Research on Drought Stress Monitoring of Winter Wheat during Critical Growth Stages Based on Improved DenseNet-121

Abstract

Drought stress has serious effects on the growth and yield of wheat in both productivity and quality and is an abiotic factor. Traditional methods of crop drought stress monitoring have some deficits. This work has been conducted in order to enhance these conventional methods by proposing a new deep learning approach. This paper has presented a deep learning-based model customized for monitoring drought stress in winter wheat during the critical growth stages. Drought-afflicted winter wheat images were captured at three crucial phases: rising–jointing, heading–flowering, and flowering–maturity. These images are correlated against soil moisture data to construct a comprehensive dataset. DenseNet121 was chosen as the network model since it extracts features from images relating to phenotypes. Several factors, like training methods, learning rate adjustment, and addition of the attention mechanism, are optimized in eight sets of experiments. This provided the final DenseNet-121 model with an average recognition accuracy of 94.67% on the test set, which means that monitoring drought stress during wheat growth’s key periods is feasible and effective.

1. Introduction

Out of the three key cereals, about 70% of wheat cultivation areas fall under arid and semi-arid agricultural zones [1]. Statistics show that droughts strike China 7.5 times on average every year, while the average afflicted crop area ranges from 20 to 30 million hm² and the average reduction in grain output amounts to 250–300 billion hm² each year, thus creating very serious challenges for grain production and security [2]. Severity, duration, timing, and location of the drought are all factors that can alter the effects of drought on wheat yield and quality. From the present research work, it is clear that the reduction in wheat yield was proportional to both the degree of drought stress and the specific growth stage during which the drought happened [3]. In particular, drought stress during jointing, heading, and grain-filling has a huge effect on wheat growth and output, reducing its quantity and quality [4]. This will require timely data on wheat drought monitoring, correct identification of drought stress, and timely irrigation for early warning and mitigation to ensure improved grain yield.

Traditional approaches to assessing soil and crop water status encompass a range of techniques such as agricultural meteorological observations, direct soil moisture assessments, thermal infrared imaging, hyperspectral analysis, chlorophyll fluorescence detection, and manual evaluation. While each of these methods offers insights into the state of drought in crops, they all come with inherent delays or constraints that may affect their accuracy or timeliness [5]. In farmland irrigation areas, the information from agricultural meteorological drought monitoring is somewhat limited. Irrigation may alter the status of soil moisture but cannot alter the humidity and temperature in time, which is monitored by meteorological monitoring systems [6]. In contrast, general indirect monitoring is soil moisture monitoring, but because it is not very accurate and does not have a large scope of coverage, application is somewhat constrained [7]. Scientists utilize a suite of advanced imaging techniques, including thermal infrared, hyperspectral, and chlorophyll fluorescence, to directly assess water stress in crops. These methods are instrumental in evaluating the hydration status of the plant canopy and foliage [8]. For example, Meng Y. et al. used thermal infrared imaging to analyze maize drought resistance and explored rapid and efficient selection of winter wheat drought-resistant varieties [9]. Mangus et al. utilized high-resolution thermal infrared images to explore the relationship between canopy temperature and soil moisture [10]. Although thermal infrared imaging is capable of detecting signs of drought stress in crops, its ability to cover large areas is limited and it can be affected by various environmental factors as well as the specific types of crops [11]. Hyperspectral technology, which reflects crop stress states through spectral features [12], is widely applied in the surveillance of crop water scarcity; the spectrum that indicates drought sensitivity typically spans from 1200 nm to 2500 nm [13]. Chlorophyll fluorescence is a responsive indicator for detecting the initial signs of water scarcity in crops, yet it faces difficulties when it comes to tracking more extreme drought conditions. The existing technology for measuring chlorophyll fluorescence is predominantly limited to the examination of small-scale plants or those in their early growth phases.

Overseeing extensive agricultural fields or the specific characteristics of crops can be quite a daunting task. Nonetheless, the evolution of computer vision and image processing has opened new avenues. Nowadays, deep learning algorithms that utilize two-dimensional digital photographs are extensively applied to distinguish and categorize various types of crop stress, whether caused by living organisms or non-living factors [14]. Deep learning combines image feature extraction and classification, automatically learning from images and improving recognition accuracy. It enables more accurate and objective stress identification and grading compared to traditional machine learning. Extensive research indicates that deep learning models surpass previous image recognition techniques [15], with extensive research indicating their high recognition accuracy and broad application advantages [16,17]. By 2050, global demand for agricultural products on current agricultural land will increase by 55%, while reducing the need for fertilizers and using water efficiently [18]. Although progress has been made in the study of drought stress phenotypes, current research predominantly focuses on a single crop. There is a notable scarcity of studies on drought stress monitoring in winter wheat, and there is a lack of publicly available datasets. Most existing monitoring efforts are confined to pot-based drought control experiments, with virtually no research conducted in field environments. To address this gap, this paper employs the DenseNet-121 network model as the foundational architecture for extracting phenotypic features of winter wheat under drought stress. By adjusting the model’s training methodology, fine-tuning the learning rate, and incorporating an attention mechanism as variables, we have conducted a series of eight integrated experiments to train and optimize the model. The resulting DenseNet-121-based identification model for winter wheat drought stress during key growth periods offers an effective strategy for monitoring drought stress in winter wheat.

2. Datasets and Methods

2.1. Data Preparation

In the study, the categorization of drought levels during the pivotal growth stages of wheat was based on the Chinese agricultural standard “Field Survey and Grading Technical Specifications for Winter Wheat Disaster”, specifically the section on winter wheat drought disaster (NY/T 2283-2012) [19]. These stages include the jointing (RJ), the heading and flowering (HF), and the flowering to maturity (FM) phases. Drought conditions were classified into five categories: Optimum Moisture (OM), Light Drought (LD), Moderate Drought (MD), Severe Drought (SD), and Extreme Drought (ED) [20], as detailed in

Table 1. Standards for drought levels of wheat at different growth stages.

Drought Level	RJ	HF	FM
OM	>65%FC	>70%FC	>70%FC
LD	60%≤~<65%FC	65%≤~<70%FC	65%≤~<70%FC
MD	55%≤~<60%FC	60%≤~<65%FC	60%≤~<65%FC
SD	45≤~<55%FC	55≤~<60%FC	55≤~<60%FC
ED	<40%FC	<45%FC	<45%FC

Note: FC (field capacity) is the field water holding capacity.

. Considering the uneven soil moisture distribution and the challenges in precise irrigation management, soil moisture sensors were strategically placed using a greedy ant colony algorithm for node deployment, ensuring an accuracy of ±1%. With the deployment of these sensors, data regarding soil moisture levels were collected. Concurrently, monitoring equipment was utilized to photograph wheat under the various drought conditions, creating a dataset that correlates images of wheat experiencing different levels of drought stress with the soil moisture monitoring data.

2.2. Dataset Description

The experiment was conducted at the Agricultural Water Efficiency Laboratory of North China University of Water Resources and Electric Power from April 2021 to June 2022. Here, there are three key growth stages of winter wheat with a large impact of drought stress: RJ, HF, and FM. Under real-time monitoring by soil moisture sensors, wheat sample images from these three key growth periods under different levels of drought were captured. After annotation and screening, a total of 12,500 images were used for model training (see

Table 2. Image quantity table of wheat growth stages under different drought levels.

Growth Stage	OM	LD	MD	SD	ED	Total Data
RJ	950	900	880	820	650	4200
HF	900	840	840	845	775	4200
FM	1220	920	850	720	490	4100

).

Table 3. Acquisition time of wheat drought images.

Year	RJ (Month/Day)	HF (Month/Day)	FM (Month/Day)
2021	3–23 April	24 April–2 May	3–23 May
2022	29 March–16 April	17–30 April	1–19 May

shows the image capture time for wheat, and Figure 1 shows samples of winter wheat images.

2.3. Deep Learning Model

The deep learning models selected in this paper are the ones with fewer parameters and higher accuracy trained on the ImageNet dataset, namely AlexNet, ResNet101, and DenseNet121 as the base network models.

AlexNet, a deep convolutional neural network proposed by Alex Krizhevsky in 2012 [21], swept the ImageNet image recognition competition that year. The structure of AlexNet is shown in Figure 2, which includes convolutional layers, pooling layers, fully connected layers, and an output layer. Several convolutional and pooling layers are alternately stacked, followed by two fully connected layers and one output layer. AlexNet led the development of deep learning in the field of computer vision. It pioneered the use of deep convolutional neural networks for image classification and laid the foundation for subsequent deep learning models.

ResNet (Residual Network) is a convolutional neural network model proposed by Kaiming He et al. in 2015 [22]. By introducing residual connections, it addresses the vanishing and exploding gradient problems encountered during the training of deep networks. The network structure of ResNet-50 is shown in Figure 3, characterized by its use of residual learning. By introducing “residual blocks”, it addresses the gradient vanishing and model degradation issues [23], enabling the network to learn deeper feature representations. ResNet101, with its deeper layers and stronger expressive power, is suitable for complex visual recognition tasks. The advent of ResNet greatly advanced the development of deep learning, making it possible to train deeper neural networks and achieving significant performance improvements in various tasks.

The Densely Connected Convolutional Network, a novel architecture within the realm of deep learning, was introduced by Gao et al. in 2019. This network is characterized by its deep convolutional structure [24]. As shown in Figure 4, the network structure of the DenseNet series is quite different from the traditional convolutional neural network because it connects the output of every layer to all previous layers’ output and forms a densely connected structure. This method of connection is good for feature transfer within the model and efficiently reduces the vanishing gradient problem, hence improving the model in terms of training efficiency and generalization capability. DenseNet-121 has 121 layers. This network is concise, and with a novel structure, it is efficient with better performance on the CIFAR index than the residual network.

2.4. Channel Attention Module ECANet

The wheat images collected in the field environment still pose challenges to the recognition ability of the network due to the complexity of the background, including but not limited to the loss of detail information and the lack of concentration of weight allocation, resulting in limited recognition accuracy. To overcome these limitations, this study introduces an attention mechanism module to enhance the feature expression ability of the model.

As the network depth and model complexity increase, which makes model training difficult, the channel attention mechanism gradually appears in the public’s field of vision. In recent years, SENet has attracted great interest in introducing information channel attention into convolution blocks, showing great potential for performance improvement. Later, research was carried out to improve SE modules by capturing more complex channel dependencies or combining additional spatial attention, but the higher the model accuracy, the higher the complexity, the higher the computational cost, and the higher the computational cost. The research shows that the dimensionality reduction operation adopted by SENet will have a negative impact on the prediction of channel attention, and the efficiency of obtaining dependencies is low. Based on this, an efficient channel attention (ECA) module for CNN is proposed [25], which avoids dimensionality reduction while effectively retaining channel feature information. The specific model is shown in Figure 5. The module effectively filters out information that is highly relevant to the target feature by dynamically adjusting the feature weight while suppressing irrelevant or redundant information, enhancing the robustness of the model for complex backgrounds and improving the accuracy of wheat drought stress degree identification.

The process ideas of the ECA model are as follows:

• First enter the feature map, whose dimension is H × W × C;
• Spatial feature compression of the input feature map; implementation: in the spatial dimension, use the global average pooling GAP to obtain a 1 × 1 × C feature map;
• Perform channel featured learning on the compressed feature map; implementation: learn the importance between different channels through 1 × 1 convolution, and the output dimension is still 1 × 1 × C at this time;
• Finally, the channel attention is combined, the feature map of the channel attention is 1 × 1 × C, the original input feature map H × W × C, the channel-by-channel multiplication is carried out, and the final output is a feature map with channel attention.

The dynamic convolutional kernel in the figure refers to the following process: the size of the convolutional kernel is adaptively changed by a function; in the layer with a large number of channels, use the larger convolutional kernel to carry out 1 × 1 convolution to make more cross-channel interactions; in the layer with a small number of channels, use the smaller convolutional kernel to carry out 1 × 1 convolution to make fewer cross-channel interactions. The convolution and AdaGrad functions are defined as follows:

$$k=\psi (C)={\left|{\displaystyle \frac{{\log }_2(C)}{\gamma }}+{\displaystyle \frac{b}{\gamma }}\right|}_{\mathrm{odd}}$$

(1)

In the above formula, $k$ represents the AdaGrad kernel size; $C$ represents the number of feature channels; $\mathrm{odd}$ indicates that $k$ can only take odd numbers; $\gamma$ and $b$ are set to 2 and 1 in the experiment to change the ratio between the number of channels $C$ and the sum of convolutional kernel sizes.

2.5. Improved DenseNet Model

This paper proposes an improved DenseNet 121 diagnostic model as shown in Figure 6. The network model consists of three parts: feature extraction layer, ECA-NET layer, and wheat drought stress degree classification layer.

In order to optimize the applicability of the DenseNet121 model in the static image identification task for the degree of drought stress in wheat, this paper makes necessary adjustments to the architecture of the model. The original DenseNet121 model is composed of four dense blocks, and the number of convolution layers inside each dense block is 6, 12, 24, and 48 in order to achieve effective classification of 1000 types of images. However, the goal of this paper is only to identify 15 kinds of drought categories. Directly applying the DenseNet121 model may lead to insufficient training of network parameters, which may lead to over-fitting phenomenon. In order to avoid this problem, this paper simplifies the model, reducing the number of dense blocks to 3 and uniformly adjusting the number of convolution layers in each dense block to 8.

In addition, in order to further improve the performance of the model, this paper introduces the attention mechanism to build an improved DenseNet121 model. The model contains 3 optimized dense blocks and 3 transition layers, and a channel attention module ECANet is integrated between each dense block and the transition layer to enhance the model’s ability to capture features. Finally, the model integrates features through a global average pooling layer and classifies them through the output layer. The specific parameter configuration of the model is shown in

Table 4. Improved network parameters of DenseNet121 model.

Layers	Input Size	Size	Output Size
Multiscale convolution layer	48 × 48 × 1	$\left[\begin{array}{cc}1\times 1 conv& 5\times 5 conv\\ 3\times 3 conv& 3\times 3 conv\end{array}\right]$ × 20	48 × 48 × 80
Dense Block 1	48 × 48 × 80	$\left[\begin{array}{cc}1\times 1& conv\\ 3\times 3& conv\end{array}\right]$ × 8	48 × 48 × 336
ECANet	48 × 48 × 336	——	48 × 48 × 336
Transition Layer 1	48 × 48 × 336	$\left[\begin{array}{cc}1\times 1& conv\\ 2\times 2& conv\end{array}\right]$	24 × 24 × 160
Dense Block 2	24 × 24 × 160	$\left[\begin{array}{cc}1\times 1& conv\\ 3\times 3& conv\end{array}\right]$ × 8	24 × 24 × 416
ECANet	24 × 24 × 416	——	24 × 24 × 416
Transition Layer 2	24 × 24 × 416	$\left[\begin{array}{cc}1\times 1& conv\\ 2\times 2& conv\end{array}\right]$	12 × 12 × 208
Dense Block 3	12 × 12 × 208	$\left[\begin{array}{cc}1\times 1& conv\\ 3\times 3& conv\end{array}\right]$ × 8	12 × 12 × 464
ECANet	12 × 12 × 464	——	12 × 12 × 464
Transition Layer 3	12 × 12 × 464	$\left[\begin{array}{cc}1\times 1& conv\\ 2\times 2& conv\end{array}\right]$	6 × 6 × 464
GAP	6 × 6 × 464	6 × 6 pool	464
FC	464	7	7

.

3. Selection of Basic Deep Learning Mode

3.1. Model Test Environment and Evaluation Metrics

3.1.1. Test Environment

The configuration of the computer used for the test is shown in

Table 5. Configuration of the computer.

Configuration	Information
Operating System	Windows 11
RAM	16 GB
CPU frequency	3.20 GHz
GPU	NVIDIA GeForce RTX 3060
GPU memory	6 GB

.

3.1.2. Model Evaluation Metrics

In the process of model training, the values of hyperparameters like optimizers and learning rates have to be continuously adjusted. While choosing the convolutional neural network model, the best-performing model among various sets of hyperparameters is used as the final model. This will ensure that the model performs at its optimum. In this research, the winter wheat drought stress identification and classification model is comprehensively evaluated by multiple indicators, including accuracy of drought stress identification, precision of drought stress classification, and comprehensive evaluation indicator F1 value. A₁ reflects the accuracy of drought identification, P₁ reflects the result of classification, and F₁ is the harmonic mean of precision and recall, which is used to evaluate a model’s accuracy of wheat drought image recognition.

(1)

Accuracy (A₁) refers to the proportion of samples correctly classified by the classifier to the total number of samples. Its formula is as follows:

$${\mathrm{A}}_1={\displaystyle \frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}}$$

(2)

(2)

Precision (P₁) refers to the proportion of actual positive samples in the samples predicted to be positive for each category. Its formula is as follows:

$${\mathrm{P}}_1={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}}$$

(3)

(3)

Recall (R₁) represents the proportion of positive samples in each category predicted as positive. The formula is as follows:

$${\mathrm{R}}_1={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}$$

(4)

(4)

The F1 value is a comprehensive evaluation indicator. It is the harmonic mean of precision and recall. A higher F1 value indicates better classifier performance. The F1 value formula is as follows:

$$\mathrm{F}1={\displaystyle \frac{{2\mathrm{P}}_1{\mathrm{R}}_1}{{\mathrm{P}}_1+{\mathrm{R}}_1}}$$

(5)

The confusion matrix for determining whether the classification of winter wheat is correct based on the recognition model is shown in

Table 6. The confusion matrix for determining whether the classification of winter wheat is correct based on the recognition model.

Label Category	Model Prediction
		0	1
Truth Label	0	True positive (TP)	False negative (FN)
	1	False positive (FP)	True negative (TN)

Table 6: TP is the number of positive examples predicted by the model to be positive. TN is the number of negative examples predicted by the model to be negative. FP is the number of actual positive examples predicted as negative. FN is the number of actual negative examples predicted as positive.

.

3.2. Training Method of Deep Learning Model

Convolutional neural networks require extensive training to seek optimal hyperparameters. With different hyperparameters as variables, multiple control experiments are conducted. The design is as follows: First, three convolutional neural networks, ResNet-101, DenseNet-121, and AlexNet, are employed to build a wheat drought stress classification identification model, and the model is trained from scratch. During the training process, all data is divided into training and test sets, randomly distributed at an 8:2 ratio, where 80% of the samples are used for model training, and 20% for model testing. During model training, the learning rate decay method is set to fixed (Fixed), with a learning rate of 0.001, optimizer set to Stochastic Gradient Descent (SGD), and a batch size of 100. Firstly, the model’s convergence is judged by comparing the model’s loss function, and at the same time, by comprehensively comparing indicators such as recognition accuracy and F1 value, the basic network model with the best generalization ability on this dataset is selected. Secondly, after selecting the basic network model, a total of eight combination experiments are carried out with training methods, learning rate, and attention mechanism as variables, comprehensively comparing the model’s recognition accuracy and F1 value indicators, to find the most suitable convolutional neural network model and corresponding hyperparameters

3.3. Training Method of Deep Learning Model

In this section, ResNet-101, DenseNet-121, and AlexNet convolutional neural networks are used to build the wheat drought stress classification identification model, respectively. The model’s generalization capability is judged by accuracy and F1 value, and the model’s convergence is verified by the loss function, thereby selecting the network model with the best generalization capability on this dataset as the basic model. The experimental results of the evaluation indicators of the three deep learning models are shown in

Table 7. Basic experimental results of three network models based on test sets.

Evaluation Metrics	ResNet101	DenseNet121	AlexNet
Drought Identification Accuracy	74.80%	87.60%	55.83%
Drought Classification Precision	79.10%	88.36%	63.38%
Drought Identification Recall	74.79%	87.59%	55.83%
F1 value	0.7173	0.8729	0.5326

. Figure 7 displays the loss values, accuracy, and F1 values of ResNet-101, DenseNet-121, and AlexNet convolutional neural networks on the test set.

From Table 6, it can be observed that, in terms of drought identification accuracy, the DenseNet-121 model has the highest precision with an accuracy of 87.60% on the test set. ResNet-101 and AlexNet models have accuracies of 74.80% and 55.83%, respectively. Judging from the model’s loss function and F1 value, the convergence speed of the DenseNet-121 model is faster than that of the ResNet-101 and AlexNet models, and the ResNet-101 model exhibits overfitting. The F1 value of the DenseNet-121 model is 0.8729, higher than the 0.7173 and 0.5326 of the ResNet-101 and AlexNet models, respectively. In a comprehensive comparison, the DenseNet-121 network model is chosen as the experimental basic network model.

4. Model Optimization Strategies and Result Analysis

4.1. Model Optimization Strategy

This section is based on the DenseNet121 network model. Using the training method of the convolutional neural network model, the changing condition of the learning rate, and the addition of the attention mechanism as variables, eight combination experiments were conducted in total. The results of the training are summarized in

Table 8. DenseNet-121 network model training results.

Parameters	Transfer Learning	Gradient Learning Rate	Attention Mechanism	Accuracy	Loss Value	F1 Score
Experiment 1	×	×	×	75.87%	2.3360	0.7581
Experiment 2	×	×	√	79.87%	2.2569	0.7894
Experiment 3	×	√	×	82.33%	0.9301	0.8110
Experiment 4	×	√	√	87.00%	0.8766	0.8536
Experiment 5	√	×	×	90.27%	0.7059	0.9015
Experiment 6	√	×	√	91.20%	0.9521	0.9087
Experiment 7	√	√	×	93.87%	0.9274	0.9383
Experiment 8	√	√	√	94.67%	0.8794	0.9438

.

4.2. Result Analysis

4.2.1. Impact of Transfer Learning on the Model

To explore the impact of transfer learning on model training, several comparison experiments were designed under the same learning rate and attention mechanism. Specifically, Experiments 1, 2, 3, and 4 were compared with Experiments 5, 6, 7, and 8. The comparative results of Experiment 1 and 5 and Experiment 4 and 8 are shown in Figure 8. As shown, the accuracy rates of Experiment 1 and 5 are 75.87% and 90.27%, respectively, and the F1 scores are 0.7581 and 0.8535. The accuracy rates of Experiment 4 and 8 are 87.00% and 94.67%, respectively, with F1 scores of 0.8536 and 0.9438. The accuracy and F1 scores of models using transfer learning techniques both improved. This indicates that the following: ① fine-tuning all layers performs better; ② compared to fine-tuning only the last layer, models that fine-tune all layers achieve a significant increase in accuracy. This suggests that the dataset used in this study differs significantly from the ImageNet dataset in terms of feature characteristics. Fine-tuning all layers can integrate feature learning, share knowledge, and provide a larger network capacity, adapting better to the target task’s requirements.

4.2.2. Impact of Gradient Learning Rate on the Model

To investigate the impact of the change in learning rate on the model training outcomes, under the conditions where the transfer learning method and attention mechanism remain consistent, several control experiments were designed, namely Experiments 1, 2, 5, and 6 and Experiments 3, 4, 7, and 8. The comparison results of Experiments 1 and 3 and Experiments 6 and 8 are shown in Figure 9. From the figure, it can be seen that the accuracy of Experiment 1 and Experiment 3 are 75.87% and 82.33%, respectively, with F1 scores of 0.7581 and 0.8110, respectively. The accuracy of Experiment 6 and Experiment 8 are 91.20% and 94.67%, respectively, with F1 scores of 0.9087 and 0.9438, respectively. The accuracy and F1 values of the model have been improved after using the gradient learning rate. The following can be concluded: ① The model using the gradient learning rate has slightly higher recognition accuracy than the one with a constant learning rate. In Experiment 1 and Experiment 3, the average recognition accuracy increased significantly. This indicates that the gradient learning rate method can effectively manage and adjust the learning speed of the model during the training process, allowing the model to use a larger learning rate in the initial stage for faster convergence and gradually reduce the learning rate in the subsequent stages. This enables the model to fine-tune its parameters more delicately, thereby enhancing its generalization performance. In contrast, the model with a fixed learning rate may experience overfitting or underfitting in the early or later stages, resulting in subpar overall performance. ② The gradient learning rate method can better balance the model’s learning speed at different training stages. In deep learning models, the learning rates of different layers may have different impacts on model performance. The gradient learning rate method can adjust the learning rates of different layers according to the specific task and network structure, allowing the model to learn and update the parameters of each layer more effectively, thereby enhancing the model’s expressiveness. In contrast, the model with a fixed learning rate may not fully utilize the hierarchical features of the network structure, thus affecting the recognition accuracy. In this section, the initial learning rate is set to 0.001, and the learning rate is halved every 10 epochs.

4.2.3. Impact of the Attention Mechanism on the Model

To investigate the impact of the attention mechanism on the model training results, under the conditions where the transfer learning method and the learning rate remain consistent, several control experiments were designed, namely Experiments 1, 3, 5, and 7 and Experiments 2, 4, 6, and 8. The comparison results of Experiments 1 and 2 and Experiments 5 and 6 are shown in Figure 10. From the figure, it can be seen that the accuracy of Experiment 1 and Experiment 2 are 75.87% and 79.87%, respectively, with F1 scores of 0.7581 and 0.7894, respectively. The accuracy of Experiment 5 and Experiment 6 are 90.27% and 91.20%, respectively, with F1 scores of 0.9015 and 0.9087, respectively. The introduction of the attention mechanism has improved the model’s accuracy and F1 values. The following can be concluded: ① Introducing the attention mechanism module can weight the importance of winter wheat to enhance the model’s focus on its features and improve the model’s ability to perceive the importance of different features. In deep learning models, the quality and importance of feature representations have a significant impact on the performance of the task. By introducing the attention mechanism, the model can dynamically adjust the weights and contributions of the features, allowing the model to focus more on key information and enhance the expression of the features. In contrast, models that do not incorporate an attention mechanism may fail to accurately distinguish the importance of different features, leading to suboptimal performance. ② Introducing the attention mechanism can enhance the model’s perception of the importance of different time steps or spatial positions. In deep learning models, for sequential problems or tasks with spatiotemporal structures, the information from different time steps or spatial positions has varying importance. By using the attention mechanism, the model can dynamically adjust the importance weights of different time steps or spatial positions, better capturing the features of sequences or spatiotemporal patterns. Compared to models that do not use attention mechanisms, models that incorporate attention can better leverage temporal or spatial information, enhancing the performance of deep learning models.

4.3. Experimental Results Validation

The confusion matrix is one of the factors used to evaluate the model. It is a heatmap in matrix form, where rows represent the actual category labels and columns represent the predictions. The confusion matrices for the results of Experiment 1, Experiment 4, and Experiment 8 are derived through calculations, as shown in Figure 11. The horizontal and vertical coordinates in the figure are composed of abbreviations with prefixes and suffixes, where the prefix represents the growth cycle of winter wheat, such as rising–jointing stage (RJ), heading–flowering stage (HF), and flowering–maturity stage (FM). Suffixes represent drought degrees, such as optimum moisture (OM), light drought (LD), moderate drought (MD), severe drought (SD), and extreme drought (ED). Analyzing Figure 9, we can discern the following: ① In the three different growth stages of wheat, certain stages are mistakenly identified as other stages. For instance, the “RJ” stage is sometimes wrongly identified as the “HF” stage. This is because the growth of wheat is a continuous process. As it transitions from the “RJ” stage to the “HF” stage, features like stem, leaves, and texture do not undergo significant changes in comparison to the “HF” stage. ② Adjacent drought levels within the same growth cycle are also easily confused. For instance, the appropriate state might be mistakenly determined as mild drought, mild drought may be wrongly classified as either appropriate or moderate drought, moderate drought might be mistakenly identified as mild or severe drought, and severe drought may be mistakenly identified as extreme drought. This is because drought stress is a continuous, dynamic accumulation process where the phenotypic differences of wheat plants under adjacent drought levels are minimal with unclear boundaries. ③ With continuous model improvements, misjudgments are alleviated. Especially evident from the confusion matrix of Experiment 8, misjudgments are limited to the “RJ” stage. This is due to the minimal physiological and morphological phenotypic differences of wheat under drought stress during this stage, leading to potential misidentification. Therefore, by utilizing transfer learning, gradual learning rate adjustments, and the introduction of the attention mechanism, this section aims to enhance the drought identification accuracy of winter wheat, thereby refining the ability to determine drought stress during the wheat growth process.

The Receiver Operating Characteristic (ROC) is one of the factors used to evaluate the model. The X-axis of the horizontal coordinate of the ROC curve is the False Positive Rate (FPR), also known as the False Diagnosis Rate (FDR). As shown in Equation (5), the closer the X-axis is to zero, the higher the accuracy; the vertical coordinate Y-axis is the True Positive Rate, TPR, also known as the sensitivity, and as shown in Equation (6), a larger Y-axis represents a higher accuracy. AUC (Area under Curve, AUC) denotes the area under the ROC curve, which is mainly used to measure the model generalization performance, i.e., how well the classification is conducted. It indicates the probability that positive examples are ranked in front of negative examples. Generally in classification models, the prediction results are expressed in the form of probability.

$$\mathrm{FPR}={\displaystyle \frac{\mathrm{FP}}{\mathrm{FP}+\mathrm{TN}}}$$

(6)

$$\mathrm{TPR}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}$$

(7)

This paper utilizes the AUC metric to reveal significant improvements in the false positive rate for the enhanced model compared to its predecessor. The enhanced model demonstrates a substantial reduction in false positives without compromising sensitivity, indicating a marked enhancement in specificity.

A comparative assessment of AUC scores across various experiments highlights the effectiveness of the improvements made in this paper. Notably, in Experiments 1 and 8, which focus on the RJ-MD classification scheme, we observed a remarkable 15% increase in AUC. This increase signifies a substantial enhancement in the model’s ability to differentiate between positive and negative cases, reflecting an overall improvement in diagnostic accuracy.

Furthermore, upon examining the performance of various improved methods, several stood out. The transfer learning approach, which involved fine-tuning the training of all layers, the model employing an asymptotic learning rate strategy, and the one integrated with an attention mechanism all performed exceptionally well in terms of AUC and specificity, suggesting that these enhancements are crucial for optimizing model performance.

To provide a comprehensive understanding, this paper conducts an in-depth analysis of the ROC curves for Experiments 1, 4, and 8, as shown in Figure 12. We analyzed the trade-off between the true positive rate and the false positive rate at various threshold settings. This detailed examination allows us to evaluate the models’ performance across a spectrum of classification decisions, offering insights into their robustness and reliability in different scenarios.

In conclusion, the enhanced model outperforms the pre-improved model across all evaluated metrics. The improvements are not only statistically significant but also practically meaningful, providing a solid foundation for further research and application in the field of winter wheat drought stress monitoring.

In this section, we explored the impacts of transfer learning methods and learning rates on model performance and investigated the effects of the attention mechanism on model training results. Experimental results show that fine-tuning all layers rather than just the final layer can further enhance model performance; models using a gradual learning rate method slightly outperform those with a fixed learning rate, indicating that the gradual learning rate method effectively manages and adjusts the learning rate of the model. Introducing an attention mechanism to models slightly improves recognition accuracy over models without it. This is because the attention mechanism enhances the model’s weighted importance of wheat features and focuses on key information, improving the model’s perception of importance across different features, time steps, and spatial positions. In this section, by leveraging the feature extraction capabilities of pre-trained models, dynamically adjusting learning rates, and enhancing attention to vital information and its importance, we aim to elevate the detection performance, accuracy, and generalization ability of the model. These strategies, applied across diverse tasks and network structures, can significantly boost the generalization capabilities of deep learning models. The integrated application of these methods in this section can improve the training results and generalization ability of the model, speed up model convergence, and enhance performance metrics such as recognition accuracy. Specifically, the test set’s average recognition accuracy for Experiment 8 reached 94.67%, demonstrating promising potential in recognizing drought levels during critical wheat growth stages.

5. Conclusions

This paper proposes a deep learning-based drought monitoring model for the growth process of winter wheat, which aims at identifying and classifying different degrees of drought stress during winter wheat’s growing period. In this paper, we focus on three key growth periods in winter wheat, namely RJ, HF, and FM, for drought stress. In these periods, we collected winter wheat drought-stress images from the field conditions and established a drought image set corresponding to soil moisture monitoring data. By comparing the drought stress recognition accuracy, classification accuracy, recall rate, and F1 value of the AlexNet, ResNet-101, and DenseNet-121 network models, the DenseNet-121 network model was chosen as the base model for this purpose. Building upon the selected DenseNet-121, we developed training and optimization strategies for the model, considering improvements like the use of transfer learning, the adoption of a gradual learning rate, and the introduction of the attention mechanism method. Eight experimental groups were carried out. The results indicate that through these improvements, the model’s convergence speed is accelerated and performance metrics, including recognition accuracy, are enhanced. This effectively improves the identification and classification of drought stress levels during the key growth stages of winter wheat. Notably, the test set’s average recognition accuracy for Experiment 8 reaches 94.67%.

A comprehensive analysis reveals that deep learning algorithms provide a more reliable and accurate method for identifying and classifying the degree of wheat drought. While advancements have been noted in the study of drought stress phenotypes, the diagnosis of crop drought stress based solely on a single phenotypic trait encounters notable challenges. Factors such as fluctuating weather conditions and variations in lighting can result in image overexposure or the loss of critical details, thereby impeding the precise identification of drought indicators. Furthermore, the prevalence of pests and diseases in agricultural settings, which can mimic the symptoms of drought stress, compounds the potential for erroneous assessments by image-based monitoring systems. To transcend these limitations, it is imperative for future research to embrace the integration of data from multiple sensors to encapsulate a holistic view of crop phenotyping. This encompasses a range of attributes—color, texture, morphology, and physiological parameters—that define the crop’s phenotype. The fusion of this rich, multi-dimensional dataset with sophisticated pattern recognition algorithms paves the way for the development of more precise and non-destructive diagnostic tools for the rapid monitoring of crop drought stress. This integrated approach is anticipated to markedly enhance the fidelity and dependability of monitoring efforts, thereby affording robust decision support for agricultural practices.