A framework for optimal agent deployment and opportunistic routing in flying Ad-Hoc networks for precision weather forecasting

Abstract

Drone based precision weather prediction and forecasting is a prevailing area of research in recent times. Modern flying ad-hoc network (FANET) leverages advanced low latency communication protocol between small flying drones and base stations. In this work, a FANET framework for efficient and precision weather prediction and forecasting that leverages ultra low latency opportunistic message transfer protocol (MTP) is proposed. An, edge empowered intelligent computing quasi-stationary sink nodes are employed to develop the FANET ecosystem further. Flying sensor nodes and edge computing quasi-stationary sink nodes are deployed in the target area in order to achieve optimal sensor coverage minimizing the cost of data transmission power thus ensuring long endurance operation of drones. An ensemble model is deployed in edge level for localized weather prediction. Opportunistic routing strategies are utilized for converge casting. The experimental results show the maximum of 0.98 message delivery ratio and a minimum of 820 ms latency is achieved through opportunistic MTP. An accuracy of 96.5% is achieved in predicting the temperature and humidity data metric.

Introduction

Severe weather conditions including thunderstorms, strong winds, fog, hail, and drought have an impact on the economy and public safety anywhere in the world. It has a major role in agriculture and farming, a study presented in Motha et al. ¹ Extreme conditions can have a negative impact on crop yield and, consequently, agricultural output. The majority of crops are directly dependent on high temperatures, less precipitation, flooding, and premature freezing during crucial growth stages. Indirect effects on crops can also arise from influences on pest organisms, nutrient dynamics, and soil processes. Using weather and climatic data effectively, one can reach optimized production levels by reducing the loss of soil, nutrients, and water resources. Weather assistance is also essential in air traffic management at every level of control, navigation, and path planning.^2,3 In Gonzalo et al. ⁴ wind modeling techniques is proposed for automation of air traffic management system.

The atmospheric boundary layer is the lowest part of the troposphere closest to Earth’s surface. This part of the atmosphere interacts with the surfaces of the Earth’s land, ocean, and ice. These interactions influence the entire climate system and global weather patterns. This can affect where and how far pollutants travel, how much rain falls during a storm and course of hurricanes and tornadoes. Wouters et al. ⁵ studied land–atmosphere interactions model which is required to accurately predict and forecast climate. In Darack ⁶ an unmanned aircraft penetrate cyclonic storm for more than 7 h and flew within Hurricane Noel to measure atmospheric condition. However, the atmospheric boundary layer has not yet provided enough observations for weather models to reliably forecast local weather occurrences. In this study, multi-drone system is used to close the data gap.

The drones are able to obtain accurate vertical profiles of wind, temperature, humidity, and atmospheric pressure at altitudes between $10$ and $800\mathrm{m}$ . In this work, drones designed explicitly for weather reasons are launched. The gathering of air data samples (like temperature, pressure, humidity, wind speed, precipitation, etc.) using drones offers a considerable improvement over the methods currently used for data collecting and has the potential to greatly improve the accuracy of weather forecasting models.

New technologies, including the Internet of drones (IoD) with UAV networks, can offer tremendous potential in Smart Farming and Precision Agriculture applications by making it possible to collect environmental data in real-time and intelligent edge enabled weather forecasting. Agricultural forecasting is the process of predicting many weather variables, such as the amount of sunshine hours, the likelihood of dew, relative humidity, rainfall, temperature, and winds, which are crucial for farming operations. For instance, as farming and agriculture as a whole mostly depend on seasons and weather, weather forecasting enables effectively plan farm operations, such as planting, irrigation, fertilizer application, pruning/weeding, harvesting, or cattle mating. One benefit of weather forecasting is its preventative nature, which gives farmers awareness of climatic changes in an area before they happen!. Smart farming system is proposed with weather forecast modeling and crop prediction in Shandilya and Khanduja. ⁷

To gather localized data predictions and leverage an overall forecasting framework it is essential to gather ambient weather data through UAVs in real-time manner. In the proposed work an IODT framework amalgamating intelligent routing and computing paradigm is proposed. The proposed framework divides the area of interest into optimized number of regions and places Quasi-stationary base stations as sink base station for localized data processing. Weighted ensemble approaches are proposed and have been tested to work on light weighted base stations. Each of the localized areas are catered through the placement of coverage and relay UAV nodes for data collection. The routing of data is catered through a proposed opportunistic routing strategy toward a main base station. The main base station is part of the application level control where forecasting models are deployed for further analytics. The proposed routing approach has been compared with other state of the art methodologies. A summary of the contribution is as follows.

1. An edge enabled flying ad-hoc network (FANET) architecture is proposed for precision weather prediction and forecasting using an opportunistic message transfer scenario.

2. Deployment of flying sensor nodes and stationary data sink node with optimum environment coverage and collision avoidance.

3. The use of flying stationary nodes as a mobile edge computing platform to execute artificial intelligence models using a low-resource edge device.

4. Designing ML model for precision weather forecasting at ground station using high end resources.

The work is presented in the following manner. Section “Related research” discusses the past works carried out in the domain of weather monitoring and drone based sensing. Section “Proposed FANET architecture” and “Quasi-stationary base station deployment and routing strategy” explains the proposed FANET architecture that discuss the sensor node placement and the proposed routing strategy employed. Section “Predictive modelling” gives the overview of the edge level prediction and forecasting procedure employed. The work showcases the extensive experimentation carried out through different comparative analysis in Section “Conclusions” Experimental Results and Concludes in Section “Future scope.”

Related research

Precise weather forecasts are critical to humans life since they affect everything from agriculture to air travel and alerts about extreme weather conditions like storms and heavy downpours to individual’s daily activities. Numerical weather prediction depends on atmospheric observations of pressure, temperature, humidity, wind direction, and wind speed. A portion of these measurements may be obtained through weather stations, aircraft, ⁸ radiosondes, ⁹ dropsondes (meteorological sensor packets dropped from high-altitude platforms; Hock and Franklin, ¹⁰ satellite-based remote sensing techniques, ¹¹ and ground-based remote sensing techniques. ¹² It is observed that weather stations and weather balloons have long been used by scientists to gather meteorological data for modeling. A weather station’s fixed location and close proximity to the earth’s surface are its two main drawbacks. It cannot obtain data from the upper reaches of the boundary layer. Weather balloons have the problem of being uncontrollable despite their much higher maximum altitude. They follow the direction of the wind. In addition, weather balloons are usually unretrievable. Drone measurements may be used to close these holes. Until now, these devices have not been considered a practical substitute for conducting measurements in the upper troposphere. The development of a drone system that can sound the atmosphere up to tropopause above the ground on its own is presented in Bärfuss et al. ¹³ for Antarctic and mid-European region. However, flight test of such scenario is subjected to take permission. In this article, drones are restricted up to an altitude of $800\mathrm{m}$ which is a low traffic zones. Data collected by other means discussed above (available in CLOUD) are fused together by LSTM algorithm at ground station for precise prediction and forecasting of weather.

The Internet of Things ecosystem is absolutely necessary to transmit and receive the weather data and control command in an automated drone environment. Recently, to facilitate autonomous devices in sixth-generation (6G) networks and ensure seamless wireless connectivity in inaccessible and hard-to-reach areas, the integration of software-defined networking (SDN) with unmanned aerial vehicles (UAVs) has demonstrated significant potential (see Siddiki Abir et al. ¹⁴ , Rahman et al. ¹⁵ , and Sharma and Mehra ¹⁶ ). In the case of a remote environment to send and receive the crucial information and control commands, a sophisticated network model with data routing and with low latency communication is highly recommended. Also, the flying sensor nodes and edge enabled data sink nodes are to be placed ensuring optimum sensor coverage and collision avoidance. Optimization of UAV assisted network coverage system has immense application such as in rescue mission, ¹⁷ in building UAV charging station. ¹⁸ A mathematical concept called Voronoi tessellation, commonly referred to as Voronoi diagrams or Voronoi cells, is employed in many disciplines, including computer science and telecommunications. ¹⁹ It is a technique for dividing up polytopic space into polytopic areas depending on their closeness to a collection of points. A Centroidal Voronoi Tessellations (CVT) is a unique kind of voronoi division where the center of a voronoi cell ensures optimum coverage to all points inside the cell. Various applications of centroidal voronoi tessellations are discussed in Faber et al. ²⁰ In Gonzalez and Hamalainen, ²¹ a CVT based cellular 6G network topology is suggested for optimal performance in time and space. In this work, a CVT based FANET topology is proposed where mobility of each sensor node is restricted to its Voronoi tessellation. The deployment objectives of UAV nodes are to achieve optimum environment coverage and collision avoidance. A detailed discussion of FANET architecture is given in next section.

In Mishra and Joshi, ²² a method to forecast temperature variation is presented using the concepts of linear regression and artificial neural networks. The study in Raimundo et al. ²³ used machine learning regression approaches such as Linear Regression, Decision Trees, Random Forests, and Neural Networks to forecast agricultural fields’ weather conditions for smart irrigation systems. Depending on the field location and day, the suggested methods can forecast the temperature, precipitation, wind speed, and evapotranspiration. A prediction model for weather forecasting based on deep learning has been built in Sathya et al. ²⁴ A comparison has been made between the created LSTM (Long Short-Term Memory) based deep learning model and classic machine learning algorithms like SVR (Support Vector Regression) and RF (Random Forest), as well as GRU (Gated Recurrent Unit), one of the recurrent neural network models. The experimental findings produced according to $R^2$ metrics, Pearsons correlation coefficient, MAE (mean absolute error), RMSE (root mean square error), and MSE (mean squared error) have been compared for each model and algorithm. The developed LSTM-based deep learning model outperforms other compared models and techniques, according to experimental results in Sathya et al. ²⁴ Accurate rainfall prediction using Artificial Neural Networks compared with Deep Learning Neural Networks is proposed in Rayudu and Roseline. ²⁵ The research in Du ²⁶ proposes an ensemble machine learning-based wind power production forecasting approach that leverages weather station meteorological observation data along with weather data anticipated by prediction model.

Proposed FANET architecture

The ad-hoc network ecosystem is divided into three logical planes. Data collected from environmental sensing is gathered by the Action Plane and sent to the Processing Plane via the Networking Plane, which connects all the system’s entities and handles data storage and processing. Figure 1 displays the UAV-based flying ad-hoc network (FANET) system considered for this paper. This system is made up of several heterogeneous components, such as unmanned aerial vehicles, quasi-stationary Base Stations, and data transmission lines. Target area is divided in an optimized manner with allocated Base Station, Coverage, and Relay UAV nodes.

Figure 1.

Proposed FANET architecture.

The sensors equipped on UAV devices collect ambient weather information and transmit the data to the quasi-stationary base station at a high altitude platform. Based on the data transmission and reception range of a node, the target region is divided into a number of uniform cells. A uniform cell area leads to uniform load distribution among the sensor nodes to cover the whole target region for collecting data. The quasi-stationary base station acts as a sink node. It is deployed onto each cell to a location to achieve optimal coverage to any point inside the cell. It also carries sensors in order to realize temperature change with altitude.

Each quasi-stationary base station acts as an edge-level processing control responsible for predicting the localized weather conditions and acts as a buffer to store localized weather parameters. A group of relay UAV nodes is utilized to route the localized weather label and parameters to the main base station for further application-level control are called data mules. Data is transmitted through an opportunistic message transfer protocol. The proposed scenario involves the consideration of three different types of message transfer protocols (MTP), namely opportunistic flooding methods, probabilistic flooding, and opportunistic forward methodology. Owing to the advantage of opportunistic message transfer protocol, the transmission range of sensor nodes can be chosen ten times smaller than the reception range of sink node in each cell. This will minimize the cost of transmission power and ensure long-duration operation of sensor nodes. Experimentation has been carried out utilizing the three MTPs. Shortest path map-based mobility is considered as the model for UAV nodes over the whole considered Voronoi region for routing. The main base stations are equipped with further forecasting and application-level control.

Quasi-stationary base station deployment and routing strategy

Consider that $k$ quasi-stationary sensors are to be placed over a 2-D convex area $\mathbb{W}$ . $x_i\in {\mathbb{R}}^2$ represents the position of sensor $i$ , while $y$ denotes a point in $\mathbb{W}$ . Let the distance function $f(\parallel y,x_i\parallel )$ denotes the performance degradation of the sensor $i$ placed at $x_i$ of measurement at point $y\in \mathbb{W}$ . Assume, the function $f$ is being monotonically increasing, $f(\parallel y,x_i\parallel )=\parallel y-x_i{\parallel }^2$ . According to Cortes et al., ²⁷ an integral over $\mathbb{W}$ is the total “sensing performance” or coverage metric. To avoid overlapping in the sensing area, the relative contributions of each sensor is weighted in the performance function as

$$J\left(x_1,\cdots ,x_n\right)=\underset{\mathbb{W}}{\int }\underset{i\in \{1,\cdots ,n\}}{\min }f\left(\parallel y,x_i\parallel \right)\rho \left(y\right)dy$$

(1)

The problem of optimal deployment of sensors $x_1,\cdots ,x_n$ to provide precise distributed sensing is known as locational optimization problem which leads to a minimization problem of the cost $J$ . Here, $\rho$ is the scalar probability density function on $\mathbb{W}$ . For exploring an environment uniformly, $\rho \left(y\right)$ can be set as $\mathrm{Volume}{\left(\mathbb{W}\right)}^{-1}$ .

By utilizing the Voronoi cell $\left({\mathbb{V}}_i\right)$ concept and expressing the measure element as $d\rho \left(y\right)=\rho \left(y\right)\mathrm{dy}$ , the locational optimization function may be reformulated as follows:

$$J\left(x_1,\cdots ,x_n\right)=\sum _{i=1}^n\underset{{\mathbb{V}}_i}{\int }f\left(\parallel y,x_i\parallel \right)d\rho \left(y\right)$$

(2)

The problem of partitioning a convex polytopic search region $\mathbb{W}$ into a number of uniform cells is addressed here by using the geometric tool of Centroidal Voronoi partition.^19,20 A quasi-stationary data sink node with edge computing ability is deployed in each cell such that its coverage area of data reception of telemetry system is optimized. This is achieved by solving the optimization problem for the coverage of a bounded target environment by $k$ sensors described above. This issue is referred to as coverage problem,^27–29 or deployment problem ³⁰ in the literature. Thus, a decentralized control laws for the placement of each stationary agents that achieve uniform and optimum coverage of the target environment $\mathbb{W}$ is proposed in this paper.

Coverage control

Decentralized control for MAS

Imagine a system with multiple agents, each of whom moves in a three-dimensional metric space. Let the manifold with boundary $\mathbb{W}$ represent the configuration space of each vehicle. On $\mathbb{W}$ , a Riemannian metric system is defined. Let $x_i\in {\mathbb{R}}^2$ represent the position of the $i$ th agent or stationary node with the $i$ th data reception sensor at time $t$ for $i\in \{1,\cdots ,n\}$ . and let

$${\stackrel{·}{\mathrm{x}}}_i=u_i$$

(3)

where the position of $x_i$ and its neighbors serve as the only sources of local knowledge that the control $u_i$ can rely on. This is defined as distributed cooperative law between vehicles given that the control law depend solely on neighbors.

The decentralized control $u_i$ is investigated to achieve uniform coverage on $\mathbb{W}$ . According to the gradient control law $u_i$ is defined as

$$u_i(x_1,\cdots ,x_n)=-{\nabla }_{i}J(x_1,\cdots ,x_n)$$

(4)

which depend on the position $x_i$ and its neighbor ${\mathcal{N}}_i$ . ${\nabla }_{i}J$ represents the gradient of the objective function $J$ with regard to the parameter $x_i$ .

Concept of Voronoi partition

The works in Aurenhammer ³¹ and Atsuyuki Okabe and Sugihara ³² provide an overview of Voronoi diagrams; Okabe et al. ¹⁹ discusses ideas and applications; and Klein ³³ discusses abstract Voronoi diagrams. The topic of centroidal Voronoi tessellations is covered in Faber and Gunzburger. ²⁰

In a metric space $\mathbb{W}$ , let $\{x_1,\cdots ,x_n\}$ be a set of points. Additionally, let ${\mathbb{V}}_i=\mathbb{V}\left(x_i\right)$ be the collection of all points $y\in \mathbb{W}$ such that, for any $j\ne i$ , $\parallel x_i-y{\parallel }_2\le \parallel x_j-y{\parallel }_2$ . This is the Voronoi region corresponding to $x_i$ . The border of any $V_i$ in a finite dimensional Euclidean space $\mathbb{W}$ is a convex polytope. For the generators $\{x_1,...,x_n\}$ , the collection of divisions $\{{\mathbb{V}}_1,\cdots ,{\mathbb{V}}_n\}$ is referred to as the Voronoi tessellation. The Voronoi cells have a mutually disjoint interior.

Also note that the interior must be non-empty. If two distinct cells ${\mathbb{V}}_i,{\mathbb{V}}_j$ have a non-empty intersection that is, they share borders

$${\mathbb{V}}_i\cap {\mathbb{V}}_j\ne Ø,i\ne j$$

(5)

then $x_i$ is referred as $x_j$ ’s Voronoi neighbor.

The partitioning of $\mathbb{W}$ and the agent neighborhoods are illustrated in Figure 2. In this figure, Voronoi partitions are generated based on the agents’s initial positions. The borders between agents are positioned in the middle of the neighboring agents. Intuitively agents are neighbors if they share a border. The set of neighbors of an agent $i$ is represented by ${\mathcal{N}}_i$ , that is,

$${\mathcal{N}}_i=\{j\ne i\in |{\mathbb{V}}_i\cap {\mathbb{V}}_j\ne Ø\}$$

(6)

Figure 2.

Agent neighborhoods. Agent positions and Voronoi cells are denoted by ${\mathbf{x}}_i$ and the ${\mathbb{V}}_i$ respectively. The union of all Voronoi cells ${\mathbb{V}}_i$ is the environment $\mathcal{W}$ .

Control solution of optimal coverage problem

According to Cortes et al., ²⁷ the partial derivative of optimization function defined in (4) is:

$$\frac{\partial J}{\partial x_i}=\underset{{\mathbb{V}}_i}{\int }\frac{\partial }{\partial x_i}\left(\parallel y-x_i{\parallel }^2\right)d\rho \left(y\right)$$

(7)

Thus the control law in (4) becomes

$$\begin{array}{c}\multicolumn{1}{c}{u_i(x_1,\cdots ,x_n)=-{\nabla }_{i}J(x_1,\cdots ,x_n)}\\ =-G^{-1}\underset{{\mathbb{V}}_i}{\int }\frac{\partial }{\partial x_i}(\parallel y-x_i{\parallel }^2)d\rho \left(y\right)\hfill \end{array}$$

(8)

An important aspect of the decision to use $f(y,x_i)=\parallel y-x_i{\parallel }^2$ is its elegant form of gradient descent approach for locating the performance function’s stationary points, which leads to optimality when using the centroidal Voronoi partition.

$$\frac{\partial J}{\partial x_i}=2M_{{\mathbb{V}}_i}\left(x_i-C_{{\mathbb{V}}_i}\right)$$

(9)

where $x_i$ is the generator of the Voronoi cell ${\mathbb{V}}_i$ and $C_{{\mathbb{V}}_i}$ is its centroid, and the mass of the corresponding voronoi cell, defined below, is denoted by $M_{{\mathbb{V}}_i}$ (for further information, see e.g. Faber and Gunzburger ²⁰ ).

$$\begin{array}{c}{M}_{{\mathbb{V}}_i}=\underset{{\mathbb{V}}_i}{\int }\rho \left(y\right)\mathrm{dy}\hfill \\ C_{{\mathbb{V}}_i}=\frac{1}{M_{{\mathbb{V}}_i}}\underset{{\mathbb{V}}_i}{\int }y\rho \left(y\right)\mathrm{dy}\hfill \end{array}$$

Using (9) in (3) and (4), following is the agent closed loop dynamics optimizing the performance function defined in (1)

$${\stackrel{.}{\mathrm{x}}}_i=\mathrm{-}\frac{\partial J}{\partial x_i}=2M_{{\mathbb{V}}_i}\left(C_{{\mathbb{V}}_i}-x_i\right)$$

(10)

Each agent is guided by these dynamics toward the center of its Voronoi cell. Stated differently, the equilibrium state is attained when every agent is situated in the centroid of its corresponding Voronoi polygon. Such a partition is referred to as a centroidal Voronoi partition. The focus is on the case where $x_i=C_{{\mathbb{V}}_i}$ , meaning the mass centroids of the Voronoi regions ${\mathbb{V}}_i$ are the points $x_i$ , which serve as generators for those areas. See Figure 3. A recursive algorithm known as Lloyd’s algorithm originally developed in another purpose ³⁴ provides iterative solution to reach an optimum partition.

Figure 3.

Voronoi partitions in $\mathbb{W}$ for final configuration; initial generation points are marked with green ‘o’; final configurations are marked with blue ‘+’ which converge to the centroid of each cell marked with red ‘o’.

Algorithm 1.

Lloyd’s Algorithm

1: Choose a starting set of $n$ points $x_1,\cdots ,x_n$ in $\mathbb{W}$ using Monte Carlo method. 2: Build the Voronoi partition ${\mathbb{V}}_1,\cdots ,{\mathbb{V}}_n$ of $\mathbb{W}$ , where $x_1,\cdots ,x_n$ are the generators. 3: Determine the mass centroids of the Voronoi regions found in step 1; these centroids are the new set of generators $x_1,\cdots ,x_n$ . 4: Terminate if an convergence requirement is met by this new set of points; if not, go back to step 1.

Moreover, the control rule is decentralized since the coverage function $J$ and its partial derivative are unique dependents on the position $x_i$ and the Voronoi polygon ${\mathbb{V}}_i$ . The above algorithm is employed to deploy the quasi-stationary sink nodes at the centroid of each cell. This guarantees optimal coverage of the receiving nodes to any point inside the cell where the coverage area is determined by the circle drawn in each cell in Figure 3. The radius of the circles are uniform for all cells, this guarantees uniform load distribution to all the sink nodes for coverage.

Opportunistic routing strategy

The collected localized weather predictions and parameters needs to be efficiently routed to the Ground Base Station for further processing and application level services. The present subsection discusses the proposed message transfer strategy that is obtained by leveraging opportunistic routing philosophy.

Three major categories of routing have been considered in the present deployment scenario, namely opportunistic flooding, probabilistic flooding, and opportunistic forwarding strategies. The opportunistic flooding is a classical flooding scheme in opportunistic networks. Majorly the epidemic routing fallen under this category. The operation represents the following operations:

• Initialize message $M$ in a buffer $B$ . Assign an ID to message which is $M<-I\left(x\right)$ .

• If the contact between two nodes happens then perform the exchange of $I\left(x\right)$ and message summary vector.

• Check the $I\left(x\right)$ for undelivered messages.

• Finally the undelivered messages has to be transferred for the delivery.

The probabilistic flooding on the other hand perform transitivity and the history based phenomena in order to transfer the message to the nodes. It is fundamentally a restricted flooding methodology. In this case the message get transfer based on the probability of the encounter of the node with one another which can be expressed as follows:

$$P\left(X,Y\right)=P{\left(X,Y\right)}_{old}+\left(1-P{\left(X,Y\right)}_{old}\right)\times P_{enc}$$

(11)

Here, $P_{\mathrm{enc}}$ is the encounter parameter. If $X$ meets more frequently with $i$ , there is a good chance to route the message via $X$ to $i$ , and from $i$ to $Y$ , toward the destination. Therefore, it can be written as.

$$P\left(X,i\right)=P{\left(X,i\right)}_{old}+(1-P{\left(X,i\right)}_{old}\times P\left(X,i\right)\times P\left(X,Y\right)$$

(12)

Forwarding technique is on the other hand restricts the number of message copy in the network. The opportunistic forwarding strategies often uses neighborhood discovery, data transfer and storage management policy. The delivery likelihood is a crucial parameter in this case. Let $n$ represent the number of nodes; each node $x\in n$ keeps track of the encounter probability of its peer $y\in n$ . now set

$$f_y^x\leftarrow \frac{1}{(\left|n\right|-1)}$$

(13)

where $f_y^x$ is the connection likelihood. If $x$ encounters $y$ , then

$$f_y^x\leftarrow f_y^x+1$$

(14)

The cost is computed as

$$c(i,i+1,\cdots d)=\sum _{x=i}^{d-1}(1-f_{x+1}^x)$$

(15)

and the destination cost is therefore

$$c\left(d\right)\leftarrow \mathrm{low}[c(i,i+1,\cdots d)]$$

(16)

The operation of localized weather prediction using quasi-stationary edge devices, with information routed to the base station via groups of relay UAVs in each Voronoi region, is limited by the energy constraints of the UAVs. This challenge can be addressed by utilizing a mobile charging station supported by an Unmanned Ground Vehicle (UGV), as suggested in Messaoudi et al. ³⁵

Predictive modeling

Each of the Vornoi cell has a edge enabled quasi-stationary UAV node for localized prediction and forecasting. While the edge-level predictive model provides real-time insights leveraging the constrained resources of flying nodes, a more comprehensive forecasting approach is achieved by integrating these edge predictions with models deployed in the cloud or main ground station.

Prediction task

Each of the localized areas has a quasi-stationary sink nodes with model processing capability. The machine learning prediction task involves predicting weather conditions based on multiple features such as temperature, rain, pressure, humidity, etc. Mathematically, this prediction task can be framed as a regression problem as well as a classification problem. The regression task is designed to estimate a continuous variable representing the weather condition.

Let $Y$ denote the predicted weather outcome, and $[f_1,...,f_n]$ represent the respective features. The prediction function $M\left(X\right)$ can be expressed as:

$$Y=M(f_1,...,f_n)$$

(17)

where:

–f(X) is the weather forecasting function based on the input features.

–n represents the number of features.

Hybrid forecasting model

The Hybrid Forecasting Model aims to combine edge-level forecasting with cloud-level forecasting using a weighted averaging technique. This model integration strategy involves executing predictive models on edge devices within a FANET and in a cloud or ground station environment. The weighted averaging approach combines the predictions from these models to generate a more accurate and comprehensive weather forecast.

Edge-level forecasting

For the edge-level forecasting, models such as Decision Trees or Linear Regression are employed due to their lightweight nature and suitability for resource-constrained devices. Let $\mathrm{Forecas}{t}_{\mathrm{Edge}}$ represent the forecast generated by the edge-level predictive model based on the features $X_1,X_2,X_3,X_4$ .

$$\mathrm{Forecas}{\mathrm{t}}_{\mathrm{Edge}}=g(X_1,X_2,X_3,X_4)$$

where $g\left(X\right)$ is the edge-level forecasting function.

Cloud-level forecasting

At the cloud or ground station level, a more sophisticated model such as Long Short-Term Memory (LSTM) is utilized for its ability to capture temporal dependencies in time series data. Let $\mathrm{Forecas}{\mathrm{t}}_{\mathrm{Cloud}}$ denote the forecast generated by the cloud-level LSTM model based on historical data and additional features.

$$\mathrm{Forecas}{\mathrm{t}}_{\mathrm{Cloud}}=h(\mathrm{Historical}\mathrm{Data},X_1,X_2,X_3,X_4)$$

where $h\left(X\right)$ represents the cloud-level forecasting function. The LSTM architecture is designed to capture temporal dependencies in time series data, making it suitable for predicting weather conditions based on historical data and additional features. Let $\mathrm{Forecas}{\mathrm{t}}_{\mathrm{Cloud}}$ denote the forecast generated by the cloud-level LSTM model. The LSTM model takes historical data sequences and the features $X_1,X_2,X_3,X_4$ as inputs to predict the weather outcome. The LSTM forecasting process involves sequential processing through memory cells and gates to capture long-range dependencies:

$$\begin{array}{c}\mathrm{Forecas}{\mathrm{t}}_{\mathrm{Cloud}}=\mathrm{LSTM}(\mathrm{Historical}\mathrm{Data},X_1,X_2,X_3,X_4)\hfill \\ \mathrm{Forecas}{\mathrm{t}}_{\mathrm{Cloud}}=h(\mathrm{Historical}\mathrm{Data},X_1,X_2,X_3,X_4)\hfill \end{array}$$

where— $\mathrm{Forecas}{\mathrm{t}}_{\mathrm{Cloud}}$ represents the forecasted weather outcome generated by the cloud-level LSTM model. — $\mathrm{LSTM}(·)$ denotes the LSTM forecasting process. — $h\left(X\right)$ represents the cloud-level forecasting function using LSTM, incorporating historical data sequences and the provided features.

LSTM components

The LSTM model processes sequential data and is programmed to learn temporal patterns to provide forecasts thus leveraging the temporal context provided by historical weather data and the input features. The LSTM based prediction process involves the following equations:

Input Gate:

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

$${\tilde{C}}_t=\tanh (W_C·[h_{t-1},x_t]+b_C)$$

Forget Gate:

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

$$C_t=f_t·C_{t-1}+i_t·{\tilde{C}}_t$$

Output Gate:

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

$$h_t=o_t·\tanh \left(C_t\right)$$

Where:

• – $i_t$ is the input gate’s activation vector at time step $t$ .

• – ${\tilde{C}}_t$ is the candidate cell state.

• – $f_t$ is the forget gate’s activation vector at time step $t$ .

• – $C_t$ is the cell state at time step $t$ .

• – $o_t$ is the output gate’s activation vector at time step $t$ .

• – $h_t$ is the hidden state or output at time step $t$ .

• – $h_{t-1}$ represents the previous hidden state.

• – $x_t$ is the input feature vector at time step $t$ . $W_i,W_f,W_o,W_C$ are weight matrices, and $b_i,b_f,b_o,b_C$ are bias vectors for respective gates.

• – $\sigma$ denotes the sigmoid activation function, and tanh is the hyperbolic tangent function.

LSTM algorithm

The LSTM process for weather forecasting can be summarized as follows:

The LSTM algorithm loops through the data, hence modifying the cell state and hidden state at each time step. This enables to capture temporal patterns and dependencies for accurate weather forecasting.

Algorithm 2.

LSTM Algorithm

Input: Historical weather data sequences- Feature vectors: precipitation, temp_max, temp_min, wind speed. 1: Initialize the LSTM model parameters (weights and biases). 2: Process historical weather data sequences and feature vectors sequentially. 3: Compute input gate activations $i_t$ and candidate cell states ${\tilde{C}}_t$ . 4: Compute forget gate activations $f_t$ and update cell states $C_t$ . 5: Compute output gate activations $o_t$ and hidden states $h_t$ . 6: Generate weather forecasts based on the final hidden states $h_t$ at each time step.

Weighted averaging fusion

The Hybrid Forecasting approach (equation (18)) combines the edge and cloud-level forecasts using a weighted averaging mechanism:

$$\mathrm{Forecas}{\mathrm{t}}_{\mathrm{Hybrid}}=\alpha ·\mathrm{Forecas}{\mathrm{t}}_{\mathrm{Edge}}+(1-\alpha )·\mathrm{Forecas}{\mathrm{t}}_{\mathrm{Cloud}}$$

(18)

where $\alpha$ is the weighting factor determining the contribution of the edge-level forecast to the final forecast. This weighted averaging approach dynamically integrates predictions from both edge and cloud-based models, optimizing for improved accuracy and reliability in weather forecasting.

Experimental results

Experimental data

The Delhi Weather Dataset (https://www.kaggle.com/datasets/mahirkukreja/delhi-weather-data/data) has been utilized for forecasting analysis. The data has been collected from wunderground utilizing their easy to use api. The dataset has $20$ features out of which following features has been utilized for analysis- temperature, pressure, humidity, rain, precipitation, dew-point etc. A class label defining weather condition- Partial Fog, Heavy Fog, Smoke etc. are defined for weather condition prediction. For current analysis data duration of 2009–2017 is considered. A histogram visualization of dataset is given Figure 4.

Figure 4.

Histogram of dataset distribution.

It is important to understand the effect of feature over the other. A correlation matrix is visualized in Figure 5. The analysis reveals several noteworthy correlations, with some features exhibiting strong positive or negative associations. It has been observed for the state of Delhi, India that temperature and humidity has strong negative correlation. This is so as the Warmer air can hold more moisture than colder air, so as the temperature in Delhi increases, the relative humidity hence may decrease. This is because the same amount of water vapor may occupy larger volumes in warmer air. A cluster analysis is given in Figure 6. Through the elbow method it has been found that the number of clusters is $3$ . It can be observed that the clusters are well formed with notable class boundary.

Figure 5.

Correlation between the features.

Figure 6.

KMeans clustering visualization of temperature versus humidity.

Cluster 0, reinforces the negative correlation. Cooler temperatures naturally hold less moisture, explaining the low humidity in this cluster. Cluster 1 deviates a little bit. Although it is pertinent that high temperatures can hold more moisture, specific factors like proximity to water bodies or strong winds contributes to the high humidity in this region hence the cluster. In Cluster 2, the temperature and humidity is the moderate range hence acting as the bridge showcasing a balanced relationship between temperature and humidity. It highlights the dynamic nature of this connection, where variations in considered features influences the balance.

Network performance analysis

To monitor the weather of a particular area the deployed UAV node will collect the data in a two-fold approach. In the first case, the data is collected from the high-altitude quasi-stationary node, and in the second case, the UAV nodes act as data collectors. In between them some of the UAV nodes are deployed as data mules for opportunistic message transfer. Two major parameters are considered in that case in order to achieve an effective message transfer between the sensing node and the destination nodes. The message delivery ratio $D_p$ is considered as a key performance metric which can be computed as the ratio between the number of messages delivered divided by the number of messages created. Therefore, one can write $D_p=M_c/M_d$ . In the case of the experimental scenario, the MQTT latency performance has been measured. In this case, the two major latency measures one can formulate. (i) latency from the sensing UAV node to the message server(broker) which is $l_{\mathrm{sb}}$ . (ii) From the broker to the base edge node (base or routing station) which is $l_{\mathrm{be}}$ . Here $l_{\mathrm{sb}}=(l_{\mathrm{tx}}+l_{\mathrm{cnack}}+l_{\mathrm{puback}}+l_{\mathrm{fwd}})$ and $l_{\mathrm{be}}=(l_{\mathrm{cnack}}+l_{\mathrm{pubrel}}+l_{\mathrm{suback}}+l_{\mathrm{fwd}})$ therefore the total latency cab achieved per node is $L=l_{\mathrm{sb}}+l_{\mathrm{be}}$ . The experimental simulation has been performed in Opportunistic Network Environment simulator which is a discrete event-based simulation environment.

The critical performance metric of delivery probability and the average latency has been measured. Figure 7 depicts the message delivery performance with varying node densities. In the proposed scenario, three major message transfer mechanisms namely Opportunistic flooding, Probabilistic flooding, and Opportunistic forwarding are considered. Figure 7 shows the message delivery performance of the opportunistic forwarding achieved the highest benchmark of message delivery in a maximum node density of $950$ . The increasing node density generally implies a better forwarding path for the packets. In the case of forwarding the message replica get controlled by the protocol itself resulting in better message delivery performance. Flooding mechanisms on the other hand give a maximum of $0.98$ in the case of probabilistic flooding which outperforms opportunistic flooding in node density of $950$ .

Figure 7.

Node density versus delivery probability achieved by the message transfer methods.

The effect of the latency is also significant in case of the increasing time. Figure 8 shows a decreasing trend of average latency with an increasing simulation time.

Figure 8.

Average latency performance of the message transfer methodology concerning the simulation time.

In Figure 8 the average latency of the opportunistic flooding is reported to be the lowest benchmark in the case of simulation time $90000$ this is due to the increasing number of replicated messages in the network causing a high degree of message overhead. If the ttl is increased even more, the latency will increase drastically. On the other opportunistic forwarding outperforms all because of the forwarding characteristics. In this case, an optimum number of message copies generated within the networks results in an average latency of $820$  ms with a simulation time of $90000$  s. Throughput performance of the different opportunistic routing shown in Figure 9 with varying quality of service level. From the figure it is evident that the $\mathrm{QoS}0$ offers a minimum throughput benchmark for all message transfer strategies where $\mathrm{QoS}2$ offers highest benchmark for relatively lesser number of message copy generation for $\mathrm{QoS}2$ . The opportunistic flooding in this case achieve minimum throughput performance due to larger number of message copy generation which leads to a congestion behavior in the whole network.

Figure 9.

Throughput performance of the different message transfer strategies under varying QoS levels.

The consumption of the power by the network node is also very crucial because most of the communication devices are battery operated. The node which consumes less power can consider as better forwarder within the whole network. Figure 10 illustrates power consumption by the node within the proposed simulation time. As the number of nodes increases with increased simulation time the cumulative power consumption goes to maximum benchmark of $290$  w and minimum of $100$  w. This is as the number of node increases the chance to get the forwarding path increase and the routing involves more number of nodes and the channel scan and forwarding operation also consumes power.

Figure 10.

Cumulative power consumption for varying number of nodes.

Figure 11 depicts average packet loss for the increasing number of nodes. The figure shows the highest number of packet loss of opportunistic flooding due to the huge amount of packet generation. On the other hand opportunistic forwarding restrict the message generation based on the selection of forwarder in an optimum way results a minimum benchmark of packet loss.

Figure 11.

Average packet loss based on number of node for flooding and forwarding techniques.

Prediction analysis

Weather condition prediction

At the edge level the trained model for weather condition prediction is deployed. Experimentation has been carried out on traditional machine learning models- Gaussian Naive Bayes, Decision Tree, K-nearest neighbor and Support Vector Machines. A weighted majority voting ensemble is deployed using the said classification models. A demonstration of accuracy metric is given in Figure 12. It is evident that the majority voting ensemble is giving the best performance as compared with all of the base learners. The details on parameter setting is given in Table 1.

Figure 12.

Weather condition prediction analysis.

Table 1.

Parameter settings for the selected classifier.

Classifier	Parameter setting
GaussianNB	var_smoothing = 1e-9, 1e-8
Decision tree	criterion = gini, max_depth = 10
kNN	neighbors = (5,7,10)
SVM	kernel = ‘poly’, degree = 4,5

Forecasting analysis

The forecasting analysis in Figures 13 and 14 showcases the effect of the proposed forecasting prediction analysis on temperature and humidity. It can be observed that the true values and the forecasted values are overlapping. This alignment underscores the effectiveness of our LSTM model in capturing the complex temporal dependencies inherent in the dataset. The congruence between actual and forecasted shows the robustness and reliability of the LSTM architecture in capturing nuanced patterns within time-series data.

Figure 13.

Proposed hybrid forecasting analysis of temperature.

Figure 14.

Proposed hybrid forecasting analysis on humidity.

In Figure 15, the test dataset has been separated from the training samples. The test sample comprised of dataset during the timeline 2015–2017. A comparative forecasting analysis is presented during the test timeline tenure using ARIMA model and the proposed LSTM model. It can be observed that the forecasted data points (marked in red) are well aligned with the actual data points (marked in green). Hence the forecasting analysis is performing better than ARIMA based model for the selected dataset.

Figure 15.

Comparison in forecasting (a) with ARIMA model (b) with LSTM model.

Conclusions

Gathering localized weather condition is an important criteria for strategic urban and rural planning. The proposed FANET framework is designed to handle areas where network coverage is very poor. The optimal segregation of area and deployment of FANET nodes creates the base coverage blueprint of the data collection framework. Localized weather parameters are recorded and predictions are made through the edge enabled base stations. The area wise localized weather parameters and predictions are routed to the main base station or cloud station. It can be observed through experimentation that the opportunistic flooding strategy utilized gives the maximum delivery probability and achieves minimum latency of $820$  ms. Extensive experimentation on the LSTM based Hybrid forecasting is presented and compared with the ARIMA model. The approach has been observed to achieve successful forecasts on futuristic predictions with RMSE: $5.79$ .

Future scope

In the proposed framework the message delivery, latency, power consumption and throughput has been considered as crucial network parameters. However some more network parameters can be considered for future analysis like energy efficiency, network lifetime, the communication overhead, weather condition, obstacle, network interference which can be considered as future work. Also, the heterogeneous UAV infrastructure using limited network bandwidth and processing capability is another major parameter to be considered. An enhance heterogeneous FANET structure also improve the scalability with varying number of UAV-UGV nodes. The network performance for the same and the deployment cost can also be considered as crucial parameter which are considered as future work. The major component of the proposed ecosystem can be developed using opensource software and the hardware platform that significantly reduce the deployment cost of the ecosystem itself.

Additionally integrating the proposed framework with 5G-enabled systems could also significantly enhance data transmission speed and reliability, allowing for more efficient, real-time weather data routing via drones, especially in remote areas. Employing advanced AI techniques, such as transformer models or attention mechanisms, could further refine forecast accuracy by capturing complex temporal patterns.