Agricultural non-point sources, as major sources of organic pollution, continue to flow into the river network area of the Jiangnan Plain, posing a serious threat to the quality of water bodies, the ecological environment, and human health. Therefore, there is an urgent need for a method that can accurately identify various types of agricultural organic pollution to prevent the water ecosystems in the region from significant organic pollution. In this study, a network model called RA-GoogLeNet is proposed for accurately identifying agricultural organic pollution in the river network area of the Jiangnan Plain. RA-GoogLeNet uses fluorescence spectral data of agricultural non-point source water quality in Changzhou Changdang Lake Basin, based on GoogLeNet architecture, and adds an efficient channel attention (ECA) mechanism to its A-Inception module, which enables the model to automatically learn the importance of independent channel features. ResNet are used to connect each A-Reception module. The experimental results show that RA-GoogLeNet performs well in fluorescence spectral classification of water quality, with an accuracy of 96.3%, which is 1.2% higher than the baseline model, and has good recall and F1 score. This study provides powerful technical support for the traceability of agricultural organic pollution.

  • The proposed RA-GoogLeNet is time-efficient, accurate, and superior to other convolutional neural network models with fewer parameters than traditional methods.

  • ECA attention mechanism and ResNet were added to improve the model's ability to identify agricultural non-point source pollution in the study area.

  • The Leaky ReLU activation function used in this model can extract more image features.

In recent years, China's efforts to strengthen pollution control in water environment management have led to improvements in agricultural surface pollution (Guo et al. 2014; Wang et al. 2019). However, this pollution is still a major problem and pain point for water environment quality improvement. Agricultural surface source pollution (Guo et al. 2014; Wang et al. 2019), which mainly consists of nutrients such as nitrogen and phosphorus, pesticides and other harmful substances, solid wastes such as straw and agricultural film, livestock and poultry farming manure and sewage, aquaculture bait and drugs, and sewage and garbage from rural life, is caused by pollution of the water environment in the forms of surface runoff, soil erosion, and farmland drainage, which is decentralised and hidden. Despite the rapid development of modern high-quality agriculture, the problem of agricultural surface pollution should not be ignored. This mainly includes pollution caused by aquaculture (e.g., shrimp ponds, fish ponds, crab ponds) and farmlands (e.g., rice cultivation), which include activities that have great impacts on water quality, biodiversity, and human health. Therefore, a method that can accurately identify agricultural non-point source pollution is particularly important. It can not only prevent the spread of non-point source pollution in time and play an important role in achieving the balance between economic development and ecological environment (Cui et al. 2020) but also provide effective support needed for water environment management (Xiao et al. 2023).

Nowadays, for the classification of agricultural pollutants, their conventional pollutants, such as TN, BOD (Ooi et al. 2022; Liu & Chen 2023), and so on, are usually measured first and then their 3D fluorescence is determined. However, most of them use traditional methods to analyse the three-dimensional excitation-emission matrix spectra (3D-EEMs) of water quality, including regional fluorescence integration (Chen et al. 2003), principal component analysis (Guimet et al. 2004; Wang et al. 2022), and parallel factorial (PARAFAC) analysis (Wünsch et al. 2017), which are usually used to analyse the water quality from the 3D-EEMs to mine meaningful information (Zhang et al. 2018). For example, the PARAFAC model is increasingly used to extract hidden fluorescence components in 3D-EEMs. PARAFAC not only estimates the number of components in a fluorescence spectrum but also identifies different components and their origins by the positions and intensities of the fluorescence peaks. However, analysis using PARAFAC is time-consuming and complex. A large amount of fluorescence spectral data is required to determine the peak intervals of different classes of fluorescent components, which can ensure the accuracy of the subsequent classification of fluorescent components. These issues affect the rapid identification and analysis of fluorescence spectra by PARAFAC.

With the continuous development of artificial intelligence (Erguzel et al. 2015), convolutional neural networks (Huang et al. 2024) act as a powerful deep learning method for processing images (Su & Zhang 2023). These deep learning models use raw image data as input to adaptively develop optimal representations rather than relying on human feature extraction (Lim et al. 2022). They have also been increasingly applied to medical and industrial fields in recent years, such as Medical image classification (Tolkach et al. 2020; Gu et al. 2021; Cao et al. 2023; Du et al. 2023), fault detection (Minu & Canessane 2022), and pest and disease recognition (He et al. 2023). In contrast to traditional neural networks, which are three-dimensional matrices, it is essentially a multilayer perceptual machine (Rahman et al. 2020), and its success is due to its use of local connectivity and weight sharing, which makes the network easy to optimise and reduces the complexity of the model. The 3D-EEM spectrum consists of excitation × emission × intensity, similar to the structure of grey scale image data (height × width × grey scale values). Therefore, many machine learning and deep learning algorithms (Wang et al. 2021) have been applied to 3D fluorescence data classification studies and have confirmed the feasibility of such algorithms. For example, Xu et al. (2022) proposed rapid identification of 3D-EEMs by a continuous convolutional neural networks (CNN), rapid prediction of the number of fluorescent components, and component fitting by inputting a single 3D-EEM spectrum, which is used for online analysis of wastewater and rapid monitoring of various water bodies. Ruan (2022) proposed VGG16 (Nijaguna et al. 2023) and VGG11 for classification and component fitting of three-dimensional fluorescence spectra, which can achieve almost the same results as PARAFAC, and the experimental conditions are less demanding than those of PARAFAC. In addition, for the first time, Cuss et al. (2016) used four machine learning approaches, such as decision trees, k-nearest neighbours, multilayer perceptrons, and support vector machines, and combined them with PARAFAC to classify the resulting fluorescent components just by osmotic fluid/river, concluding that the support vector machines had the best classification results. The previous methods of fluorescence spectral (Yadav et al. 2019) analysis using machine learning, although improved in terms of efficiency over traditional methods, had a higher overall experimental time complexity.

The objectives of this study are as follows: (1) to identify a pollution source identification model suitable for Changdang Lake Basin in view of the problems of the existing methods; and (2) to find a method that can more accurately identify the source of pollution in a shorter period of time and block pollution in a timely manner, taking into account the disadvantages of time-consuming and labour-intensive traditional methods. In this study, a GoogLeNet-based convolutional neural network model (RA-GoogLeNet) was used to rapidly identify agricultural surface source pollution in fish, shrimp, crab, and farmland around Changdang Lake. 3D-EEMs collected from major aquaculture ponds were divided into training sets, and the RA-GoogLeNet model was trained and evaluated for performance. Finally, the RA-GoogLeNet model was compared with the base models such as GoogLeNet, VGG, and AlexNet.

Study area and sample collection

Changzhou City, Jiangsu Province (31°09′–32°04′N, 119°08′–120°12′E), with a total area of 4,385 km2, is situated at the confluence of two major water systems, the Yangtze River and the Huaihe River, and is endowed with abundant water resources and a well-developed water system structure. The study area of this experiment is located in Changzhou City, Jiangsu Province. Changdang Lake (31°33′–31°40′N, 119°30′–119°37′E), also known as the Tao Lake, is the third largest lake in the Taihu Lake Basin, which belongs to the typical over-water, shallow grass-type lakes (as shown in Figure 1). Its water area is about 90 km2, about 16.0 km long and 5.6 km wide, with a water depth of 0.8–1.2 m. The lake has a complex topography, with upstream and downstream topography interspersed with mountains, dikes, and ditches. Because of its favourable geographical location and the richness of nutrients in the water, the region has used the high-quality water of Changdang Lake to nourish the numerous breeding ponds and farmlands in the vicinity, promoting the development of local aquaculture and agriculture. The study area calculated by a Geographic Information System (ArcGIS) shows that the area around Changdang Lake is about 870,000 acres, of which about 130,000 acres are farmed ponds and about 450,000 acres are farmlands. At the time of collecting water samples, it was the season of planting rice, and for the farmland aspect, the receding water of rice was collected. Therefore, in this study, a large number of agricultural surface pollution sources around the Changdang Lake, including aquaculture (shrimp ponds, fish ponds, crab ponds), and farmland (rice) were collected.
Figure 1

Land use types around the Changdang Lake.

Figure 1

Land use types around the Changdang Lake.

Close modal

To obtain the fluorescence spectra of agricultural fouling sources around Changdang Lake, this study conducted multiple collections of data sources from aquaculture ponds and agricultural fields in various regions of Changzhou City from July to August in the summer of 2023, and 550 agricultural collection sources were used cumulatively (100 fish ponds, 100 shrimp ponds, 100 crab ponds, and 250 agricultural fields (mainly rice-based)). During field collection, all water samples were placed in brown glass vials and kept at low temperature and darkness in a cryostat, before being transported to the laboratory on the same day and stored in a 4°C freezer.

The 3D-EEM spectra of the samples were measured by a fluorescence spectrometer RF-6000 (Shimadzu, Japan). The main parameters of the device are shown below: excitation wavelength range of 200–500 nm with 5 nm spacing, emission wavelength range of 250–600 nm with 1 nm step spacing, and a scan rate of 12,000 nm. All water samples were filtered using a 0.45-nm glass fibre filter membrane prior to measurement, while ultrapure water was used as a reference. The measurement process for all samples must be completed within 48 h.

Technical approaches

The main objective of this study was to identify the fluorescence spectra of water samples from culture ponds and agricultural fields around the Changdang Lake. By learning the characteristics of the fluorescence spectra of each category, water samples from polluted lakes and rivers around the Changdang Lake were collected for classification and prediction to determine the type of pollution to which the water body was subjected and to provide a basis for subsequent prevention and control of pollution sources. The specific process (as shown in Figure 2) consists of two main modules: data pre-processing module and model construction and presentation of classification results module. In the data pre-processing stage, the collected back data need to be measured and a parallel factor analysis is performed to eliminate noise and other scattering interferences. Next, the dataset is expanded and the images are resized to the size required for network training. For model construction, we used GoogLeNet and added the ECA attention mechanism and residual network (ResNet). The attention mechanism is used to adaptively adjust the weights of the channels to enhance attention to the intensity and peaks of the fluorescence spectrum to improve network performance. The ResNet is introduced to address the problem of vanishing gradients and loss of information in the fluorescence spectrum. Finally, we will show the classification results, which will be displayed at the top of the image.
Figure 2

Technology roadmap.

Figure 2

Technology roadmap.

Close modal

PARAFAC method

After the fluorescence spectra have been measured and analysed by PARAFAC, all measured EEM spectra must be pre-processed prior to analysis, taking into account the effects of Raman and Rayleigh scattering on the fluorescence components, prior to modelling for PARAFAC. The pre-processing is mainly to minimise the effect of Rayleigh and Raman scattering by subtracting a blank (pure water) from the spectra. To facilitate PARAFAC modelling and analysis, outliers in the samples need to be removed; outliers are mainly samples whose leverage is significantly different from the rest of the samples (leverage is calculated by installing the DOM flour toolbox in MATLAB). The 3D fluorescence dataset was modelled using PARAFAC in MATLAB R2021a software. It uses alternating least squares to minimise the sum of squares of the residuals of the cubic matrix model, reducing the 3D fluorescence dataset to a cubic matrix and a residual matrix as shown in Equation (1).
formula
(1)
where ; ; k indicates the number of the sample; is an element of the cubic array with group fraction Z representing the fluorescence intensity of sample k at excitation wavelength i and emission wavelength j; Z is the group fraction; , and are the elements of the three base profile matrices and of X; denotes an element of the model unfit three-way residual matrix E.

Characteristics of 3D-EEM data

The data used in this study are fluorescence spectra of agricultural surface sources. Fluorescence spectra are plots obtained by projecting the fluorescence intensity as contours on a plane with excitation and emission wavelengths as vertical and horizontal coordinates. The fluorescence spectra vary according to the type and content of organic matter and have a one-to-one correspondence with the water sample. Since the composition of fluorescence spectra is similar to the grey scale image structure, a combination of deep learning and fluorescence spectra is used. The essence of deep learning for fluorescence spectral feature extraction is the delineation of fluorescence peak positions and intensities of fluorescence spectra. As shown in Figure 3, due to the different types of organic matter inside different water samples, it will lead to a significant difference in the position and intensity of the fluorescence peaks in the fluorescence spectra of different categories of pollution sources. In addition, the position and intensity of the fluorescence peaks in the fluorescence spectra of the same category of water samples collected in different places are slightly deviated but roughly the same due to the influence of a number of external factors, including errors caused by the instrumentation during the measurements, the differences in the specific locations of the collected water samples, and so on.
Figure 3

Fluorescence spectra of different water pollution sources.

Figure 3

Fluorescence spectra of different water pollution sources.

Close modal

Image processing

The image size of the database created in this study is 1,024 × 1,024 × 3. However, due to the large size of the image, it will lead to slow model training if it is placed directly in the network, so the size of the image will be set to 224 × 224 × 3 in the experimental imported images. Second, to ensure the adequacy of the samples required for training the convolutional neural network model, the experimental data were subjected to data augmentation (Wei et al. 2022), such as flipping, shifting, and colour enhancement. The number of images before and after data expansion is shown in Table 1.

Table 1

Pollution source data sets

Categories of pollution sourcesBefore expansionAfter expansion
Crab pond 100 800 
Paddy field 250 800 
Fish pond 100 800 
Shrimp pond 100 800 
Categories of pollution sourcesBefore expansionAfter expansion
Crab pond 100 800 
Paddy field 250 800 
Fish pond 100 800 
Shrimp pond 100 800 

The dataset by extension has 800 images per class and 3,200 images in total. The dataset is divided into a training set (including validation set) and a test set in the ratio of 8:2. To ensure consistency of data distribution, stratified sampling is used. After dividing the training set and test set, the number of data in each class is 640 and 160, respectively.

To better distinguish each category of fluorescent fingerprint images, the images of the dataset were labelled, here one-hot coding was used, with 0, 1, 2, and 3 representing the four categories of crab ponds, paddy fields, fish ponds, and shrimp ponds, respectively.

Construction of an RA-GoogLeNet model based on CNN

GoogLeNet (Alex et al. 2012) is a new deep learning framework proposed by Christian Szegedy in 2014. The network is constructed based on the Inception network, which is a sparse and high-performance network structure that incorporates feature information at different scales. It is mainly characterised by the ability to use its resources efficiently and obtain high classification accuracy and robustness through fewer parameters. In this study, the attention mechanism and ResNet are implanted on the basis of GoogLeNet, so that the fluorescence spectra of various types of pollutants can be identified more accurately. The model is based on GoogLeNet, and the 7 × 7 convolutional kernel inside the first convolutional layer is firstly replaced with three 3 × 3 convolutional kernels, which not only increases the nonlinear representation ability of the network but also enhances the network's ability to learn about the features, and at the same time reduces the parameters in the model. Secondly, an attention mechanism (A-Inception) is added to the Inception module. Each A-Inception module is connected through a ResNet. In addition to that, also in this model, the original ReLU activation function is replaced by using the Leaky ReLU activation function, and more image features can be extracted because Leaky ReLU can compensate for the problem of the vanishing gradient of the ReLU function. The specific structure of the RA-GoogLeNet model is shown in Figure 4.
Figure 4

Structure of RA-GoogLeNet.

Figure 4

Structure of RA-GoogLeNet.

Close modal

ECA attention mechanism

Attentional mechanisms (Niu et al. 2021) are essentially a set of weighting coefficients learnt autonomously by the network and ‘dynamically weighted’ to emphasise regions of interest while suppressing irrelevant background regions. The current mainstream attention mechanisms can be categorised as follows: channel attention mechanisms, spatial attention mechanisms, branching attention, temporal attention, and hybrid attention.

The traditional squeeze-and-excitation (SE) attention mechanism (Hu et al. 2020) is a pooling operation based on spatial dimensions, which obtains dependencies between different channels through two fully connected layers. This kind achieved good results in convolutional neural networks, but also on the network's computation of the emergence of some problems. The ECA attention mechanism (Shi et al. 2022) is also the channel attention mechanism and is an improvement on SE. This module has a huge potential to improve the performance of deep neural networks; the ECA module removes the fully connected layer of the SE module and learns by a one-dimensional convolution on the globally averaged pooled features, by using a one-dimensional convolution. Then the size of the convolution kernel becomes more important, and the size of the convolution kernel directly affects the performance of the attention mechanism. The specific implementation of one-dimensional convolution to obtain the dependencies between different channels is given by , where denotes the one-dimensional convolution, denotes the Sigmoid activation function, z denotes the global average pooling layer, the module involves only k parameters, and when , the ECA module can achieve similar results as SE-var3, which also employs fewer parameters but ensures the efficiency and effectiveness of the local inter-channel interactions.
formula
(2)
where C is the dimension of the channel and denotes the closest odd distance to m. and b are hyperparameters, which are 2 and 1 in the experiments.
The Inception module in deep neural networks or deep convolutional networks was proposed by Christian Szegedy and others at Google. The core idea of the Inception module is to combine different convolutional layers together by concatenation, and the resultant matrix processed by different convolutional layers is spliced together in the dimension of depth to form a deeper matrix. The success of GoogLeNet is mainly due to the Inception module, and the whole GoogLeNet main architecture can be seen as a stack of multiple Inception modules. Its internal Inception structure is shown in Figure 5(a).
Figure 5

Inception and A-Inception structure.

Figure 5

Inception and A-Inception structure.

Close modal

The main idea of the Inception structure is to approximate the optimal local sparse structure with dense components. The inner structure has a 3 × 3 pooling layer and 1 × 1, 3 × 3, and 5 × 5 convolutional layers, and multiple 3 × 3 convolutional layers are added to the original Inception module, which acts as a dimensionality reduction in the direction of the channel. In this study, the ECA attention mechanism is added on top of the Inception module to become the A-GoogLeNet module. It is actually an average pooling layer and a one-dimensional convolution added on top of the basic module. Finally, the output is made by a Sigmoid activation function. The specific structure is shown in Figure 5(b).

RA-GoogLeNet module

By adding the attention mechanism after the Inception module, it allows for more features to be extracted; however, with the addition of the ECA attention mechanism, also known as the A-Inception module, more layers of the network than before are created. Then as the number of network layers becomes larger, a series of problems such as loss of information, vanishing gradients, and so on, can occur. ResNets (Wang et al. 2023) were created to solve problems such as vanishing gradients or gradient explosion due to the increase in the number of network layers. Therefore, this study mitigates the problem due to the inclusion of an attention mechanism by employing a ResNet. This is implemented by adding a ResNet between every two A-Inception modules. The ResNet does not introduce additional parameters and computational complexity, allowing the features to be passed directly to the output without going through all the convolutions, and therefore does not increase the error even if the number of network layers is increased. The structure is named RA-GoogLeNet module and the connection of the specific ResNet is shown in Figure 6.
Figure 6

ResNet connection diagram.

Figure 6

ResNet connection diagram.

Close modal
The activation function used in this study was also changed from the original ReLU activation function in the GoogLeNet network to the Leaky ReLU (Liu et al. 2019) activation function. Because the ReLU function has a gradient of zero for inputs less than or equal to zero, this can lead to the phenomenon of neuronal death, i.e., once a neuron's output is zero, it will not be able to recover and will not be able to update its weights. This may affect the learning ability and expressiveness of the model. The Leaky ReLU activation function used in this study can be a good solution to this problem, and its expression is shown in the following equation:
formula
(3)
where a is a fixed parameter in the interval . From the equation, Leaky ReLU assigns a non-zero slope to all negative values, while the ReLU function tends to activate rather than die with all negative values set to zero on this interval.

Evaluation indicators

In this study, the effectiveness of this network model for 3D fluorescent fingerprint recognition is evaluated by using accuracy, precision, recall, and F1 score. The formulas for accuracy, precision, recall, and F1 score are shown below.
formula
(4)
formula
(5)
formula
(6)
formula
(7)
where is the number of samples that were actually positive and predicted to be positive, is the number of samples that were actually negative but predicted to be positive, is the number of samples that were actually positive but predicted to be negative, and is the number of samples that were actually negative and predicted to be negative. is the proportion of correctly categorised samples out of the total number of samples. is the proportion of samples predicted positive by the model that are actually also positive to the proportion of samples predicted positive. is the proportion of samples that are actually positive to the proportion of samples predicted positive. is the weighted average of precision and recall.

Spectral pre-processing based on the Delaunay triangulation interpolation method

In general, under the influence of the bias of the spectrometer system, the fluorescence spectra measured directly without correction will have some deviation from the real spectra in terms of peak position and fluorescence intensity. Therefore, the fluorescence spectra of the water samples collected from agricultural surface sources after filtering and scanning by the spectrometer will be affected by Raman scattering and Rayleigh scattering. These scattering phenomena can lead to biased trilinear component models using the PARAFAC algorithm, which can seriously affect the analysis of the fluorescence characteristics of pollution sources. Scattering produces a fluorescence peak that is too high and masks the fluorescence peak of the contamination itself, so the scattering needs to be eliminated before analyses can be performed. To solve this problem, this experiment can effectively remove the interference of scattering by using Delaunay triangulation interpolation. According to Figure 7, it can be observed that in the absence of pre-processing, there is primary and secondary scattering in the fluorescence spectrum, where the primary scattering obscures part of the sign region. After the pre-processing of the 3D fluorescence spectroscopy data, the scattering of the contaminants was successfully removed, allowing the peaks of their own fluorescence characteristics to come to the fore. However, there may be clipped regions in the pre-processed data, resulting in gaps. To solve this problem, we used an interpolated fitting method to fill in the surrounding data.
Figure 7

Original image and processed image.

Figure 7

Original image and processed image.

Close modal

Results of the given model

Effect of different parameters on model performance

The overall performance of a deep learning network architecture depends on the composition of the dataset and the set of parameters used for training and testing, so it is important to choose a network architecture that limits overfitting and underfitting. The magnitude of the network learning rate in the parameters has a significant impact on the recognition performance of the model. Too large a value of the learning rate leads to the phenomenon of oscillation during the search process, which may skip the optimal solution; while too small a value of the learning rate increases the number of times the model learns and the convergence time. Therefore, a higher or lower learning rate will lead to a lower accuracy of training and testing, and an increase in the loss value, which is not favourable to model learning.

In this study, four different sizes of learning rates (0.00001, 0.0001, 0.001, and 0.01) were tried for the experiment. As shown in Figure 8, through several experimental comparisons, it is found that when the learning rate is 0.0001, the validation accuracy of the model is high, the curve tends to be stable, and the loss value continues to decrease and eventually converges without overfitting. However, when the learning rate is 0.001, the accuracy instead decreases and the loss function rises significantly. This suggests that the optimal value of the learning rate is within an interval, and too large or too small values can adversely affect the network performance. Therefore, based on the experimental results, it can be concluded that the present model is best for the recognition of fluorescence spectra of agricultural pollutants at a learning rate of 0.0001.
Figure 8

Learning curves for training accuracy at different learning rates.

Figure 8

Learning curves for training accuracy at different learning rates.

Close modal

Results of the model after parameterisation

For the model proposed in this study, various hyperparameters were used for training, which include having an epoch, a learning rate, a batch size, and an optimiser. The optimal results of the model were obtained through continuous attempts. The model is optimal for recognition of fluorescence spectra from agricultural surface sources when using the Adam optimiser with an epoch of 150, a learning rate of 0.0001, and a batch size of 32. The learning curve of the model in identifying fluorescence spectra of agricultural pollutants with specified hyperparameters is shown in Figure 9. These two graphs show that there are some fluctuations in the early stages of training, but they gradually level off after an epoch of 90, and the model's learning becomes smoother with no overfitting. The experimental results show that the model has an accuracy of 97.9% for the training set and 97.3% for the validation set.
Figure 9

Accuracy and loss function for training and validation of the model after determining the parameters.

Figure 9

Accuracy and loss function for training and validation of the model after determining the parameters.

Close modal
To get a clearer picture of the model's performance in classifying over the data domain, this study presents it in the form of a confusion matrix, as shown in Figure 10. Rice has the lowest error rate during the classification process with only seven samples misclassified. As the fluorescence spectra of various types of farmed sources have similarities in the position of the fluorescence peaks, resulting in more misidentified samples during the training phase, most of the misclassifications shown in the confusion matrix are for farmed sources, with only a small number of misclassifications for rice.
Figure 10

Confusion matrix of the proposed model.

Figure 10

Confusion matrix of the proposed model.

Close modal
In this study, the weights obtained from training the RA-GoogLeNet network model were used to predict the fluorescence spectra of four randomly selected agricultural pollutants, and the prediction results reached 100% accuracy, indicating that the improved RA-GoogLeNet can better identify the fluorescence spectra of agricultural pollutants. The prediction results of the fluorescence spectra of specific agricultural pollutants are shown in Figure 11.
Figure 11

Predicted results of fluorescence spectra of pollutants.

Figure 11

Predicted results of fluorescence spectra of pollutants.

Close modal

Performance comparison of different models

Comparison experiments using different attentional mechanisms

With the continuous development of deep neural networks over the years, the attention mechanism module has gradually become an additional neural network that is widely used in major fields and has been shown to be used to improve the performance of models. The attention mechanism can dynamically generate weights for different connections and these weights are variable. The three commonly used attentional mechanisms to be used in this study were compared, including SE, ECA, and convolutional block attention module (CABM) (Qu et al. 2023), and the effects of different attentional mechanisms on the model performance were tested while keeping other hyperparameters constant. Table 2 demonstrates the performance comparison of networks using these three attention mechanisms. The experimental results show that the identification of pollutant fluorescence spectra is optimal when using the ECA attention mechanism, whereas the results are similar using SE and CABM, and ECA is more suitable for GoogLeNet.

Table 2

Experimental comparison of different attention mechanisms

Batch sizeEpochActivation functionAccuracy (%)
RE-GoogLeNet 32 150 Leaky ReLU 95.7 
RM-GoogLeNet 32 150 Leaky ReLU 95.8 
RA-GoogLeNet 32 150 Leaky ReLU 96.3 
Batch sizeEpochActivation functionAccuracy (%)
RE-GoogLeNet 32 150 Leaky ReLU 95.7 
RM-GoogLeNet 32 150 Leaky ReLU 95.8 
RA-GoogLeNet 32 150 Leaky ReLU 96.3 

Confusion matrix (Li et al. 2004) is a way to show the performance of the network. Figure 12 shows the confusion matrices of the models for the three attentional mechanisms described above (RE-GoogLeNet for the SE attentional mechanism, RM-GoogLeNet for the CABM attentional mechanism, and RA-GoogLeNet for the ECA attentional mechanism). The confusion matrix (with predictions correct on the main diagonal and classifications incorrectly predicted on both the top and bottom of the main diagonal) is a table describing the realisation of the classification model on a test set of data with known true values, which can be related to the evaluation metrics, and allows for a more intuitive view of the recognition effects of the models incorporating the different attentional mechanisms, as indicated by the markings on the figure.
Figure 12

Confusion matrix for different attention mechanisms.

Figure 12

Confusion matrix for different attention mechanisms.

Close modal

These results suggest that the attention mechanism plays an important role in improving model performance. The experimental results of this study further validate the superiority of the ECA attention mechanism in the fluorescence spectral identification of pollution sources.

Comparison of the network before and after the improvement

In view of this study's proposal to add the two components of ECA attention mechanism and ResNet on top of GoogLeNet, it needs to be verified that including the addition of these two components, the model will improve the recognition accuracy of fluorescence spectra of pollution sources. In this study, the basic Inception module is replaced with the A-Inception module in the original GoogLeNet model, and the ECA attention mechanism and ResNet are added to the RA-GoogLeNet module. The comparison results of the three networks (GoogLeNet, A-GoogLeNet with the addition of an attention mechanism, and RA-GoogLeNet with the addition of a ResNet to the current one) with respect to accuracy, recall, and F1 score are shown in Table 3 (crab, cropland, fish, and shrimp). It can be concluded from the results that the accuracy of adding the ECA attention mechanism is 95.8%, which is 0.7% higher than the base model, and it is obvious that the addition of the attention mechanism strengthens the model's attention to some important features such as fluorescence intensity and position in the fluorescence spectrum. On this basis, the ResNet was added so that the fluorescence spectral image information loss was reduced. The accuracy of this network is 96.3% and the recall, F1 score of this model is above 95%.

Table 3

Comparison of results between improved networks

EvaluationCrabCroplandFishShrimpAccuracy (%)
GoogLeNet Precision 0.95 0.96 0.95 0.95 95.1 
Recall 0.96 0.96 0.95 0.95 
F1 0.96 0.96 0.95 0.96 
A-GoogLeNet Precision 0.96 0.96 0.96 0.95 95.8 
Recall 0.96 0.96 0.96 0.95 
F1 0.96 0.96 0.96 0.96 
RA-GoogLeNet Precision 0.97 0.97 0.96 0.96 96.3 
Recall 0.97 0.96 0.97 0.97 
F1 0.97 0.96 0.96 0.97 
EvaluationCrabCroplandFishShrimpAccuracy (%)
GoogLeNet Precision 0.95 0.96 0.95 0.95 95.1 
Recall 0.96 0.96 0.95 0.95 
F1 0.96 0.96 0.95 0.96 
A-GoogLeNet Precision 0.96 0.96 0.96 0.95 95.8 
Recall 0.96 0.96 0.96 0.95 
F1 0.96 0.96 0.96 0.96 
RA-GoogLeNet Precision 0.97 0.97 0.96 0.96 96.3 
Recall 0.97 0.96 0.97 0.97 
F1 0.97 0.96 0.96 0.97 

These results show that the model proposed in this study is able to better identify different categories of agricultural pollution sources with the addition of the ECA attention mechanism and ResNet. This further validates the importance of these two components in improving the performance of the model.

Comparison of different fluorescence spectral recognition models

Comparison of time complexity of different models

In this section, this study investigates the time complexity in image classification by using different network structures (as shown in Table 4). Different models were applied to an image dataset of size 224 × 224 for training and the corresponding times were recorded. The GPU used for local training is RTX3080Ti and the system platform used is cuda11/Python3.10.

Table 4

Comparison of the structure of the models

VGGAlexNetGoogLeNetRA-GoogLeNet
Conv1 Conv1 Conv1 Conv1_1 + Conv1_2 + Conv1_3 
Conv2 MaxPooling1 MaxPooling1  
MaxPooling1 Conv2 Conv2  
Conv3 MaxPooling2 MaxPooling2  
Conv4 Conv3 Inception(3a) A-Inception(3a) 
MaxPooling2 Conv4 Inception(3b) A-Inception(3b) 
Conv5 Conv5 MaxPooling3  
Conv6 MaxPooling3 Inception(4a) A-Inception(4a) 
Conv7 FC1 + FC2 + FC3 Inception(4b) A-Inception(4b) 
MaxPooling3  Inception(4c) A-Inception(4c) 
Conv8  Inception(4d) A-Inception(4d) 
Conv9  Inception(4e) A-Inception(4e) 
Conv10  MaxPooling4  
MaxPooling4  Inception(5a) A-Inception(5a) 
Conv11  Inception(5b) A-Inception(5b) 
Conv12  FC  
Conv13  Softmax  
MaxPooling5    
FC1 + FC2 + FC3    
VGGAlexNetGoogLeNetRA-GoogLeNet
Conv1 Conv1 Conv1 Conv1_1 + Conv1_2 + Conv1_3 
Conv2 MaxPooling1 MaxPooling1  
MaxPooling1 Conv2 Conv2  
Conv3 MaxPooling2 MaxPooling2  
Conv4 Conv3 Inception(3a) A-Inception(3a) 
MaxPooling2 Conv4 Inception(3b) A-Inception(3b) 
Conv5 Conv5 MaxPooling3  
Conv6 MaxPooling3 Inception(4a) A-Inception(4a) 
Conv7 FC1 + FC2 + FC3 Inception(4b) A-Inception(4b) 
MaxPooling3  Inception(4c) A-Inception(4c) 
Conv8  Inception(4d) A-Inception(4d) 
Conv9  Inception(4e) A-Inception(4e) 
Conv10  MaxPooling4  
MaxPooling4  Inception(5a) A-Inception(5a) 
Conv11  Inception(5b) A-Inception(5b) 
Conv12  FC  
Conv13  Softmax  
MaxPooling5    
FC1 + FC2 + FC3    

The results of the study show that the time required to train the model increases as the parameters increase. Based on the data in Table 5, it can be seen that the training time for the VGG model is about 3 h, for the AlexNet model it is 2.3 h, and for the GoogLeNet and RA-GoogLeNet models it is 1.2 and 1.5 h, respectively. Obviously, the model parameters in this study are less than the other models and training time has been reduced. However, in this experiment, increasing the number of layers and the number of parameters in the model did not always improve the model performance. In fact, when the number of layers and parameters exceeds a certain threshold, the accuracy decreases instead. This problem is evident in the current experiment, where the first model in the table has an accuracy of 94.6%, while the second and third models have an accuracy of 95.4 and 95.1%, respectively. This validates the statement above that increasing the number of parameters does not necessarily improve the accuracy of the model. Although the RA-GoogLeNet proposed in this study introduces two modules, the attention mechanism and the ResNet, on top of GoogLeNet, which increases the corresponding parameters and training time, the experimental results show that this increase in parameters and training time brings a proportional enhancement that is acceptable in the aggregate. However, the other models have no significant advantage in terms of the number of parameters or accuracy.

Table 5

Effect of number of parameters on training time and accuracy for different models

ModelParameters (M)Image sizeTraining time (h)Accuracy (%)
VGG 138.3 224 × 224 94.6 
AlexNet 62.3 224 × 224 2.3 95.4 
GoogLeNet 6.9 224 × 224 1.2 95.1 
RA-GoogLeNet 9.2 224 × 224 1.5 96.3 
ModelParameters (M)Image sizeTraining time (h)Accuracy (%)
VGG 138.3 224 × 224 94.6 
AlexNet 62.3 224 × 224 2.3 95.4 
GoogLeNet 6.9 224 × 224 1.2 95.1 
RA-GoogLeNet 9.2 224 × 224 1.5 96.3 

Therefore, a good balance between accuracy and time complexity is needed when selecting models for image classification tasks. The results of this study show that RA-GoogLeNet strikes a good balance in this regard with high accuracy and moderate time complexity.

Performance comparison of different models

To verify the superiority of the performance of the classification models proposed in this study, four models, AlexNet, VGG, GoogLeNet, and RA-GoogLeNet, proposed in this study were selected for the comparison experiments.

AlexNet (Luo et al. 2022) consists of five convolutional layers, three fully connected layers, and a Softmax output layer, using the ReLU nonlinear activation function, which is a structure that allows for fast training of the neural network, and prevents the loss of gradient while increasing the nonlinearity of the network. VGG (Que et al. 2023) consists of 13 convolutional layers and 3 connectivity layers, and is characterised by stacking the network through the extensive use of 3 × 3 convolutional kernels. The main feature of GoogLeNet is the introduction of Inception modules, where the network is a stack of multiple Inception modules.

The experimental results are shown in Figure 13; in terms of model recognition accuracy, VGG has the lowest recognition accuracy of 94.6%. The RA-GoogLeNet network proposed in this study has the best recognition result of 97.3%. This is a 1.2% improvement over the base network GoogLeNet, which improves the recognition accuracy of the network without significantly increasing the training cost. From the learning curves of the models, the fastest convergence was observed in the RA-GoogLeNet model and several other models converged at about the same rate when the models were trained on the agricultural pollution fluorescence spectral image dataset. From an epoch of 30 to the end, RA-GoogLeNet consistently had the highest accuracy, with the other models having relatively low accuracy. In addition, the accuracy of VGG, GoogLeNet, and AlexNet has been fluctuating and not small in magnitude, indicating the poor applicability of the model to the present fluorescence spectral dataset. In conclusion RA-GoogLeNet's method is the most effective and more applicable.
Figure 13

Accuracy and loss rate of different fluorescence spectral recognition models.

Figure 13

Accuracy and loss rate of different fluorescence spectral recognition models.

Close modal

In this study, combining the fluorescence spectra of agricultural organic pollutants around Changdang Lake, we proposed a method (RA-GoogLeNet) that can rapidly identify agricultural organic pollutants, which can effectively solve the problem of organic pollutants from agricultural surface sources in the Changdang Lake basin. The method introduces the ECA attention mechanism and ResNet on top of GoogLeNet, which not only improves the performance of the model but also reduces the computational cost. The experimental results show that:

  • (1)

    The RA-GoogLeNet model has a classification accuracy of 96.3%, which is better than other classification models.

  • (2)

    Incorporating the ECA attention mechanism in the Inception module can improve the model's recognition of fluorescence spectra from agricultural surface sources.

  • (3)

    The ResNet connected between each A-Inception module prevents the problem of gradient vanishing and information loss due to the increase in the number of network layers.

  • (4)

    The Leaky ReLU activation function used prevents neuronal necrosis, which allows more image features to be extracted.

  • (5)

    RA-GoogLeNet has fewer parameters and a smaller network computational size than conventional models, resulting in reduced training time.

The model performs well when this dataset is classified, but the number of categories is currently low, including only the farming and agriculture categories, and the amount of data is not sufficient to cover other sources of pollution, such as printing and dyeing and domestic wastewater. In addition, the categorisation results of this study can only be output in a single way and cannot identify mixed pollution from multiple categories. Future work will include collecting more data to validate and optimise the model and improve the efficiency of the network parameters to obtain better and finer classification results.

This study was funded by the Chinese National Natural Science Foundation (61803050, 52070023), City-based Tracking Study on Ecological Environment Protection and Restoration of the Yangtze River (2022-LHYJ-02-0502-02-11), Changzhou the Key Research and Development Plan (Science and Technology Support for Social Development) project (CE20225061, CE20235071), the Postgraduate Research & Practice Innovation Program of Jiangsu Province (KYCX22_3079), and the Innovation and Entrepreneurship Training Program of Jiangsu College Student (202310292293B, 202310292078Y).

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.

Alex
K.
,
Ilya
S.
&
Geoffrey
E.
2012
ImageNet classification with deep convolutional neural networks
.
Commun. ACM
60
,
84
90
.
https://doi.org/10.1145/3065386
.
Cao
Y.
,
Zhou
W.
,
Zang
M.
,
An
D.
,
Feng
Y.
&
Yu
B.
2023
MBANet: A 3D convolutional neural network with multi-branch attention for brain tumor segmentation from MRI images
.
Biomed. Signal Process. Control.
80
.
https://doi.org/10.1016/j.bspc.2022.104296.
Chen
W.
,
Westerhoff
P.
,
Leenheer
J. A.
&
Booksh
K.
2003
Fluorescence excitation-emission matrix regional integration to quantify spectra for dissolved organic matter
.
Environ. Sci. Technol.
37
,
5701
5710
.
https://doi.org/10.1021/es034354c
.
Cui
J. Y.
,
Cui
J.
,
Peng
Y.
,
Yao
D. R.
,
Chan
A.
,
Chen
Z. Y.
&
Chen
Y. M.
2020
Sources and trends of oxidized and reduced nitrogen wet deposition in a typical medium-sized city of eastern China during 2010–2016
.
Sci. Total Environ.
744
,
140558
.
https://doi.org/10.1016/j.scitotenv.2020.140558
.
Cuss
C. W.
,
McConnell
S. M.
&
Guéguen
C.
2016
Combining parallel factor analysis and machine learning for the classification of dissolved organic matter according to source using fluorescence signatures
.
Chemosphere
155
,
283
291
.
https://doi.org/10.1016/j.chemosphere.2016.04.061
.
Du
Y.
,
Hu
L.
,
Wu
G.
,
Tang
Y.
,
Cai
X.
&
Yin
L.
2023
Diagnoses in multiple types of cancer based on serum Raman spectroscopy combined with a convolutional neural network: Gastric cancer, colon cancer, rectal cancer, lung cancer
.
Spectrochim. Acta A Mol. Biomol. Spectrosc.
298
,
122743
.
https://doi.org/10.1016/j.saa.2023.122743
.
Erguzel
T. T.
,
Ozekes
S.
,
Sayar
G. H.
,
Tan
O.
&
Tarhan
N.
2015
A hybrid artificial intelligence method to classify trichotillomania and obsessive compulsive disorder
.
Neurocomputing
161
,
220
228
.
https://doi.org/10.1016/j.neucom.2015.02.039
.
Gu
X. Q.
,
Shen
Z. X.
,
Xue
J.
,
Fan
Y. Q.
&
Ni
T.
2021
Brain tumor MR image classification using convolutional dictionary learning with local constraint
.
Front. Neurosci.
15
,
679847
.
https://doi.org/10.3389/fnins.2021.679847
.
Guo
W.
,
Fu
Y.
,
Ruan
B.
,
Ge
H.
&
Zhao
N.
2014
Agricultural non-point source pollution in the Yongding River Basin
.
Ecol. Indic.
36
,
254
261
.
https://doi.org/10.1016/j.ecolind.2013.07.012
.
He
Y. M.
,
Xiang
H.
,
Zhou
H.
&
Chen
J. H.
2023
In-situ fault detection for the spindle motor of CNC machines via multi-stage residual fusion convolution neural networks
.
Comput. Ind.
145
,
103810
.
https://doi.org/10.1016/j.compind.2022.103810
.
Hu
J.
,
Shen
L.
,
Albanie
S.
,
Sun
G.
&
Wu
E.
2020
Squeeze-and-excitation networks
.
IEEE Trans. Pattern Anal. Mach. Intell.
2020
(
42
),
2011
2023
.
https://doi.org/10.1109/TPAMI.2019.2913372
.
Huang
L.
,
Zhao
Y. G.
&
Yang
T. J.
2024
Skin lesion image segmentation by using backchannel filling CNN and level sets
.
Biomed. Signal Process. Control.
87
,
105417
.
https://doi.org/10.1016/j.bspc.2023.105417
.
Li
Y. X.
,
Tan
C. L.
,
Ding
X. Q.
&
Liu
C. S.
2004
Contextual post-processing based on the confusion matrix in offline handwritten Chinese script recognition
.
Pattern Recognit.
37
,
1901
1912
.
https://doi.org/10.1016/j.patcog.2004.03.002
.
Lim
W. X.
,
Chen
Z. Y.
&
Ahmed
A.
2022
The adoption of deep learning interpretability techniques on diabetic retinopathy analysis: A review
.
Med. Biol. Eng. Comput.
60
(
3
),
633
642
.
https://doi.org/10.1007/s11517-021-02487-8
.
Liu
Y. Z.
&
Chen
Z. Y.
2023
Prediction of biochemical oxygen demand with genetic algorithm-based support vector regression
.
Water Qual. Res. J. Can.
58
(
2
),
87
98
.
https://doi.org/10.2166/wqrj.2023.015
.
Liu
Y. S.
,
Wang
X. J.
,
Wang
L.
&
Liu
D. L.
2019
A modified leaky ReLU scheme (MLRS) for topology optimization with multiple materials
.
Appl. Math. Comput.
352
,
188
204
.
https://doi.org/10.1016/j.amc.2019.01.038
.
Luo
X. X.
,
Wen
W. H.
,
Wang
J. R.
,
Xu
S. Y.
,
Gao
Y. Y.
&
Huang
J. L.
2022
Health classification of Meibomian gland images using keratography 5M based on AlexNet model
.
Comput. Methods Programs Biomed.
219
,
106742
.
https://doi.org/10.1016/j.cmpb.2022.106742
.
Minu
M. S.
&
Canessane
R. A.
2022
Deep learning-based aerial image classification model using inception with residual network and multilayer perceptron
.
Microprocess. Microsyst.
95
,
104652
.
https://doi.org/10.1016/j.micpro.2022.104652
.
Nijaguna
G. S.
,
Babu
J. A.
,
Parameshachari
B. D.
,
De Prado
R. P.
&
Frnda
J.
2023
Quantum Fruit Fly algorithm and ResNet50-VGG16 for medical diagnosis
.
Appl. Soft Comput.
136
,
110055
.
https://doi.org/10.1016/j.asoc.2023.110055
.
Niu
Z. Y.
,
Zhong
G. Q.
&
Yu
H.
2021
A review on the attention mechanism of deep learning
.
Neurocomputing
452
,
48
62
.
https://doi.org/10.1016/j.neucom.2021.03.091
.
Ooi
K. S.
,
Chen
Z. Y.
,
Poh
P. E.
&
Cui
J.
2022
BOD5 prediction using machine learning methods
.
Water Supply
22
(
1
),
1168
1183
.
https://doi.org/10.2166/ws.2021.202
.
Qu
J.
,
Zhang
Y. H.
&
Zhang
Z. G.
2023
PMA-Net: A parallelly mixed attention network for person re-identification
.
Displays
78
,
102437
.
https://doi.org/10.1016/j.displa.2023.102437
.
Que
Y.
,
Dai
Y.
,
Ji
X.
,
Kwan
L. A.
,
Chen
Z.
,
Jiang
Z. L.
&
Tang
Y. C.
2023
Automatic classification of asphalt pavement cracks using a novel integrated generative adversarial networks and improved VGG model
.
Eng. Struct.
277
,
115406
.
https://doi.org/10.1016/j.engstruct.2022.115406
.
Rahman
C. R.
,
Arko
R. S.
,
Ali
M. E.
,
Lqbal Khan
M. A.
,
Apon
S. H.
,
Nowrin
F.
&
Wasif
A.
2020
Identification and recognition of rice diseases and pests using convolutional neural networks
.
Biosyst. Eng.
194
,
112
120
.
https://doi.org/10.1016/j.biosystemseng.2020.03.020
.
Ruan
K.
2022
3D Fluorescence Component Prediction and Recognition Based on Convolutional Neural Network VGG
.
MS thesis
,
Sichuan Agricultural University, Agricultural Information Engineering
,
Sichuan, China
.
http://doi.org/10.27345/d.cnki.gsnyu.2022.000732
.
Shi
Y. K.
,
Wang
Z. W.
,
Du
X. J.
,
Ling
G. B.
,
Jia
W. C.
&
Lu
Y. R.
2022
Research on the membrane fouling diagnosis of MBR membrane module based on ECA-CNN
.
J. Environ. Chem. Eng.
10
,
107649
.
https://doi.org/10.1016/j.jece.2022.107649
.
Su
T. F.
&
Zhang
S. W.
2023
Object-based crop classification in Hetao irrigation zone by using deep learning and region merging optimization
.
Comput. Electron. Agric.
214
,
108284
.
https://doi.org/10.1016/j.compag.2023.108284
.
Tolkach
Y.
,
Dohmgorgen
T.
,
Toma
M.
&
Kristiansen
G.
2020
High-accuracy prostate cancer pathology using deep learning
.
Nat. Mach. Intell.
2
,
411
.
https://doi.org/10.1038/s42256-020-0200-7
.
Wang
H. L.
,
He
P.
,
Shen
C. Y.
&
Wu
Z. N.
2019
Effect of irrigation amount and fertilization on agriculture non-point source pollution in the paddy field
.
Environ. Sci. Pollut. Res. Int.
26
,
10363
10373
.
https://doi.org/10.1007/s11356-019-04375-z
.
Wang
F.
,
Diao
X. M.
,
Chang
S.
&
Xu
L.
2021
Recent progress of deep learning in drug discovery
.
Curr. Pharm. Des.
27
(
17
),
2088
2096
.
https://doi.org/10.2174/1381612827666210129123231
.
Wang
X.
,
Zhang
M.
,
Liu
L.
,
Wang
Z.
&
Lin
K.
2022
Using EEM-PARAFAC to identify and trace the pollution sources of surface water with receptor models in Taihu Lake Basin, China
.
J. Environ. Manage.
321
,
115925
.
https://doi.org/10.1016/j.jenvman.2022.115925
.
Wang
Z. W.
,
Zeng
J. X.
,
Shi
Y. K.
&
Ling
G. B.
2023
MBR membrane fouling diagnosis based on improved residual neural network
.
J. Environ. Chem. Eng.
11
,
109742
.
https://doi.org/10.1016/j.jece.2023.109742
.
Wei
S. Z.
,
Chen
Z. Y.
,
Arumugasamy
S. K.
&
Chew
I. M. L.
2022
Data augmentation and machine learning techniques for control strategy development in bio-polymerization process
.
Environ. Sci. Ecotechnol.
11
,
100172
.
https://doi.org/10.1016/j.ese.2022.100172
.
Wünsch
U. J.
,
Murphy
K. R.
&
Stedmon
C. A.
2017
The one-sample PARAFAC approach reveals molecular size distributions of fluorescent components in dissolved organic matter
.
Environ. Sci. Technol.
51
,
11900
11908
.
https://doi.org/10.1021/acs.est.7b03260
.
Xiao
Y.
,
Han
D. M.
,
Currell
M.
,
Song
X. F.
&
Zhang
Y. H.
2023
Review of endocrine disrupting compounds (EDCs) in China's water environments: Implications for environmental fate, transport and health risks
.
Water Res.
245
,
120645
.
https://doi.org/10.1016/j.watres.2023.120645
.
Xu
R. Z.
,
Cao
J. S.
,
Feng
G. Y.
,
Luo
J. Y.
,
Feng
Q.
,
Ni
B. J.
&
Fang
F.
2022
Fast identification of fluorescent components in three-dimensional excitation-emission matrix fluorescence spectra via deep learning
.
Chem. Eng. J.
430
,
132893
.
https://doi.org/10.1016/j.cej.2021.132893
.
Yadav
M. K.
,
Aryal
R.
,
Short
N. D.
&
Saint
C. P.
2019
Fluorescence excitation-emission spectroscopy: An analytical technique to monitor drugs of addiction in wastewater
.
Water
11
,
377
.
https://doi.org/10.3390/w11020377
.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY 4.0), which permits copying, adaptation and redistribution, provided the original work is properly cited (http://creativecommons.org/licenses/by/4.0/).