3.1 Mid-infrared spectroscopy

Spectroscopy is a technique that investigates the interaction between matter and electromagnetic radiation (e.g., light), focusing on the relationship between the radiation’s wavelength or frequency and its interaction with the material. The method involves measuring and analyzing the spectrum of absorbed radiation by a substance under analysis.

Infrared spectroscopy is used in entomological surveillance to identify mosquito species by quantifying how their cuticles absorb infrared light. Most works generally consider near-infrared spectroscopy (NIRS) for examining insects, in which the spectrum is restricted from 10,000 to 4,000 cm^–1 (Barbosa et al. 2018; Johnson 2020; Mayagaya et al. 2009). Although efficient in identifying insect species, age prediction using NIRS is still challenging (Siria et al. 2022). In order to obtain better age predictions, researchers have investigated the use of mid-infrared spectroscopy (MIRS) (González Jiménez et al. 2019). MIRS considers the absorption spectrum in the mid-infrared region from 400 to 4,000 cm^–1. In Figure 1, we illustrate the mid-infrared absorption spectra obtained by Anopheles gambiae and Anopheles arabienses mosquitoes, studied in this work.

As we note in Figure 1, the data generated by MIRS are time series, that is, real-valued continuous data with a length of 3,600 observations. These series are used as input for machine learning methods to predict species and age of insects.

3.2 Data and preprocessing

Our study considers an MIRS dataset collected by Siria et al. (2022) and publicly available in the Enlighten database.² This dataset contains spectra of 41,151 female mosquitoes aged 1 to 17 days from the following malaria vector species: Anopheles gambiae (AG), Anopheles arabiensis (AA), and Anopheles coluzzii (AC). In order to increase variability, the data was collected from mosquitoes with different physiological states and the spectroscopy was measured in laboratories in the UK, Tanzania, and Burkina Faso. From this data, we perform a set of preprocessing procedures to train supervised algorithms from different approaches to predict the species and age of malarial mosquitoes, as discussed in the following.

Initially, we discarded the examples from the Anopheles coluzzii species to reduce the class imbalance since this species represents only 3% of the examples, making the steps of training and testing classifiers more difficult. Thus, our experiments consider 39,716 examples from two species, which we split into 70% for training and 30% for testing. Since we have two tasks (i.e., species and age prediction), we split the data into stratified training and test sets according to the labels of each task. For age prediction, we grouped the examples spanning 17 days into three age classes: 1–4, 5–10, and 11–17 days. Figure 2 illustrates the class distribution for both tasks considered in our holdout evaluation.

We have noted that the original data have varying sampling rates, with examples from 1,701 to 14,932 observations due to the different equipment used in the analysis. However, we need a fixed length to use the data as input to train the machine learning models. In this direction, we downsampled all examples to have 1,000 observations after applying an anti-aliasing filter (Broersen & de Waele 2000). Such a dimensionality reduction also contributes to the computational cost spent on model training or feature extraction.

Finally, the last preprocessing step was the normalization to guarantee that all examples are in the same range of values. Without normalization, observations with larger magnitudes or variances may dominate the learning process, leading to biased model training and inaccurate classification results. Normalization also reduces the impact of outliers.

In this phase, we employed z-normalization (Lima & Souza 2023). This method rescales the time-series values to have zero mean and standard deviation close to one. The normalized values X’ of a time series X are obtained according to Equation 1, in which μ and σ are, respectively, the mean and standard deviation of a series with n observations.

In Figure 3, we illustrate the same time series previously shown in Figure 1 after resampling to 1,000 observations and z-normalization, responsible for changing both horizontal and vertical original axes.

3.3 Time-series classification

A time series X = (x₁, x₂,…,x_n), such as the MIRS data previously illustrated in Figure 3, is an ordered sequence of n real-values x_i measured at equal time intervals, in which x_i represents a value observed at time i. In the classification task, we aim to assign a class label ŷ (e.g., AG or AA) to an unknown query time series using a supervised model built with a labeled training set $\mathcal{D}\text{=}\{(X_1,\hspace{0.17em}{Y}_1),\hspace{0.17em}(X_2,\hspace{0.17em}{Y}_2),\hspace{0.17em}\dots ,\hspace{0.17em}(X_N,\hspace{0.17em}{Y}_N)\}$M2 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\mathcal D}{ = }\{ ({X_1},\;{Y_1}),\;({X_2},\;{Y_2}),\; \ldots,\;({X_N},\;{Y_N})\}\] \end{document} of N time series where Y_i denotes the target variable for each X_i.

Since the features of time-series examples are ordered and correlated, classification models for this data differ from traditional models in which the features are independent. In the following, we introduce the main approaches for time-series classification based on the categories discussed in Bagnall et al. (2017) and other recent proposals (Dempster, Schmidt & Webb 2023; Middlehurst et al. 2021).

3.3.1 Feature-based

Feature-based classifiers consider a preliminary step of extracting descriptive statistics from the time series to use them as features to train conventional machine learning algorithms, such as Random Forests, Support Vector Machines (SVM), and XGBoost. The process is illustrated in Figure 4 given an input time series. The most straightforward approach is to use the raw data comprising the 1,000 observations of the time series as descriptive features for each example.

Figure 4 

General process of the feature-based approach.

Another strategy is to consider a subset of values as features instead of all observations. For example, González Jiménez et al. (2019) consider the values observed at 17 specific positions of the MIRS time series that represent the absorption of pre-selected wavenumbers. These wavenumbers contain well-defined intense peaks, which are easily identifiable as coming from the chemical components of the cuticle, being discriminant to identify the species. Figure 5 illustrates the positions of such wavenumbers in an MIRS time series.

Figure 5 

Wavenumbers selected as features to train supervised machine learning algorithms.

For more general time-series classification problems, we can consider statistical features in the time domain such as mean, maximum, minimum, standard deviation, mean absolute deviation, skewness, and kurtosis or structural features such as trend, seasonality, periodicity, and self-similarity. Besides, we can transform the series from time to frequency domain by Discrete Fourier transform (DFT) to extract features such as spectral irregularity, flux, roll-off, and energy (Silva et al. 2015). We can also transform the time series to a two-dimensional representation, such as a Recurrence Plot, then extract statistical descriptive features related to texture (Souza, Silva & Batista 2014).

Highly Comparative Time-Series Analysis (HCTSA) is a library that computes over 7,700 features from time series (Fulcher & Jones 2017). Recently, Lubba et al. (2019) introduced the 22 CAnonical Time-series CHaracteristics (Catch-22), which identified a reduced subset with the 22 most discriminatory features from the HCTSA that provides competitive results for different time-series problems. Among the selected features, we can highlight: longest period of consecutive values above the mean, time intervals between successive extreme events above and below the mean, first minimum of autocorrelation function, centroid of the Fourier power spectrum, change in correlation length after iterative differencing, periodicity, among others. It is noteworthy to mention that such features are suitable for the particularities of time series and implicitly consider the temporal dependence of the observations.

In this work, we also combine the 22 features computed by Catch-22 with the 17 values observed at specific wavenumbers, composing a new feature set with 39 descriptive features to train the classifiers.

3.3.2 Interval-based

While feature-based methods extract features from the whole time series, interval-based methods select one or more intervals (i.e., contiguous subsequences) of the series to derive features. An advantage of this approach is to discard regions of the series with noise that could confound the classifier through irrelevant features. For example, in Figure 6, we show five examples from the same class (AG) and highlight two different regions. The region on the left represents a promising candidate interval to extract discriminatory features. On the other hand, the region on the right side contains a significant variation between the examples of the same class, and the extraction of features in this interval can confound the classifier.

Figure 6 

Example of promising and unpromising intervals on MIRS data of Anopheles gambiae.

For a given time series with n observations, there are $\frac{n(n–1)}{2}$M3 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\textstyle{{n(n-1)} \over 2}}\] \end{document} possible intervals, and a challenge is finding the best interval. In general, these algorithms generate different random intervals and classifiers on each one, ensembling the resulting predictions.

The most representative interval-based algorithms are Time Series Forest (TSF) (Deng et al. 2013), Canonical Interval Forest (CIF) (Middlehurst, Large & Bagnall 2020), and interval-based Diverse Representation Canonical Interval Forest (DrCIF) (Middlehurst et al. 2021).

TSF is an ensemble of tree classifiers based on Random Forest (Breiman 2001), built on the summary statistics of randomly selected intervals from the time series. For each tree, k intervals with a random start position and length are randomly selected. The mean, standard deviation, and slope are extracted from each interval and concatenated into a feature vector with 3k length. These features are then used to build a tree, which is added to the ensemble. A TSF predicts a testing instance as the majority class according to the votes from all time-series trees in the ensemble. TSF considers 500 trees and $\sqrt{n}$M4 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\sqrt n\] \end{document} random intervals by default.

CIF classifier is an adaptation of TSF that embeds the Catch-22 features. The classifier considers the three TSF features along with the 22 features from Catch-22. In order to add diversity to the ensemble, eight of the 25 features are randomly chosen for each tree.

DrCIF is an extension of CIF that extracts features from multiple intervals taken from the original series, the first-order difference series, and the periodograms of the whole series. Seven basic summary statistics are extracted from the interval of any of the three representations: mean, standard deviation, slope, median, interquartile range, minimum, and maximum. DrCIF adds the Catch-22 features to form a set of 29 features. As performed by CIF, eight of the 29 features are randomly selected for each tree.

3.3.3 Convolution-based

The convolution-based methods are computationally efficient classifiers that use many random convolutional kernels to extract features from the time series and use them as input for linear classifiers such as logistic regression. For time series, a kernel is a vector of weights (e.g., [–1, 0, 1]) convolved with a time series through a sliding dot product operation to produce a feature map. Such convolution kernels at different and random lengths and weights can capture complex patterns and shapes to discriminate the series. The main methods of this category are Rocket (Dempster, Petitjean & Webb 2020) and MiniRocket (Dempster, Schmidt & Webb 2021).

Instead of learning a convolutional kernel as performed by convolutional neural networks, Rocket (RandOm Convolutional KErnel Transform) generates more than 10,000 random convolutional kernels that capture relevant features from time series when combined. Each one-dimensional kernel with random weights, length, bias, dilatation, and padding slides through a time series, performing the dot product and producing a transformed time series named feature map. Rocket computes two aggregate features from this transformed time series by a pooling operation: the maximum value (i.e., global max pooling) and the proportion of positive values (PPV). Figure 7 illustrates the whole feature extraction process by convolution considering a single kernel. Since Rocket produces two features per kernel (MAX and PPV) and there are 10,000 random kernels, the algorithm produces 20,000 features per time series. These features are then used to train a linear classifier, such as ridge regression or logistic regression using stochastic gradient descent.

Figure 7 

Process of kernel convolution for feature extraction in which two features (MAX and PPV) are extracted from the transformed time series (or feature map). Rocket performs such a process for 10,000 random kernels generating 20,000 features for training a linear classifier.

A limitation of Rocket is the large number of kernels required to make it accurate. Besides, Rocket is non-deterministic due to the use of random kernels. In this direction, MiniRocket (MINImally RandOm Convolutional KErnel Transform) follows the main steps of Rocket, reducing the computational cost and removing the randomness by using a small and fixed set of kernels. MiniRocket uses kernels of length 9, with weights restricted to two values, and computes only PPV from the transformed time series. Thus, the algorithm uses 10,000 features to train a linear classifier.

3.3.4 Deep learning-based

Conventional machine learning algorithms such as SVM or Random Forests learn a predictive model from handcrafted features extracted from data, as discussed in Section 3.3.1, requiring domain experts to design adequate extractors according to the problem. On the other hand, deep learning algorithms consider an end-to-end approach in which the entire predictive task is addressed as a single and integrated system without the need for explicit feature engineering. These algorithms receive raw data as input and generate the answer as output after transforming the data into multiple levels of representations that contain relevant and discriminative features (LeCun, Bengio & Hinton 2015). In general terms, a deep neural network is a composition of multiple hierarchical parametric functions (in practice, the layers of a network) where each layer is a different representation of input data (Sarker 2021).

While deep neural networks, commonly referred to as deep learning, have a well-established track record in computer vision applications like face recognition (Lawrence et al. 1997) and object detection (Zhao et al. 2019), their application to one-dimensional data as time-series classification problems has only gained prominence in recent times, as demonstrated by the competitive results discussed in Ismail Fawaz et al. (2019) in a comprehensive comparison considering time-series benchmark datasets from different domains.

Currently, InceptionTime Ismail Fawaz et al. (2020) is the state-of-the-art neural network for time-series classification. InceptionTime is an ensemble of five deep learning models, each one created by cascading multiple Inception modules. The Inception Network is inspired by the Inception-v4 architecture Szegedy et al. (2017) and contains two residual blocks in which each residual block’s input is transferred via a shortcut linear connection to be added to the next block’s input. Following these residual blocks, a Global Average Pooling (GAP) layer averages the outputted multivariate time series by the previous layers. The network contains a softmax activation function for label prediction in the last layer.

Previously to the proposal of InceptionTime, Ismail Fawaz et al. (2020) showed that the Residual Network (ResNet) is a strong deep learning baseline for time series, followed by the Fully Convolutional Network (FCN). In this work, we also evaluate the performance of the Time Convolutional Neural Network (Time-CNN) (Zhao et al. 2017).

FCN is a type of neural network architecture designed for semantic segmentation tasks (e.g., image segmentation), where the goal is to label the regions of an image based on its semantic category. The use of FCNs for time-series classification was first proposed by Wang, Yan, and Oates (2017). The key characteristic of FCNs is that they are composed entirely of convolutional layers without fully connected layers at the end, different from a typical Convolutional Neural Network (CNN). In the context of time series, a convolution consists of sliding one-dimensional filters over the series, enabling the extraction of non-linear discriminant features. Precisely, the FCN architecture proposed by Wang, Yan, and Oates consists of three convolutional blocks with the filter sizes {128, 256, 128}, where each block contains three operations: a convolution followed by batch normalization, and the result is then fed into a Rectified Linear Unit (ReLU) activation function. The convolution operation is fulfilled by three one-dimensional kernels with the sizes {8, 5, 3} without striding. The output features of the third convolutional block are fed into a Global Average Pooling (GAP) layer. Finally, the prediction is produced by a softmax function.

ResNet increases the depth of conventional neural networks by introducing residual shortcut connections between consecutive convolutional layers (Wang, Yan & Oates 2017). Figure 8 illustrates the ResNet architecture evaluated in this work (Ismail Fawaz et al. 2019). The network contains 11 layers, of which the first nine are convolutional layers, followed by a GAP layer that averages the time series across the time dimension. The network has three residual blocks followed by a GAP layer and a final softmax classifier with the number of neurons equals the number of classes. Each residual block comprises three convolutions whose output is added to the residual block’s input and then fed to the next layer. The number of filters for all convolutions is 64, with the ReLU activation function. In each residual block, the filter’s length is set to 8, 5, and 3 for the first, second, and third convolution.

Figure 8 

Residual Network (ResNet) architecture for time-series classification (Lima & Souza 2023).

A standard CNN architecture consists of two main components: i) a feature extraction module composed of convolution layers followed by pooling operations to reduce the dimension of feature maps, and ii) a fully connected network that receives the previously extracted features as input. In the context of time-series classification, the Time-CNN architecture comprises two consecutive convolutional layers with 6 and 12 filters, respectively, followed by a local average pooling operation of length 3. In this network, the convolutional layers use the sigmoid activation function. Time-CNN diverges from the typical CNN in two aspects: firstly, it employs the mean squared error (MSE) as the loss function instead of the traditional categorical cross-entropy, and secondly, the network incorporates a local average pooling operation instead of local max pooling.

4.1 Settings

We performed the tasks of species and age prediction of malarial mosquitoes using the mid-infrared spectroscopy data provided by Siria et al. (2022). We split the dataset into training and testing sets following the 70/30 ratio. For both tasks, we use classification accuracy to measure the performance of our models. This measure represents the ratio of correct predictions to the total number of examples in the testing set.

Our evaluation considers 14 machine learning algorithms from four time-series learning approaches: i) feature-based, ii) interval-based, iii) convolution-based, and iv) deep learning-based. In Table 4.1, we show the set of algorithms from each approach and their parameter values. For feature-based methods, we found the values by a grid search procedure performed in the training data (Syarif, Prugel-Bennett & Wills 2016). For the remaining algorithms, we consider default values.

Since algorithms such as TSF, CIF, DrCif, Rocket, and MiniRocket are unsuitable for non-time-series data, we categorize these algorithms as non-traditional along the paper. On the other hand, algorithms such as KNN, LR, SVM, RF, and XGB used in the feature-based approach are categorized as traditional since they can be employed on conventional and time-series data.

For approaches such as feature-based, we trained five algorithms considering the following feature sets: i) raw data, in which we consider the 1,000 observations of a time series after the preprocessing phase as features, ii) wavenumbers, in which we extract the time-series values at 17 representative positions, iii) Catch-22, in which we extracted 22 predictive features and iv) Catch-22 + wavenumbers, in which we concatenate the 22 features extracted from Catch-22 and 17 wavenumbers, composing a feature vector with 39 features. The second option for feature sets (wavenumbers) is the same as considered by the related works previously discussed. The other feature sets are our attempts to improve the results from the literature.

Given the 14 supervised learning algorithms and feature sets, we have evaluated 29 different settings. Figure 9 shows an overview of our experimental setting considering varying approaches, feature sets, and supervised learning algorithms.

4.2 Species prediction

The species prediction task evaluated in this work aims to distinguish two species from the Anopheles genus. The Anopheles gambiae and Anopheles arabiensis are morphologically indistinguishable in the adult stage (Zianni et al. 2013), and both are the most broadly distributed and efficient vectors of malaria (Coetzee, Craig & Le Sueur 2000).

We begin our analysis by showing the results of the feature-based algorithms. In Figure 10, we show the accuracies obtained by the five algorithms evaluated considering four different feature sets. For all machine learning algorithms, we note that the best feature set for this task is composed of the 1,000 observations of the time series (i.e., raw data), outperforming the wavenumbers as proposed by González Jiménez et al. (2019) in their evaluation with similar MIRS data. Using raw data, Logistic Regression (LR) and SVM are the most accurate classifiers, with 93% and 92% of correct predictions, respectively. Besides the performance, it is essential to note an advantage of this feature set: the lack of an additional step for feature extraction as required by other methods. On the other hand, the use of features provided by Catch-22 obtained the worst results, varying from 58% (LR) to 72% (RF) of accuracy.

The feature-based is a well-known approach that allows the use of traditional machine learning algorithms. However, interval and convolution-based classifiers are well-suited approaches for time-series data that still need to be evaluated in the mosquito prediction, as performed in this work. In Figure 11, we show the results obtained by both approaches. For the interval-based methods, CIF achieved the best result with 90% of accuracy, being surpassed by the best results obtained by feature-based algorithms, while the convolution-based method Rocket achieved a competitive result of 93%.

Finally, the results of the fourth approach are shown in Figure 12. InceptionTime was the most accurate with 97%, followed by ResNet with 96%. Such algorithms presented the best results among all approaches. On the other hand, Time-CNN is outperformed by simple algorithms such as Rocket and Logistic Regression, requiring more time and costly hardware (i.e., GPU) for model training.

Table 1 shows a ranking with the 15 best results for better comparison between the approaches, in which we note deep learning and convolutional based methods at the top.

In Figure 13, we show the error distribution between the classes AA (Anopheles arabiensis) and AG (Anopheles gambiae) carried out by the best algorithm from each approach. The normalized confusion matrix of InceptionTime illustrated in Figure 13d) demonstrates the superior results of this algorithm with predictions close to perfect for both species.

4.3. Age prediction

The ability of female mosquitoes to transmit diseases such as malaria or yellow fever is age dependent. Due to the incubation period of the parasites and pathogens that mosquitoes transmit, only older mosquitoes are potential vectors of diseases (Dowell, Noutcha & Michel 2011). Thus, predicting the age of malarial mosquitoes holds significant importance in the ongoing efforts to control mosquitoes. Besides, younger mosquitoes are often more susceptible to insecticides, making them prime targets for control interventions. For age prediction, we train our models to distinguish the mosquitos’ age using mid-infrared spectroscopy data in three ranges of values: 1–4, 5–10, and 11–17 days.

In Figure 14, we show the results obtained by feature-based algorithms. As well as for species prediction, using raw data as a feature set provided the best results. However, the most accurate classifier for species prediction (Logistic Regression) did not prove adequate for age prediction. Specifically, the XGBoost algorithm achieved the best result with 75% accuracy, followed by Random Forest (72%), while Logistic Regression shows results close to 50%. Comparatively, we note that age prediction is a more challenging task than species identification.

Figure 15 illustrates the results of interval and convolution-based methods. For age prediction, we noted that the interval-based method CIF slightly outperformed the results of rivals with 76% accuracy.

As well as for species prediction, the deep learning algorithm InceptionTime achieved the best result among the approaches, achieving 83% accuracy. Figure 16 shows a performance comparison among the deep learning methods. Interestingly, although InceptionTime achieved the best result, the other deep learning algorithms showed results inferior to interval and convolution-based methods.

In the ranking shown in Table 2 with the best 15 algorithms for age prediction, we note only InceptionTime as a deep learning approach at the top, followed by algorithms from the interval, convolution, and feature-based approaches.

In Figure 17, we show the normalized confusion matrix of the best algorithm from each approach. In these results, we can note that although XGB (raw data) and CIF achieved similar accuracies of 75% and 76%, respectively, the algorithms make different mistakes. For example, both algorithms concentrate most of their errors in predicting the age of younger insects (i.e., 1–4 days). However, XGB makes more wrong predictions for insects comprising 5–10 days, while CIF makes mistakes for the 11–17 days range. For InceptionTime, the errors vary from 0.15 to 0.18 for all age grades, being an adequate algorithm for age prediction.

Feature-based approach

The main advantages of this approach are the interpretability and flexibility. In some problems, the time-series shape of raw data can be well-defined according to the class (e.g., with peaks and valleys located in specific regions), helping to visually understand the classification result provided by a similarity-based algorithm as the nearest neighbor. However, the difference between the shapes of time series from different species is subtle for MIRS data, as illustrated in Figure 3. In this case, analyzing specific wavenumbers or the information extracted by the Catch-22 can support understanding the data and results. Regarding flexibility, the features can be used to train virtually any machine learning algorithm, such as Naive Bayes, Decision Trees, and Artificial Neural Networks. The limitation of this approach is that its performance heavily relies on selecting and engineering relevant features, which can be time-consuming and dependent on a domain expert. Moreover, most cases involve manually executing an additional step in the learning process. For example, the selection of 17 wavenumbers provided by González Jiménez et al. (2019) was performed through the careful analysis of chemistry and ecologist experts.

Interval-based approach

Such an approach is robust to noise and outliers since it discards most observations after selecting relevant and minor subsequences from the raw data. Once the intervals are defined, any machine learning algorithm can be trained using them, bringing flexibility to the approach. Optionally, we can extract features from the intervals to train a more robust classifier, as performed by CIF and DrCIF. An essential advantage of this approach is its efficiency and scalability in processing high-dimensional data, a usual characteristic of time series. A limitation of the interval-based approach is the loss of potentially relevant information during the interval selection phase and the influence of interval size, a parameter that requires careful attention.

Convolution-based approach

Rocket achieves comparable accuracy to the state-of-the-art ensemble methods composed of three dozen classifiers, such as HIVE-COTE (Lines, Taylor & Bagnall 2018) or complex deep learning methods, with a fraction of processing time. However, since the algorithm uses 10,000 random convolutional kernels in terms of length, weights, bias, dilation, and padding to capture relevant features, it is a non-deterministic solution, which could be considered a limitation of this approach regarding reproducibility and predictability. However, we show that MiniRocket provided similar results with a deterministic solution. We also point out that a limitation of this approach is the requirement for a lengthy time series, which is not an issue for MIRS data.

Deep learning-based approach

The classification accuracy is the main advantage of the deep learning-based approach, mainly for InceptionTime, an algorithm adapted for time series data. However, deep learning algorithms such as FCN and Time-CNN performed inferiorly than simple feature-based algorithms. The unified framework of deep learning algorithms that perform both feature extraction and classification steps is advantageous, removing the need for manual feature engineering. This approach’s limitations are the need for a large amount of labeled data for training, lack of model interpretability, and time costs for model training. While models from alternative approaches were trained using CPU processing, the algorithms employed in this methodology necessitated GPU utilization for model training within an acceptable timeframe.