Intelligent identification method for rice seedling growth stages and its application in laser supplementary lighting control research

Abstract

Accurately identifying the growth stages of rice seedlings is crucial for managing factory nurseries and ensuring consistent seedling quality. This study introduces PGL-ShuffleNetV2, a lightweight and advanced model designed for efficient and accurate recognition of rice seedling growth stages. The proposed model achieves a streamlined architecture by: 1). removing the second 1 × 1 convolution in the downsampling block’s right branch. 2) Reducing the repetition of basic units for improved efficiency. Additionally, the GELU activation function replaces ReLU to enhance nonlinear representation capabilities, and a parallel weighted hybrid attention module (PWMAM) is incorporated to improve feature extraction. Experimental results demonstrate that PGL-ShuffleNetV2 achieves a remarkable 98.80% recognition accuracy and a 98.82% F1 score, with a compact model size of just 0.84 MB. Its optimal balance between accuracy and parameter efficiency makes it highly suitable for deployment on resource-constrained devices, enabling effective monitoring and management of rice seedlings in factory nursery environments. Based on the advantages of this model, this study further applied it to rice seedlings under laser supplementary lighting conditions to investigate the impact of laser on the growth stages of seedlings, providing technical support for the application of laser technology in intelligent seedling cultivation.

Introduction

Rice is one of the most important staple crops worldwide, providing a primary food source for nearly half of the global population¹. Economically, the global rice market is valued at several hundred billion US dollars, and worldwide rice production is projected to increase by 5.8% to 567 million tons (Mt) by 2030². The seedling stage plays a critical role in rice cultivation, as seedling quality directly influences subsequent yield. During this stage, temperature and humidity are regulated according to the specific environmental requirements of seedlings at each growth phase. Therefore, accurate identification of rice seedling growth stages is essential for implementing efficient and optimized seedling management practices.

The growth stages of rice seedlings are typically classified according to the BBCH (Biologische Bundesanstalt, Bundessortenamt und Chemische Industrie) scale³, an internationally recognized decimal system for crop development. During the seedling period, particular attention is given to key phases within growth stage 1 (leaf development), including the full expansion of the first leaf (BBCH11), second leaf (BBCH12), and third leaf (BBCH13). Rice transplanting is typically performed no later than BBCH13⁴. Currently, the identification of seedling growth stages primarily relies on manual field observations, which are both time-consuming and labor-intensive. Therefore, there is an urgent need for low-cost, efficient, and accurate methods for seedling growth stage recognition.

Optimal light conditions can significantly improve rice seedling quality. Laser technology has attracted considerable attention in agriculture owing to its superior photoelectric conversion efficiency. Studies have shown that laser stimulation of seeds or seedlings can affect crop yield⁵. Laser irradiation can accelerate starch degradation, thereby promoting plant growth and development⁶. However, research on the effects of laser treatment on vigorous rice seedlings remains limited. Therefore, investigating the effects of laser treatment on rice seedling growth is imperative for developing novel approaches to cultivate robust seedlings.

The application of computer vision in agriculture has become increasingly important⁷. Previous studies have applied Otsu’s thresholding to segment rapeseed seedlings⁸ and combined morphological processing with dual-thresholding to monitor early-stage corn growth⁹. In multispectral remote sensing, features such as the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI) are commonly used for weed identification¹⁰. However, most of these approaches rely on thresholding and require manual parameter tuning, thereby limiting their automation and practical applicability.

Quan et al.¹¹ proposed an enhanced Faster R-CNN model for monitoring corn seedlings at various growth stages. Wan et al.¹² optimized the U-Net architecture to segment and identify corn growth stages from field images. Hussain et al.¹³ and Teimouri et al.¹⁴ utilized deep convolutional neural networks (DCNNs) to identify weed characteristics across growth stages. The models, including Faster R-CNN, belong to the two-stage object detection family. Although these models achieve high accuracy, they impose substantial computational demands and exhibit slow inference speeds, rendering them unsuitable for real-time applications.

In contrast, single-stage detection models such as YOLO have been widely applied in agricultural scenarios, integrating object localization and classification into a single regression task. These models have been applied to fruit detection in crops such as apples¹⁵, tomatoes¹⁶, corn¹⁷, and peppers¹⁸, as well as to pest and disease identification¹⁹. Paul et al. employed YOLOv9 for nighttime chili pepper detection in greenhouse environments²⁰ and YOLOv8 for bell pepper detection and growth stage classification²¹. Moreover, Vision Transformers (ViTs) have demonstrated accuracy comparable to, or even surpassing, that of CNNs in tasks such as plant disease identification²² and corn seed classification²³.

However, agricultural applications are often deployed on resource-constrained edge devices, such as drones and field robots, which impose strict constraints on power consumption, computational capacity, and storage. Consequently, lightweight architectures, such as MobileNet²⁴ and ShuffleNet²⁵, are favored over large-scale ViT or standard YOLO models. For example, ShuffleNetV2 achieved over 96% accuracy in corn seed sorting²⁶ and tomato disease detection²⁷, while maintaining extremely low FLOPs and model size.

Therefore, this study focuses on rice seedlings as the primary research subject and employs laser illumination as the control method. By integrating image recognition technology, it achieves synergistic optimization across three key components: lightweight network architecture, activation functions, and attention mechanisms. On this basis, an improved model, termed PGL-ShuffleNetV2, (where PGL represents the Parallel Weight Mixing Attention mechanism, Gaussian Error Linear Unit, and L emphasizes the Lightweight structure). The specific research objectives of this study are as follows:

1. Simplify the structural design of the lightweight ShuffleNetV2 architecture to improve model inference speed and satisfy the low-latency deployment requirements of edge devices.

2. Incorporate the Gaussian Error Linear Unit (GELU) activation function to enhance nonlinear representational capacity, achieve smoother feature mapping, and improve the ability to capture complex characteristics of seedlings at different growth stages;

3. Develop a Parallel Weight Mixing Attention Mechanism (PWMAM) to enhance the extraction and fusion of key features in dense regions, thereby improving recognition robustness;

4. Integrate laser supplementary lighting control technology with lightweight recognition models to evaluate the impact of laser light on rice seedlings, offering a novel strategy for promoting robust seedling growth.

The research process consists of four main stages: (1) image acquisition and preprocessing; (2) model design and optimization; (3) model training and performance evaluation; and (4) identification of rice seedling growth stages under laser treatment (Fig. 1).

Fig. 1.

Analysis process of the effects of the laser on rice seedlings.

Materials and methods

Data acquisition and preprocessing

Data acquisition

Data collection for rice seedling growth stages in this study was divided into two components: model training data and laser effect analysis data. Model training data were collected in the greenhouse facilities at Jilin Agricultural University’s research Base (Fig. 2).

Fig. 2.

Data acquisition.

Rice seedling imaging was performed between April and May 2023, encompassing the critical growth stages from BBCH11 to BBCH13. The Honor X10 smartphone, featuring a 40 MP main camera, an 8 MP ultra-wide camera, and a 2 MP macro lens, was used for imaging (Table 1). For each sampling session, 100 seedlings were randomly selected from each plot. When more than 80 seedlings exhibited consistent growth stages, the corresponding stage was designated as the representative growth stage for that plot. Through meticulous annotation of these seedling growth stages, a dataset comprising three distinct growth stages, totaling 458 images, was constructed. The dataset included 120, 185, and 153 images for each respective growth stage.

Table 1.

Data acquisition protocol for capsicum image acquisition.

Parameter	Details
Camera Type	High-resolution digital camera
Camera Specification	featuring a 40 MP main camera, an 8 MP ultra-wide camera, and a 2 MP macro lens
Image width x height	2400 × 1080
Lighting Conditions	Natural lighting sources
Image Format	JPG
Scene Selection Criteria	Different growth stages; vertical angle

The experimental setup for laser supplemental lighting during seedling cultivation is illustrated in Fig. 3. The experiment employed laser plant growth lamps, developed by Zhejiang Changxin Optoelectronics Co., Ltd., as the light source, with the red laser having a central wavelength of approximately 660 nm and the blue laser approximately 450 nm. The equipment is classified as a Class 2 laser according to IEC 60825-1 standards, posing no safety risks to either plants or operators²⁸.

Fig. 3.

Time series diagram of laser treatment on rice seedling growth. (a) Laser Seedling Rearing Trial Site; (b) Laser treatment of rice seedlings; (c) Rice seedling growth process.

Six distinct treatment groups were established in the experiment, as summarized in Table 2. The CK group served as the control, utilizing natural light without supplemental illumination. The R and B groups received supplemental lighting with only red or blue lasers. The R1B1, R3B1, and R3B2 groups were treated with red-to-blue laser ratios of 1:1, 3:1, and 3:2, respectively.

Table 2.

Color ratio composition of the laser in each treatment.

Laser treatment	Explanation
CK	No laser light irradiation (natural light)
R	Red laser (660 nm)
B	Blue laser (450 nm)
R1B1	50% red laser + 50% blue laser
R3B1	75% red laser + 25% blue laser
R3B2	60% red laser + 40% blue laser

To examine the effects of laser exposure on the growth stages of rice seedlings, this study employed fixed-interval continuous monitoring. Data were collected over 15 consecutive days, encompassing the critical developmental stages BBCH11 through BBCH13.

The training images contained redundant information, including roads, seedling tray edges, and regions with relatively sparse seedlings, which could adversely affect model training accuracy. Furthermore, large image dimensions decrease model training efficiency. To mitigate these issues, the images were cropped to 224 × 224 pixels, and those containing redundant information were removed. The dataset of rice seedling growth stages was categorized into three classes—BBCH11, BBCH12, and BBCH13, as illustrated in the Fig. 4.

Fig. 4.

Growth stages of rice seedlings.

To maintain relative balance among categories in the dataset and improve the model’s generalization capability, random image augmentation was applied²⁹. The augmentation strategy comprised adjustments to image brightness³⁰, contrast, and chrominance³¹. Figure 5 illustrates the effects of brightness, contrast, and chrominance adjustments at factors of 1.2, 0.7, and 1.3, respectively. After augmentation, the dataset comprised 19,928 images. The dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio, a benchmark approach widely adopted in deep learning research³².

Fig. 5.

Effects of image enhancement: (a) Original image; (b) Brightness modification; (c) Contrast modification; (d) Chroma modification.

Model construction

To achieve fast and accurate recognition of rice seedling growth stages, this study adopts the lightweight ShuffleNetV2 network and optimizes it by eliminating redundant layers, replacing activation functions, and integrating an attention mechanism, resulting in the proposed PGL-ShuffleNetV2 model, as shown in Fig. 6.

Fig. 6.

PGL-ShuffleNetV2 structure. PWMAM: Parallel Weighted Mixed Attention Module. The The number of basic units in ShuffleNetV2’s stage 1, stage 2, and stage 3 is set to 2, 4, and 2.

The model leverages ShuffleNetV2 as a foundation for constructing a backbone feature extraction network. It reduces the number of parameters by eliminating 1 × 1 convolutional layers and decreasing network depth. Additionally, it incorporates the GELU activation function to address the gradient vanishing issue and enhances feature extraction performance. By introducing a parallel weight fusion attention mechanism, the model improves the representation of relevant features while mitigating the impact of extraneous ones.

ShuffleNetV2

In 2018, Ma et al. proposed ShuffleNetV2. The fundamental structure of ShuffleNetV2 consists of two main types of units. The basic unit divides the input channels into two parts: the left branch remains unchanged, without any operation; the right branch performs a 1 × 1 convolution, a 3 × 3 depthwise convolution, and another 1 × 1 convolution sequentially, and finally concatenates the two parts at the bottom of the module. Subsequently, a random mixing operation is applied to the output feature map channels, as shown in Fig. 7a. The downsampling unit assigns all feature maps to the left and right branches, concatenates the output feature maps from both branches at the bottom of the module, and then performs a channel shuffle operation, as shown in Fig. 7b.

Fig. 7.

The basic feature extraction unit in the ShuffleNetV2 model consists of several components: ‘Conv’ refers to standard convolution, ‘BN’ denotes batch normalization, ‘ReLU’ represents the activation function, and ‘Concat’ refers to channel concatenation. (a) Basic unit; (b) Downsampling unit.

GELU activation function

The GELU activation function is defined as a smooth, continuous S-shaped curve, offering a more gradual transition compared to the ReLU function. GELU shares similarities with the Sigmoid function in terms of nonlinear transformations, allowing it to produce outputs over a broader range. This property enhances its ability to capture complex patterns and relationships within the data. The mathematical formulation of the GELU activation function is presented in Eq. (1):

1

Where x represents the input value, while X denotes a Gaussian random variable with zero mean and unit variance. P(X≤x) Indicates the probability that X is less than or equal to a given value x. Φ(x) corresponds to the cumulative distribution function of the Gaussian normal distribution for x.

The GELU activation function can be approximated using the standard normal distribution, which facilitates its computation. This approximation is mathematically represented as shown in Eq. (2):

2

Where 0.5 acts as a scaling factor, and tanh represents the hyperbolic tangent function, which applies a nonlinear transformation to the input.

PWMAM module

The Convolutional Block Attention Module (CBAM) is an attention mechanism that integrates the Channel Attention Module (CAM) and the Spatial Attention Module (SAM). In the CAM, the input feature map 𝐹 undergoes global average pooling and global maximum pooling operations, generating two distinct feature vectors. These vectors are then combined using a shared multilayer perceptron (MLP), after which feature extraction is completed. Finally, the channel attention vector 𝑀𝑐 is obtained by applying a sigmoid activation function, as described in Eq. (3):

3

Where 𝜎 denotes the sigmoid activation function, AvgPool indicates average pooling, and MaxPool indicates max pooling.

The Channel Attention Module processes the input feature map 𝐹 to generate a weighted feature map 𝐹′, as shown in Eq. (4):

4

After the CAM-adjusted feature maps are passed into the SAM, the feature map 𝐹′ undergoes global maximum pooling and global average pooling, and the two resulting feature maps are merged. Subsequently, the merged feature map is convolved with a 7 × 7 convolution kernel and processed through a sigmoid activation function to produce the spatial attention vector 𝑀𝑠, as described in Eq. (5):

5

Where f ^7 × 7denotes a 7 × 7 convolutional layer.

The feature map 𝐹′ is element-wise multiplied with the spatial attention vector 𝑀𝑠 to obtain the weighted feature map 𝐹′′, as shown in Eq. (6):

6

It can be observed that CBAM sequentially connects the channel attention mechanism and the spatial attention mechanism. This sequential approach may cause the inputs to the spatial attention mechanism to be based on features that have already been adjusted by the channel attention mechanism, rather than the original features. Such a situation may, to some extent, impact the feature learning of the spatial attention mechanism.

To address the aforementioned issue, this study proposes a module called the Parallel Weighted Mixed Attention Module (PWMAM), as illustrated in the Fig. 8. In this module, CAM and SAM are designed as parallel structures, ensuring that the inputs to both the channel attention and spatial attention mechanisms remain independent and unmodified. During the feature fusion process, two learnable parameters, α and β are introduced to regulate the weight distribution between channel attention and spatial attention. These two types of attention are then merged through a weighted sum, and the entire fusion process is expressed in Eq. (7):

7

Fig. 8.

The Parallel mixed weight attention mechanism (PWMAM) network structure. Mc is the channel attention vector; Ms is the spatial attention vector;  α  is the channel attention fusion weight; and β is the spatial attention fusion weight.

Where α denotes the weight assigned to channel attention, where β denotes the weight assigned to spatial attention.

Experimental environment setup

The experimental training platform was configured with an Intel^® Xeon^® Gold 6246R CPU running at 3.40 GHz and 128 GB of RAM. It also featured an NVIDIA Quadro RTX 8000 GPU with 48 GB of dedicated graphics memory. The GPU’s substantial parallel computing capabilities and high-capacity memory made it well-suited for training and debugging deep neural networks, thereby significantly accelerating experimental iterations. The software environment comprised Python 3.8 and CUDA 11.3, as detailed in Table 3.

Table 3.

The parameters of the experimental environment.

Environment	Parameters
OS	Win 10
CPU	Intel(R) Xeon(R) Gold 6246R@3.40 GHz 32 cores
Memory	128
GPU	NVIDIA Quadro RTX 8000
Deep learning framework	Pytorch-GPU 1.12.1
CUDA version	CUDA Toolkit 11.3
Python version	Python3.8

This study employed the Adam optimization algorithm, a variant of stochastic gradient descent, along with the cross-entropy loss function to assess model performance. Based on preliminary experiments, the learning rate was set to 0.001 and the batch size to 64. Detailed hyperparameter settings are provided in Table 4.

Table 4.

The model parameter settings.

Parameter name	Value	Importance
Learning Rate	0.001	The learning rate affects the efficiency and performance of model training, determining the magnitude of weight updates and directly influencing convergence speed and results.
Batch size	64	The number of samples processed before updating the model impacts the stability and speed of the training.
Loss Function	Cross Entropy	The loss function measures how well the model’s predictions match the actual results.
Optimizer	Adam	Optimizer helps the model learn and improve.

Evaluation metrics

In this study, Accuracy (Acc), Precision (P), Recall (R), and F1-score (F1) were adopted as the performance evaluation metrics for the model, as defined in Eqs. (8)–(11).

8

9

10

11

TP (True Positive): The number of seedling images correctly classified by the model as belonging to a specific growth stage.

FP (False Positive): The number of seedling images that belong to other growth stages but are incorrectly classified by the model as belonging to the current target stage.

TN (True Negative): The number of images correctly classified by the model as not belonging to the current target stage.

FN (False Negative): The number of seedling images that belong to the current target stage but are incorrectly classified by the model as belonging to other stages.

Additionally, FLOPs serve as a key metric for assessing a model’s computational efficiency. Lower FLOPs indicate fewer required floating-point operations, which directly enhance the model’s capability to perform rapid, real-time identification of seedling growth stages on mobile field devices. Simultaneously, reducing the number of model parameters decreases hardware deployment costs and improves the model’s practical applicability.

Results

Model improvement and comparison results

Model lightweight test

To reduce the model’s complexity, it was designed with fewer parameters, and the results of this optimization are presented in Fig. 9. The term ‘_conv’ refers to the removal of the second 1 × 1 convolutional layer in the right branch of ShuffleNetV2, while the accompanying numbers represent the number of base units in Stage2, Stage3, and Stage4. The ShuffleNetV2_conv variant achieves an accuracy of 95.9% with 0.92 million parameters. Compared with the base model, this variant demonstrates a 0.70% improvement in accuracy and a parameter reduction of 0.34 million. Adjustments to the number of base units in Stage 2, Stage 3, and Stage 4 were made following the removal of convolutional layers. Setting the number of base units in Stage 2, Stage 3, and Stage 4 to 2, 4, and 2, respectively, results in an accuracy of 96.2% with 0.75 million parameters. This configuration improves accuracy by 1% while reducing model parameters by 0.51 million. Based on a trade-off between accuracy and parameter count, this lightweight scheme was selected for optimizing the model’s performance in this study.

Fig. 9.

Model lightweight results.

Comparison of different activation functions

This study introduces ReLU, ReLU6, Hard Swish, and GELU activation functions into the ShuffleNetV2 unit, with the results summarized in Table 5. The model achieves accuracies of 95.2%, 97.23%, 97.4%, and 97.8% when using the aforementioned activation functions. Compared to ReLU, the accuracy of the model improves by 2.03%, 2.20%, and 2.60% when ReLU6, Hard Swish, and GELU activation functions are used, respectively. With the GELU activation function, the model achieves an accuracy of 97.80% and an F1 score of 97.84%. The smooth nonlinear properties of the GELU function, coupled with its ability to effectively handle negative values, enable it to capture more intricate features in rice seedling images, thereby improving model accuracy.

Table 5.

Experimental results using different activation functions.

Activation function	Acc(%)	P(%)	R(%)	F1(%)
RELU	95.20	95.31	95.27	95.28
RELU6	97.23	97.23	97.29	97.26
HardSwish	97.40	97.42	97.40	97.40
GELU	97.80	97.81	97.87	97.84

Comparison of different attention mechanisms

As shown in Table 6, the accuracy of each model is improved by adding the attention mechanism. After incorporating SE, ECA, CBAM, and PWMAM, the model achieves accuracies of 97.3%, 97.45%, 97.7%, and 98.2%, respectively, with corresponding parameters of 1.30 M, 1.26 M, 1.42 M, and 1.30 M. ECA captures fine texture features of the rice seedlings through local feature aggregation, while CBAM selects and focuses on key channel and spatial features to improve recognition accuracy. The PWMAM attention mechanism fuses spatial and channel attention in parallel with different weights to improve the accuracy of model parameters by 3% compared to the base model. Experimental results indicate that the integration of the PWMAM module enhances the model’s focus on seedling-specific features, thereby improving recognition accuracy, as illustrated in Fig. 10.

Table 6.

Experimental results using attention mechanisms.

Model	Acc(%)	P(%)	R(%)	F1(%)	Params (M)	Flops(M)
ShuffleNetV2	95.20	95.30	95.27	95.28	1.26	151.7
ShuffleNetV2_SE	97.30	97.35	97.38	97.36	1.30	152.1
ShuffleNetV2_ECA	97.45	97.50	97.50	97.50	1.26	152.1
ShuffleNetV2_CBAM	97.70	97.76	97.74	97.75	1.42	152.3
ShuffleNetV2_PWMAM	98.20	98.25	98.24	98.26	1.30	152.7

Fig. 10.

Comparison curves of attentional mechanisms: (a) Attentional accuracy curve, (b) Attentional loss curve.

Ablation experiment

To evaluate the effects of architectural adjustments, activation function substitutions, and improvements in the PWMAM module on ShuffleNetV2 model performance, ablation experiments were performed, and the results are presented in Table 7.‘Lite’ refers to the architectural tuning method that removes the second 1 × 1 convolutional layer from the right branch and reduces the number of basic units. This tuning method optimizes the ShuffleNetV2 model, improving accuracy by 1% and reducing model parameters by 0.45 M. The GELU activation function is employed to probabilistically scale the input size either up or down. The smoothing property enhances the network performance, increasing accuracy by 1.9% while keeping model parameters unchanged. The incorporation of the PWMAM module leads to a 0.7% improvement in model accuracy. Consequently, the improved PGL-ShuffleNetV2 model achieves an accuracy of 98.80%, representing a 3.60% increase over the original model, with a reduction of 0.42 M parameters and 58.4 M FLOPs. The accuracy and loss curves of both the original and improved models are shown in Fig. 11. The improved model outperforms the original model in terms of convergence speed and training stability.

Table 7.

Experimental results from the ablation test.

	Lite	GELU	PWMAM	Acc (%)	Params (M)	FLOPs (M)
ShuffleNetV2	×	×	×	95.20	1.26	151.6
	√	×	×	96.20	0.81	92.5
	×	√	×	97.80	1.26	151.6
	×	×	√	98.20	1.30	152.7
	√	√	×	98.10	0.81	92.5
	×	√	√	98.49	1.30	152.7
	√	×	√	98.30	0.84	93.2
	√	√	√	98.80	0.84	93.2

Fig. 11.

Comparison of accuracy and loss curves between the original model and the improved model.

As illustrated in Fig. 11, both the training and validation losses of the model stabilized after approximately 160 epochs. Despite the early convergence, the total number of training epochs was set to 200. The validation loss did not exhibit any notable increase in the subsequent epochs, indicating that overfitting did not occur during training.

Comparison with different models

The objective of this study was to evaluate the effectiveness of the PGL-ShuffleNetV2 model by comparing it to the classical network, as presented in Table 8. Compared to the classical models ResNet18 and DenseNet121, PGL-ShuffleNetV2 achieves an accuracy increase of 0.9% and 0.8%, respectively. When compared to lightweight convolutional networks such as EfficientNet, MobileNetV2, and GhostNet, the accuracy of PGL-ShuffleNetV2 improves by 0.75%, 0.9%, and 3.2%. Additionally, the number of parameters in PGL-ShuffleNetV2 is only 21%, 38%, and 21.4% of the parameters in EfficientNet, MobileNetV2, and GhostNet. Overall, PGL-ShuffleNetV2 demonstrates the highest accuracy, the lowest number of parameters, and the smallest FLOPs, reflecting the most efficient resource consumption.

Table 8.

Experimental results comparing different networks.

Model	Acc (%)	P(%)	R(%)	F1 (%)	Params (M)	FLOPs (M)
ShuffleNetV2	95.20	95.30	95.27	95.28	1.26	151.6
ResNet18	97.90	97.95	97.96	97.95	11.2	1823.5
DenseNet121	98.00	98.07	98.07	98.06	7.0	2895.7
GhostNet	95.60	95.72	95.69	95.69	3.91	154.6
MobileNetV2	97.90	97.96	97.96	97.96	2.2	326.3
EfficientNet	98.05	98.13	98.09	98.10	4.0	411.6
PGL-ShuffleNetV2	98.80	98.81	98.83	98.82	0.84	93.2

The performance of the model is assessed using the data obtained from the confusion matrix³³. Figure 12 illustrates the confusion matrix for the validation set of each model. At the BBCH11 stage, ResNet18, DenseNet121, and PGL-ShuffleNetV2 all achieved 100% classification accuracy. However, as the seedlings grow, the ground cover increases, and the texture of the seedling leaves becomes more complex. This complexity is particularly pronounced in the BBCH12 stage, where the differences from the other two stages are less distinct, leading to lower accuracy. Compared to other networks, the recognition method proposed in this study demonstrates superior accuracy and precision.

Fig. 12.

Confusion matrix of classical CNN model and PGL-ShuffleNetV2 model: (a) ShuffleNetV2; (b) ResNet18; (c) DenseNet121; (d) GhostNet; (e) MobileNetV2; (f) EfficientNet; (g) PGL-ShuffleNetV2.

Visualization of model results

Gradient-weighted class activation mapping (Grad-CAM)³⁴ is a technique for visualizing features by evaluating the contribution of image regions to the prediction results through class activation heatmaps, highlighting the regions the model focuses on. In the resulting heatmaps, red areas indicate regions with strong activation, while blue areas correspond to weakly activated regions³⁵. As shown in Fig. 13, the ShuffleNetV2 model tends to focus on a broader area, whereas the PGL-ShuffleNetV2 model demonstrates more concentrated attention on key features, such as the seedling leaf region and the soil. This enhanced focus allows the PGL-ShuffleNetV2 model to effectively capture and extract the texture features of rice seedlings, leading to more precise identification of critical areas.

Fig. 13.

Heat maps.

Laser control of rice seedling growth stages

Intelligent growth stage identification models and strategies

The PGL-ShuffleNetV2 model was employed to identify the growth stages of rice seedlings subjected to different laser treatments. The specific identification process was as follows:

Image preprocessing and subimage generation

To address potential issues such as angular deviations and uneven illumination in daily-collected top-view images of rice seedlings, a sliding-window cropping strategy was employed. Specifically, a fixed-size window of 224 × 224 pixels was moved across each original image with a 50% overlap, producing a series of subimages that together comprehensively capture the entire seedling canopy.

Model inference and prediction

Each subimage was input into the pre-trained PGL-ShuffleNetV2 model to generate a predicted probability distribution corresponding to the seedling growth stages, as expressed in Eq. (12):

12

Where p_{i1}, p_{i2}, and p_{i3} represent the probabilities that the subgraph is predicted to be in BBCH stage 11, BBCH stage 12, and BBCH stage 13, respectively. Each subgraph will output a definitive growth stage prediction result.

Voting ensemble and final decision

To integrate comprehensive information from the seedlings, a majority voting mechanism was employed. By aggregating the prediction results of all subimages derived from a single original image, the growth stage with the highest frequency was selected as the final predicted label. This process is expressed as in Eq. 13:

13

Where N represents the total number of sub-images, and mode denotes the statistical mode function.

Effects of laser supplemental lighting on rice seedling growth

Based on the aforementioned identification method, growth records of rice seedlings were obtained. The growth rate was quantified as the number of days required for transitions between key developmental stages, defined as follows:

T1: the number of days required to progress from BBCH stage 11 to BBCH stage 12;

T2: the number of days required to progress from BBCH stage 12 to BBCH stage 13.

Table 9 summarizes the median number of days required for seedlings in the laser treatment groups and the control group (CK) to complete each developmental stage transition.

Table 9.

Median days required for rice seedlings at different growth stages.

Laser treatment	T1(days)	T2(days)
CK	5	7
B	5	7
R	5	6
R1B1	5	6
R3B1	5	6
R3B2	5	6

As shown in Table 9, no differences were observed among treatments for T1, where as significant variation was evident in T2. Seedlings subjected to R, R1B1, R3B1, and R3B2 treatments developed faster than those in the CK and B groups. Specifically, these treatments required only 6 days to progress from BBCH12 to BBCH13, compared with 7 days in the CK and B groups, representing an acceleration of 1 day and an improvement of 14.3% in growth rate.

Discussion

Discussion on model performance improvement

Lightweight modelling

Lightweight models are essential for deploying deep learning algorithms directly on edge computing devices. Convolutional neural networks (CNNs) typically extract object features by increasing network depth. However, merely increasing network depth can result in overfitting to irrelevant features, while simultaneously increasing the number of parameters and computational complexity. As a lightweight architecture, ShuffleNetV2 integrates an adequate number of convolutional operations within its network design. Accordingly, we optimized the number of units across stages, achieving high accuracy while reducing model parameters. These results are consistent with the findings reported by Xu et al.³⁶ and Jin et al.³⁷. This parameter-reduction strategy provides a more efficient approach for deploying models on edge devices.

Improved attention mechanisms

Attention mechanisms are extensively employed to enhance a model’s ability to focus on salient features within regions of interest, thereby improving overall performance³⁸. The Squeeze-and-Excitation (SE) attention mechanism captures the relative importance of each channel, thereby endowing the network with enhanced discriminative capability³⁹. The Efficient Channel Attention (ECA) mechanism achieves substantial performance gains while introducing minimal additional parameters by facilitating appropriate cross-channel information exchange⁴⁰. The Convolutional Block Attention Module (CBAM) integrates attention across both channel and spatial dimensions, thereby enhancing the model’s contextual representation capability.

This study proposes the PWMAM, which adaptively integrates features by connecting the CAM and SAM in parallel, incorporating a weighted fusion strategy. Compared with existing approaches⁴¹, PWMAM autonomously adjusts the contribution weights of channel and spatial attention, thereby more effectively mitigating feature dimensionality imbalance. This significantly enhances the feature representation of rice seedlings, constituting a critical factor underlying the model’s performance improvement.

Although PGL-ShuffleNetV2 has been optimized for lightweight deployment, its real-time performance on resource-constrained embedded devices remains limited. Future work could investigate advanced techniques, including neural architecture search (NAS) and knowledge distillation, to identify more computationally efficient architectures.

Effects of laser on the growth stages of rice seedlings

Light significantly influences the growth of rice seedlings. Li et al.⁴² reported that exposure to combined red and blue light significantly increased total leaf number in Brasenia schreberi seedlings. Ren et al.⁴³ demonstrated that both leaf length and leaf number under red light exceeded those under blue light. These observations are consistent with the results of the present study, in which red light and its combinations with blue light (R1B1, R3B1, and R3B2) significantly promoted vegetative growth. The present study investigated the effects of laser supplemental lighting on rice seedlings from a phenotypic perspective. Nonetheless, the underlying mechanisms by which laser treatment influences photosynthesis, hormone regulation, and gene expression in rice seedlings remain to be elucidated.

Conclusions

This study introduces the PGL-ShuffleNetV2 model, specifically designed to accurately identify the growth stages of rice seedlings. With an impressive accuracy rate of 98.8%, this model represents a substantial improvement of 3.6% over its predecessor, all while reducing the number of parameters by 33.3%. When compared to notable models such as ResNet18, GhostNet, MobileNetV2, and EfficientNet, the enhanced PGL-ShuffleNetV2 stands out for its compact size and superior accuracy, making it an ideal candidate for deployment on small mobile devices. This remarkable performance underscores its significant potential for monitoring rice seedling growth in agricultural production. In this study, the PGL-ShuffleNetV2 model was integrated with laser treatment experiments to investigate the effects of laser exposure on rice seedling growth stages. The results indicate that rice seedlings subjected to laser treatments exhibit distinct developmental changes across growth stages. Red light and red-blue combination laser treatments effectively enhance leaf development, resulting in accelerated leaf growth. These findings provide important insights for the optimization of rice seedling cultivation techniques.

Despite the aforementioned achievements, this study presents several limitations. First, the experimental data utilized in this study are relatively idealized. Additional data accounting for factors such as uneven lighting, weather variations, and shadow occlusion is required to further enhance the model’s generalization ability and robustness under complex scenarios. Second, future research will focus on integrating this model into lightweight IoT devices and embedded development platforms. This integration will facilitate the development of a real-time monitoring system that combines recognition, analysis, and decision-making functionalities, thereby serving as a decision-support tool for agricultural production processes.

Author contributions

Xiaoyan Liu : Methodology, Conceptualization, Validation, Resources, Writing—review and editing, Funding acquisition.Leijinyu Zhou: Supervision, Software, Data curation, Writing—original draft preparation.Bei Jiang : Software, Data curation, Visualization.Helong Yu : Formal analysis, Visualization, Data curation, Supervision, Project administration.

Funding

This research was supported by the Science and Technology Development Programme of Jilin Province (NO:20240601079RC,202202032NC); the Science and Technology Research Project of Jilin Provincial Department of Education ( No: JJKH20241706KJ).

Data availability

The datasets used during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.