for fine-grained phenological transitions.
G. Precision and Recall curve
- The baseline YOLOv5s precision–recall curve tracks the relationship between precision and recall across confidence thresholds on validation data.Table III is identified as the baseline YOLOv5s detection-performance table.Source: G. Precision and Recall curve · p. 6, Table III · p. 6
- The curves show stable performance over a wide recall range, indicating consistent object localization and classification capability.The passage points to Fig. 5 for this curve analysis.Source: G. Precision and Recall curve · p. 6
- The P–R analysis identifies limits in capturing fine-grained phenological transitions, motivating architectural refinements.Source: G. Precision and Recall curve · p. 6
H. Model Complexity and Computational Efficiency of base model
The discussion compares YOLOv5s and Pheno-Lite + ECA across detection and efficiency metrics, describing a more favorable overall balance for the smaller proposed model and identifying future development directions.
H. Model Complexity and Computational Efficiency of base model
- YOLOv5s has approximately 7 million parameters and 15.8 GFLOPs at 640 × 640, and its resource demand may limit use in constrained agricultural settings.The discussion presents the baseline as a reference for assessing lightweight modifications that reduce complexity while preserving detection performance.Source: H. Model Complexity and Computational… · p. 6
IV. DISCUSSION
- The framework targets greenhouses with limited computational resources, power availability, and automation infrastructure.The discussion frames its design as efficiency-aware architectural refinement rather than exhaustive network expansion.Source: §IV · p. 6
- Pheno-Lite + ECA records 90.6% precision versus 89.8% for YOLOv5s, while YOLOv5s records 93.6% mAP@50 versus 92.6% for the proposed model.The proposed model has approximately 4.0 million parameters and 10.9 GFLOPs, compared with YOLOv5s at approximately 7 million parameters and 15.8 GFLOPs; the discussion characterizes recall as comparable and the overall balance as more favorable for Pheno-Lite + ECA.Source: §IV · p. 7, Table IV · p. 8
- The radar chart compares normalized precision, recall, mAP@50, parameters, and GFLOPs, with lower resource demands receiving higher efficiency scores.Normalization puts all metrics on a common higher-is-better scale; the chart is described as showing a more favorable overall balance for Pheno-Lite + ECA.Source: §IV · p. 7, Figure 6 · p. 7
- Future work includes adaptive feature modeling, time-series greenhouse data, and environmental-sensor integration for decision-support systems.The proposed directions address detection improvement, temporal phenological transitions and seasonal variability, and irrigation and fertigation scheduling.Source: §IV · pp. 7–8
V. CONCLUSION
The conclusion presents Pheno-Lite + ECA as a lightweight, phenology-aware detector for tomato growth stages in constrained greenhouse environments. It reports a compact model and frames the work as a foundation for climate-resilient precision agriculture.
V. CONCLUSION
- Pheno-Lite + ECA is designed for tomato growth-stage recognition in greenhouses with limited automation and constrained computational resources.Source: §V · p. 8
- The model achieves high detection precision with approximately 4.0 million parameters and 10.9 GFLOPs, offering an accuracy–efficiency balance relative to YOLOv5s.Source: §V · p. 8
- The conclusion describes the framework as supporting real-time inference and practical deployment on low-power edge devices.It presents the work as a foundation for stage-specific greenhouse monitoring in Bhutan and similar developing agricultural settings.Source: §V · p. 8