Detection of Forest Fire Using Modified LSTM Based Feature Extraction with Waterwheel Plant Optimisation Algorithm Based VAE-GAN Model

Forests are crucial to maintaining our way of life because they provide a wealth of priceless resources, such as minerals and materials required for several industrial operations [1]. Beyond their obvious benefits, trees considerably improve the environment by purifying the air naturally, collecting carbon dioxide, and releasing oxygen that sustains life. Additionally, trees provide crucial habitat for a variety of species and act as a barrier against sandstorms, safeguarding crops besides maintaining ecological balance. But the widespread effects of climate change [2, 3] are mostly to blame for the increasing frequency of forest fires in recent years. High temperatures and dry circumstances encourage the spread of flames, causing significant damage to ecosystems, animal habitats, natural reserves, and a clear danger to people's lives. Notably, coniferous woods, which are distinguished by their needle- or cone-shaped foliage, are especially vulnerable to fires because the sap found in their branches is flammable [4]. Coniferous trees' dense growth patterns also contribute to the fast-moving spread of flames. Millions of acres of forest are annually destroyed as a result of this worrying trend, with catastrophic economic consequences.

Many nations, including the fires. In particular, the horrific Australian bushfires in 2020 provide as a sobering example of the intensity of these occurrences, resulting in the irreparable loss of forest resources, innumerable animal deaths, and human casualties. These flames destroyed 1,500 dwellings, approximately 500,000 animals, about 14 million acres of forest, and nearly a third of all living things [5]. Similarly, in 2018 and 2019, devastating wildfires of equal size burned large areas of the Amazon rainforest and California's forests, causing enormous losses [6]. Surprisingly, between 1992 and 2015, human activity was responsible for a whopping 85% of forest fires in the United States, with natural causes like lightning strikes and the effects of climate change accounting for the remaining 15%. More stringent regulations and reasonable practises may have prevented many of the human-caused fires [7]. It is noteworthy to note that during the worldwide COVID-19 epidemic, the frequency of forest fires decreased as several nations instituted lockdown measures that restricted human activity, hence decreasing the risk of human-induced fires [8].

DL networks have been a very successful method for addressing the crucial problem of forest fire detection. DL has shown its aptitude in a number of fields, including autonomous machine translation [9] and image and video categorization. This was made possible by DL's capacity to automatically extract and categorise properties from data stored on the same network [10]. DL is especially good at detecting forest fires because to large datasets and improved processing capacity. For both ground-based and aerial images, DL-based systems have shown their superior performance over conventional machine learning techniques in handling the complexity of forest fire categorization and detection [11, 12]. The rise of automated DL fire detection systems holds enormous potential for the development of AI fire models are a useful tool for addressing this pressing environmental issue since they can quickly and precisely identify and track flames inside the camera's field of vision [13].

In subsequent work, we intend to look into the following possible improvements:

Attention Mechanisms: By including attention mechanisms in the LSTM architecture, the model may be better able to identify significant patterns in the data by focusing on pertinent elements at certain time steps.

Ensemble approaches: By integrating several LSTM models or other recurrent neural network (RNN) types, one can improve the model's prediction presentation and robustness by using ensemble approaches.

Hybrid Architectures: Investigating hybrid architectures that integrate transformer models or convolutional neural networks (CNNs) with LSTM may offer supplementary benefits in sequence modelling and feature extraction.

This paper's primary contributions are:

-The dataset for this investigation was first pre-processed. This research then improves the quality of information related to forest fires by using cutting-edge DL methods and extracts useful temporal characteristics using LSTM networks.

-The performance and convergence of the LSTM model are improved by the application of the MFFA for weight optimisation.

-As a tool for classification, the VAEGAN model is used. The WWPA for hyperparameter tuning promises to greatly improve the ACC and efficacy of systems for detecting forest fires. Results are analyzed using five parameter metrics.

The rest of the research is structured like shadows: The literature review is presented in Section 2, followed by a brief explanation of the proposed model in Section 3, the results and validation analysis in Section 4, and finally, a conclusion and summary are given in Section 5.

1.png

3.1 Dataset description

As exposed in Figure 2, the dataset utilised in this study consists of 3000 images of forest fires that were captured by drones and video surveillance equipment in various forest environments. It also includes additional forest fire datasets discovered using web crawling techniques and publicly available forest fire datasets [23]. This collection's 1000 images all have hand annotations. Then, these 1000 annotated images were divided into two subsets: 300 served as a specialist test set to assess the model's ACC, and 700 were set aside for training purposes to construct a prototype forest fire detection model. 2000 images of unlabelled forest fires were also included in the dataset and utilised in the training process.

2.png

Figure 2. Schematic diagram of forest fire data set

3.2 Preprocessing

The dataset provides a wide range of photos taken from different perspectives, enabling the algorithm to identify between forest fire and non-fire events with greater ACC. With the help of this information, the model is equipped to recognise forest fires based on two key criteria: the existence of fire flames and the presence of fire flames mixed with smoke. Up to this point, our main attention has been on the exacting standards used to divide up the dataset into an equal number of images with fire (1) and those without fire (0) [24]:

Fire (1): Images of forests and mountain ranges that are enveloped in flames and/or smoke brought on by fires.

3.png

Figure 3. Images from (a) and (b) Fire class and (c) and (d) No-Fire class

No-Fire (0): Consisting of a wide variety of pictures showing forest and mountain vistas devoid of fire. This categorization method was developed to make it easier to train models with a variety of images while avoiding misunderstanding with situations that could seem similar, such mountain sunsets.

The goal of this meticulous dataset refining method was to improve overall model performance and streamline the model training procedure. As they were the most contextually relevant to our study objective at this dataset curation phase, we particularly cropped pictures of fires in mountainous or forest settings. After then, every image was scaled consistently such that it had the same size, 250×250 pixels. These preprocessing methods were crucial in helping the model successfully include important information about forest fires. Figure 3 shows visual representations of both the fire and no-fire categories within forest fire dataset.

3.3 LSTM feature extraction

After the preprocessing phase, the input characteristics are then passed to the LSTM module, a crucial component of our methodology [25]. Due to the huge quantity of data gathered from the dataset, a typical RNN would not be sufficient for our purposes. Gradient disappearance and explosion issues are addressed via a customised RNN iteration known as LSTM. During the training phase, RNN creates temporal connections between prior states and the inputs to provide predictions. RNN, on the other hand, finds it challenging to maintain the past because to its limited memory capacity, especially when dealing with massive volumes of time series data. However, LSTM excels at classifying enormous time series datasets and locating temporal correlations. Its use covers several sectors and yields outstanding results for tasks like speech recognition and image classification.

The LSTM architecture seen in Figure 4 has special memory cells designed to make use of prior knowledge and maintain key characteristics from massive volumes of time series data. These memory cells may store and apply the information that was learnt, allowing the model to process and classify input effectively.

4.png

Figure 4. Architecture of LSTM

The output gate, also known as the o_t gate, the forget gate, also known as the ft gate, and the input gate all play distinct roles in regulating information flow in the LSTM architecture.

The major responsibility of the forget gate is to decide what data to preserve and what to discard within the cell state. It does this by conducting a pointwise multiplication operation using the inputs $x_t$, the current input, and $h_{t-1}$, the previous hidden state information. By using the sigmoid activation function, the forget gate generates an output that is either 0 or 1. Keeping important information in the cell state is indicated by a value of 1 , whilst removing unimportant information is indicated by a value of 0 . The forget gate, input gate, and output gate's core characteristics are explained in literature [26] by Eq. (1) to Eq. (6).

$f_t=\sigma\left(W_f\left[h_{t-1}, x_t\right]\right)+b_f$               (1)

The forget gate, or $f_t$, is subjected to a bias called $b_f$, and the process weight $W_f$ stands for that. The forget gate is activated using the activation function, represented by the letter "s," to allow choice. Next, a critical decision on whether data should be kept in the cell state, $C_t$, must be made by the input gate. It considers both the input, $x_t$, and the preceding hidden state, $h_{t-1}$ to arrive at this conclusion. Eq. (2) and Eq. (3) describe the pointwise multiplication of the forget gate, $f_t$, and hyperbolic tangent (tanh) activation functions in this decision-making process:

$i_t=\sigma\left(W_i\left[h_{t-1}, x_t\right]\right)+b_i$            (2)

$C_t=\tan h\left(W_c\left[h_{t-1}, x_t\right]\right)+b_0$            (3)

While $b_i$ and $b_c$ stand for the biases of the neural network, $W_i$ and $W_c$ stand for the weights connected to the input gate $\left(i_t\right)$ and output gate $\left(o_t\right)$. The information about the previous concealed cell state is denoted by the word $C_t$. Eq. (4) shows how we combine Eq. (2) and Eq. (3) to conduct a pointwise addition operation in order to update the current cell state, $C_t^{\prime}$:

$C_t^{\prime}=\left(f_t \times\left(C_t\right)\right)+\left(i_t \times\left(C_t\right)\right)$              (4)

The output gate is calculated in Eq. (5). The current input, represented as $x_t$, and the prior hidden state, $h_{t-1}$, which incorporates the activation function $s$, are both used in this gate. In order to further hone the output network, a bias term called $b_0$ is included.

$o_t=\sigma\left(W_o\left[h_{t-1}, x_t\right]\right)+b_o$            (5)

A pointwise multiplication operation is used to combine the information from the cell state, $C_t$, with the updated output gate, indicated as $o_t$. The hidden state that follows, $h_t$, is produced by this procedure and is represented in Eq. (6).

$h_t=\sigma\left(O_t \times \tan h\left(C_t^{\prime}\right)\right)$              (6)

The LSTM model may be improved significantly by using optimal parameters. The efficacy of feature extraction is substantially influenced by these characteristics. With the exception of the completely linked dense layer, 50 more neurons have been added to each layer to aid in better training. Furthermore, a 20% dropout rate has been used to allay any overfitting concerns.

3.3.1 Weight optimization in LSTM using MFFA

Firefly algorithm. In situations where it is necessary to optimise not one, but several competing objectives at once, WWPA can handle these difficulties. By efficiently exploring the trade-off between different objectives, WWPA can discover Pareto-optimal solutions that represent the best compromise between competing goals.

Efficient exploration and exploitation. WWPA balances exploration (searching diverse regions of the hyperparameter space) and exploitation (exploiting promising 0refine solutions) effectively. This balanced exploration-exploitation trade-off enables WWPA to quickly converge to high-quality solutions while avoiding premature convergence to suboptimal regions.

5.png

6.png

$v=Dis(u) \epsilon[0,1], u=Gen(w)$                (14)

$J_{G A N}=\log ({Dis}(u))+\log (1-{Dis}({Gen}(w)))$                  (15)

In this instance, "w" is a random variable represented by the probability density function p(w), and "u" represents a genuine sample. While it's true that GANs may produce synthetic data without using density functions in certain circumstances, such as when dealing with unbalanced data, there are other situations when getting fresh samples from the generator with specified distributions might be advantageous. In order to do this, a generative model is built by using the VAE's decoder component as the GAN's generator. A visual illustration of the VAE-GAN model's structure is shown in Figure 6.

The following is a representation of the loss function of the VAE-GAN [32].

$J_{V A E-G A N}=J_{\text {prior }}+J_{D i s_l}+J_{G A N}$             (16)

$J_{D i s_l}=-E_{q(Z \mid x)}\left[\log p\left(D i s_l(x) \mid z\right)\right]$             (17)

where, $\operatorname{Dis}_l(x)$ denotes a Gaussian observation model with an identical covariance and $D i s_l(\tilde{x})$ as the mean.

3.4.1 Hyper parameter tuning using WWPA

The WWPA is described in this section. Here, we explore the motivation behind the method and provide a thorough mathematical explanation of how it works.

7.png

Figure 7. Image of the WW plant. (a) A side view of a shot that is free-floating and loaded with traps. (b) Frontal view with both open and shut traps. (c) Just one open trap. (d) An open trap schematic illustration

Inspiration of WWPA. Aldrovanda vesiculosa, the alternative name for the waterwheel (WW) plant, has broad petioles that contain its unusual traps, which are barely 1/12 inches in size and resemble tiny transparent flytraps [33]. The interactions with other aquatic plants won't cause these traps to deteriorate or unintentionally activate because of their skilled design. They are protected by a ring of bristles that resemble hair. These traps include a variety of hook-like teeth along their edges that interlock when the trap catches its victim, much like the teeth seen in a typical flytrap. The Aldrovanda trap has more than 40 elongated trigger hairs, compared to the normal 6-8 trigger hairs on a Venus flytrap. When one or more triggers are pulled, these trigger hairs allow the trap to shut. These carnivorous plants have trigger hairs as well as glands that emit acid to help in the digesting of their caught prey. The sealant and the plant's interlocking teeth trap the prey. By leading the prey towards the hinge at the base of the trap, the seal successfully catches the prey. The body fluids of the prey are extensively broken down by the plant's digestive secretions, and any leftover material is excreted. Similar to a flytrap, an Aldrovanda trap can hold and digest two to before filling to capacity. The infrastructure of the waterwheel plant is shown in Figure 7.

The WWPA mathematical model. This section describes how WWPA is set up before going into detail about how the WW's location is updated throughout both the exploration and exploitation phases using a classical based on the actual behaviour of WWs.

Initialization. The population-based approach of the WWPA tries to locate the ideal solution by using the collective search capabilities of its population members within the solution space. Each of the WWs that make up the population of this algorithm represents a potential resolution to the problem and has a specific set of problem-related variables. Vectors may be used to formally represent these responses. The whole population of the WWPA, which consists of all WWs as given in Eq. (18), may be represented by a matrix. At the beginning of WWPA, the positions of these WWs inside the solution space are initialised at random using Eq. (19).

$P=\left[\begin{array}{c}P_1 \\ \vdots \\ P_i \\ \vdots \\ P_N\end{array}\right]=\left[\begin{array}{ccccc}p_{1,1} & \cdots & p_{1, j} &\cdots & p_{1, m} \\ \vdots & \ddots & \vdots & \ddots & \vdots \\ p_{i, 1} & \cdots & p_{1, j} & \cdots & p_{i, m} \\ \vdots & \ddots & \vdots & \ddots & \vdots \\ p_{N, 1} & \cdots & P_{N, j} & \cdots & P_{N, M}\end{array}\right]$               (18)

$\begin{aligned} & p_{i, j}=l b_j+r_{i, j .}\left(u b_j-l b_j\right) \\ & i=1,2, \ldots, N, j=1,2, \ldots, m\end{aligned}$             (19)

where, $\mathrm{N}$ stands for the quantity of WWs, and $\mathrm{m}$ stands for the quantity of variables. The limits for the $\mathrm{j}$-th issue variable are represented by $l b_j$ and $u b_j$, while the variable $r_{i, j}$. has random values between $[0,1]$. The population matrix of WW locations is designated as $\mathrm{P}$, where $p_i$ is the $\mathrm{j}$-th $\mathrm{WW}$, which corresponds to a problem variable, and $p_{i, j}$ denotes its i-th dimension. The target function may be calculated for each WW as they each stand in for a possible answer to the issue. Studies have proven that a vector may be used to properly represent the variables that make up the objective function in Eq. (20).

$F=\left[\begin{array}{c}F_1 \\ \vdots \\ F_i \\ \vdots \\ F_N\end{array}\right]=\left[\begin{array}{c}F\left(X_1\right) \\ \vdots \\ F\left(X_i\right) \\ \vdots \\ F\left(X_N\right)\end{array}\right]$              (20)

where, F denotes the vector containing all of the values for the objective functions, and $F_i$ is the i-th WW. Most of the time, objective are used to choose the best keys. The best candidate solution, indicated by the highest objective function value, and the worst candidate solution, indicated by the lowest value, are thus the most crucial metrics. Given that WWs pass across the search region at varying speeds throughout each iteration, the optimal solution could evolve over time.

Stage 1: Recognising positions and hunting for insects

WWs flourish as skilled hunters capable of locating pests because to their exceptional sense of smell. A WW attacks as soon as it notices an insect nearby, focusing on the bug's particular location and starting a chase to trap it. For the first stage of its populace update procedure, the WWPA simulates the behaviour of a WW. The WWPA improves its exploration skills by modelling the hunting behaviours of WWs, allowing it to find ideal places while avoiding being caught in local optima. This is accomplished by simulating the large motions of the WW as it approaches the insect within the solution space. This simulation of the WW's approach to the insect is integrated using an Eq. (21) as shown below, to predict the WW's new location. If moving the WW to the newly determined position increases the charge of the target function, as shown in Eq. (21) and Eq. (22), the old position is abandoned in favour of the new one.

$\vec{W}=\vec{r}_{1 .}(\vec{P}(t)+2 K)$               (21)

$\vec{P}(t+1)=\vec{P}(t)+\vec{W} .\left(2 K+\vec{r}_2\right)$              (22)

Alternately, the WW's location may be changed using the following Eq. (23) if the results do not improve after three consecutive repetitions:

$\vec{P}(t+1)=\operatorname{Gaussian}\left(\mu_p, \sigma\right)+\vec{r}_1\left(\frac{\vec{P}(t)+2 K}{\vec{W}}\right)$            (23)

In this Eq. (23) the random variables $\vec{r}_1$ and $\vec{r}_2$ may, respectively, have values of 0 or 2 and 0 or 1 . The vector $\vec{W}$ represents the radius of each circle that the WW plant evaluates as possible regions, and $\mathrm{K}$ is a variable with values between 0 and 1 .

Stage 2: Carrying the insect in the suitable tube (exploitation)

The behaviour of insects being transported to feeding tubes by waterwheels serves as the model for the second stage of population updates in WWPA. By simulating this behaviour, WWPA may improve its convergence towards answers that are very similar to those it has already collected. WWPA modifies the location of the WW inside the problem region by simulating the insect's journey to the proper tube for ingestion. To do this, each WW was first placed in a fresh, arbitrary position that represented a "favourable region for insect consumption," according to the WWPA designers. Eq. (24) and Eq. (25) show that if the objective function produces a better value at the new position, the WW is moved.

$\vec{W}=\vec{r}_3\left(K \vec{P}_{\text {best }}(t)+r_3 \vec{P}(t)\right)$                    (24)

$\vec{P}(t+1)=\vec{P}(t)+K \vec{W}$                 (25)

After three iterations, if the response still doesn't indicate progress, the method incorporates a mutation process akin to the exploration stage. The algorithm undergoes certain alterations throughout this mutation phase to avoid being stuck in local minima. In this adaptive method, the current answer, indicated as $\vec{P}$ at iteration $t$, is represented as $(\mathrm{P})$ at iteration $\mathrm{t}$, and the ideal solution is written as $\vec{P}_{\text {best }}$. A random mutable with values between $[0,2]$ is $\vec{r}_3$. This strategy aids the algorithm's robustness and effective escape from local maxima.

$\vec{P}(t+1)=\left(\vec{r}_1+K\right) \sin \left(\frac{F}{C} \theta\right)$                 (26)

where, [-5, 5] is a range for the random variables F and C's values. Additionally, using the Eq. (27), K's value falls down rapidly.

$K=\left(1+\frac{2 * t^2}{T_{\max }}+F\right)$              (27)

The proposed WWPA's pseudocode

The iterative process used by the WWPA has the following three phases. Once the first and second steps are complete, each WW is moved in the third and final stage. This adjustment, which results in the key adjustments of the best candidate solution, is based on a comparison of target function values. The WW positions are then adjusted in preparation for the next iteration. This repeated process is carried out till the algorithm achieves its conclusion. Implement WWPA as instructed, following the detailed instructions in Algorithm 1. Based on its iterative development, WWPA offers the most promising candidate solution after being completely deployed.