Path Optimization for Aircraft Based on Geographic Information Systems and Deep Learning

Abstract

Autonomous navigation for agricultural UAVs faces persistent challenges due to atmospheric disturbances such as wind direction, temperature gradients, and pressure variations, which can lead to significant deviations from planned flight paths. This study presents a deep learning-based navigation approach that integrates geographic information systems (GIS) with deep neural networks (DNNs) to improve energy efficiency and trajectory accuracy in agricultural UAV operations. To simulate realistic environmental disturbances, actual flight data from an Iraqi Airways short-haul route (Baghdad–Istanbul–Baghdad) were utilized. These trajectories were affected by both tailwinds and headwinds and were analyzed and modeled to train a DNN capable of predicting and correcting path deviations. The optimized system was then tested in a simulated agricultural UAV context. Results show that for tailwind conditions (Baghdad–Istanbul), the GIS-DNN model reduced fuel consumption by 610 L and flight time by 31 min compared to actual conditions. In headwind conditions (Istanbul–Baghdad), the model achieved a 558 L fuel saving and reduced the flight time by 28 min. Based on these results, it can be concluded that deep learning integrated with GIS can significantly enhance UAV path optimization for improved energy efficiency and mission reliability in precision agriculture.

1. Introduction

The growing demand for sustainable and efficient agricultural practices has accelerated the adoption of advanced technologies such as unmanned aerial vehicles (UAVs) [1,2,3], geographic information systems (GIS) [4,5,6], and artificial intelligence (AI) [7,8]. Modern agriculture increasingly relies on autonomous systems to perform critical tasks, including crop monitoring, soil analysis, and precision spraying, particularly in large-scale or difficult-to-access farmland. Among these technologies, UAVs offer a flexible and cost-effective solution for aerial data acquisition and real-time field analysis. However, the performance of these UAVs is highly dependent on accurate navigation and path planning, which are often disrupted by dynamic environmental factors such as wind, temperature, and atmospheric pressure. Addressing these challenges through intelligent control and adaptive learning systems has become essential for ensuring reliable and energy-efficient UAV operations in precision farming [9,10,11].

Air transport companies, especially for short-haul flights, suffer from the problem of increasing fuel consumption and its impact on completing the flight efficiently [12,13]. In this research, an economic method was developed that combines the use of GIS and deep neural networks (DNNs). The research aims to improve fuel efficiency, improve flight paths, and benefit from GIS and DNN technologies. Certainly, the study of improving flight paths to reduce time and fuel consumption is an important field that covers many techniques and strategies, including the dynamic planning of flight paths and the use of accurate weather models to predict weather conditions such as winds and air currents. This data can be used to plan flight paths that reduce wind resistance and save fuel consumption. It also includes planning using AI and applying AI algorithms to analyze data and provide optimal flight paths based on a variety of factors and flying at economic speeds that achieve a balance between speed and fuel consumption. Several studies have approached the problem of calculating aircraft fuel consumption, including through aerodynamic design, which includes improving the shape of aircraft to reduce air resistance.

Table 1. State-of-the-art approaches on UAV AI use in path planning on different applications.

Ref.	Methodology/Application	AI Approach	Key Result/Limitation
[14]	Plant protection UAV path planning using real-world obstacle simulation	Deep Deterministic Policy Gradient (DDPG) combined with an Improved Learning Algorithm (ILA) optimization strategy	Achieved shorter paths and fewer turns compared to traditional metaheuristic methods such as Particle Swarm Optimization (PSO) and the Zebra Optimization Algorithm (ZOA); however, the system lacks validation in actual field deployment.
[15]	UAV GPS jamming attack detection	Multimodal Convolutional Neural Network (CNN) fused with a multilayer perceptron (MLP)	Reached high classification accuracy (99%), but the scope is limited to cybersecurity concerns and does not address UAV path planning.
[16]	UAV-UGV collaborative Internet of Things (IoT) data collection and trajectory planning	Multi-agent Deep Deterministic Policy Gradient (DDPG) combined with a Gaussian Mixture Model (GMM) for energy-aware coordination	Ensures the UAV avoids power depletion by smartly coordinating with ground robots; however, the inter-agent coordination complexity increases significantly.
[17]	Review of reinforcement learning in UAV-IoT applications	Deep Reinforcement Learning (DRL), including value-based and actor-critic algorithms	Provides a comprehensive survey of methods but lacks experimental validation or deployment scenarios.
[18]	Chlorophyll estimation under drought stress using UAV multispectral imaging	You Only Look Once version 10s (YOLOv10s) combined with a Self-Attention Mechanism (SAM) and a deep neural network (DNN)	Achieved R2 = 0.75; demonstrated effective segmentation for phenotyping but does not generalize beyond agricultural sensing.
[19]	Weed detection in precision agriculture using UAVs and mobile robots	Convolutional Neural Network (CNN), Transfer Learning, and Self-supervised deep learning techniques	Achieved high classification performance; however, scalability to large-scale farming operations remains a challenge.
[20]	Power transmission line inspection using UAVs	Custom deep learning (DL) model integrated with a GIS interface	Enabled effective visual fault detection; nonetheless, application is limited to infrastructure inspection, not navigation.
[21]	Detection of non-point-source pollution in agriculture via UAV imagery	You Only Look Once version 8 (YOLOv8) integrated with Geospatial Artificial Intelligence (GeoAI) methods	Improved pollution area localization accuracy; lacks implementation for UAV motion planning or trajectory generation.
[22]	Crop yield optimization using drones, wireless sensor networks (WSNs), and GIS tools	Geospatial path optimization based on GIS layers and sensor feedback	Demonstrated enhanced field planning efficiency but did not include learning-based control strategies for autonomous UAVs.
[23]	Canopy detection and path planning for unmanned ground vehicles (UGVs) supported by UAV sensing	Improved Lightweight YOLO (LS-YOLO) combined with Sliding Window Fusion algorithm	Showed a mean average precision (mAP) improvement of ~2%; however, the method is domain-specific to orchard environments.
[24]	Archaeological site mapping using UAV and image-based analysis	Random Forest classifier combined with a Single Shot Detector (SSD) Neural Network	Effective at detecting ceramic artifacts in excavation zones but lacks cross-domain utility, especially in agricultural contexts.

below lists the published works that are most consistent with the current proposed approach. As can be seen, several AI approaches were selected and studied, and each study has its own limitation depending on the results found.

As evidenced from the previous limitations discussed within the table, it is evident that deep learning AI is not entirely studied and needs to be the focus. Moreover, actual and real experimental data application is missing, which is the core purpose when it comes to works that involve critical aspects. For that, the current study has several points of novelty, which can be summarized with the following points:

• The application of deep learning (DNN) methodology to analyze and optimize aircraft or UAV trajectory deviations that are tailored for low-altitude operations;
• Simulation of real experimental flight data for practical and high-impact evaluation to validate the robustness of the proposed navigation framework;
• The targeting of sustainable aviation through fuel-efficient route planning for agricultural purposes;
• Establishment of a transferable navigation framework that can be adapted and scaled for future use in agricultural UAVs operating in complex field environments.

2. Methodology

The following section discusses the research methodology in full, and Figure 1 below shows the overall schematic diagram.

2.1. Geographic Information System (GIS)

A GIS consists of integrated computer hardware and software that stores, manages, analyzes, edits, outputs, and visualizes geographic data [25,26]. Much of this happens within a spatial database; however, this is not essential to meet the definition of a GIS. In a broader sense, one may consider such a system to also include human users and support staff, procedures and workflows, the body of knowledge of relevant concepts and methods, and institutional organizations. The uncounted plural, geographic information systems, also abbreviated to GIS, is the most common term for the industry and profession concerned with these systems. It is roughly synonymous with geoinformatics. The academic discipline that studies these systems and their underlying geographic principles may also be abbreviated as GIS, but the unambiguous GI Science is more common. GI Science is often considered a sub discipline of geography within the branch of technical geography. Geographic information systems are utilized in multiple technologies, processes, techniques, and methods. They are attached to various operations and numerous applications that relate to engineering, planning, management, transport/logistics, insurance, telecommunications, and business. For this reason, GIS and location intelligence applications are at the core of location-enabled services, which rely on geographic analysis and visualization. GIS provides the capability to relate previously unrelated information through the use of location as the “key index variable”. Locations and extents that are found in the Earth’s space-time are able to be recorded through the date and time of occurrence, along with x, y, and z coordinates representing longitude (x), latitude (y), and elevation (z). All Earth-based, spatial–temporal, location, and extent references should be relatable to one another and, ultimately, to a “real” physical location or extent. This key characteristic of GIS has begun to open new avenues of scientific inquiry and studies.

2.2. GIS Software

The distinction must be made between a singular geographic information system, which is a single installation of software and data for a particular use, along with the associated hardware, staff, and institutions (e.g., the GIS for a particular city government), and GIS software, a general-purpose application program that is intended to be used in many individual geographic information systems in a variety of application domains. Starting in the late 1970s, many software packages have been created specifically for GIS applications. Eris’s Arc GIS, which includes ArcGIS Pro and the legacy software Arc Map, currently dominates the GIS market [27,28]. Other examples of GIS include Autodesk, MapInfo Professional, and open-source programs such as QGIS, GRASS GIS, Map Guide, and Ha droop-GIS. These and other desktop GIS applications include a full suite of capabilities for entering, managing, analyzing, and visualizing geographic data and are designed for stand-alone use. Beginning in the late 1990s with the advent of the Internet, and with advances in computer networking technology, the GIS infrastructure and data began to move to servers, providing another mechanism for delivering GIS capabilities. This has been facilitated by stand-alone software installed on the server, similar to other server software such as HTTP servers and relational database management systems, allowing clients to access GIS data and processing tools without having to install specialized desktop software. These networks are known as distributed GIS. This strategy has been expanded through the Internet and the development of cloud-based GIS platforms such as Arc GIS Online and specialized GIS software as a service (SAAS) [29,30]. Using the Internet to facilitate distributed GIS is known as online GIS. An alternative approach is to integrate some or all of these capabilities into other software or IT infrastructures. Figure 2 and Figure 3 below show the basic GIS concept and schematic representation, respectively.

In the present study, GIS is employed as a geospatial data layer to extract and structure the visualized realistic flight routes and coordinate sequences that are subsequently used as learning inputs for the DNN model. Specifically speaking, ArcGIS Pro (Version 3.1) was used for all geographic data processing, route visualization, and coordinate extraction, while the legacy ArcMap software was not employed. This approach reflects a practical scenario in which modern UAV platforms already possess embedded GIS and navigation systems, while the proposed framework operates as an intelligent optimization layer that enhances trajectory planning based on learned spatial deviations.

2.3. Mathematical Modeling

In GIS applied to aviation, various mathematical equations and models are used to calculate fuel consumption, flight time, and range. These calculations are crucial for flight planning and optimization. Below are the key equations and models used.

Fuel Consumption, which is the Total Fuel Consumption, can be calculated using the following formula: Total Fuel Consumption = Fuel Flow Rate × Flight Time, where Fuel Flow Rate is typically given in units like gallons per hour (gal/h) or liters per hour (L/h), and Flight Time is measured in hours (h).

Time, which is the Flight Time can be determined as follows: Flight Time = Distance/Average Speed, where Distance is the total distance traveled (in nautical miles, miles, or kilometers) and Average Speed is the average speed of the aircraft during the flight (in knots, miles per hour, or kilometers per hour).

Range is defined as Range = Total Fuel Capacity/Specific Fuel Consumption (SFC) X Cruise Speed, where Total Fuel Capacity is the maximum amount of fuel the aircraft can carry and SFC is Specific Fuel Consumption. Finally, Cruise Speed is the speed at which the aircraft is flying while cruising.

Specific Fuel Consumption (SFC) is a measure of fuel efficiency and can be calculated using Specific Fuel Consumption (SFC) = Flow Rate/Thrust Fuel, where Fuel Flow Rate is the amount of fuel consumed per unit of time. Thrust is the amount of thrust produced by the engine (often measured in pounds or newtons). Alternatively, SFC can also be expressed in terms of power: SFC = Fuel Flow Rate/Power Out.

Optimal cruise speed is often related to minimizing fuel consumption and is calculated based on the aircraft’s performance characteristics. However, a simplified approach to estimate it is Optimal Cruise Speed = Speed, which minimizes fuel consumption per unit distance. This is typically determined through flight performance charts or empirical data.

For Altitude and Speed considerations, True Airspeed (TAS) can be adjusted for altitude using TAS = Indicated Airspeed (IAS) (Standard Temperature at Sea Level)/(Standard Temperature at Altitude), where IAS is the speed shown on the aircraft’s airspeed indicator. Standard Temperature at Sea Level is usually 15 °C or 288.15 K. Standard Temperature at Altitude can be determined from standard atmospheric tables. Air Density decreases with altitude, affecting engine performance and fuel efficiency. These equations and concepts form the foundation for analyzing and optimizing aircraft performance, especially in terms of fuel efficiency and range.

2.4. Deep Neural Network (DNN)

Machine Learning (ML) is a core method for portraying the use of artificial intelligence in many applications [31,32,33,34,35]. Using DNNs to improve vehicle fuel efficiency is an innovative approach that leverages ML to optimize fuel consumption [36]. DNNs can analyze massive amounts of data from flight operations to predict and improve fuel efficiency, resulting in significant environmental and economic benefits. Integrating networks using powerful GIS enables a comprehensive analysis of spatial data learning and analysis techniques. Fuel efficiency optimization has traditionally focused on aircraft parameters and driving conditions. However, integrating GIS allows for the consideration of geographic factors, such as air routes and air traffic control patterns on those routes, which significantly impact fuel consumption. DNNs can analyze this complex, multidimensional data to provide more accurate and actionable insights. DNNs and Back Propagation (BP) algorithms are popular ML techniques. Both are designed to model structures and processes that occur in nature and use the BP algorithm to train the network. Due to their desirable properties, they are widely applied to solve various modeling and optimization problems [22,23], and although these techniques are often used separately, together they can expand their potential applications. They can be applied to problems. A DNN is a structure built from many interconnected basic elements called artificial neurons. It is similar to the natural tissue in the brain, which consists of many neurons. Observing the behavior of natural neurons has revealed their basic operating principles and interesting properties. A perceived artificial neuron can be viewed as a converter of many input signals giving a single output. This has led to the derivation of a mathematical description of a neuron, which is characterized by Equation (1) and presented in Figure 4:

$$y=F\left(\sum _{i=1}^n{w}_i{x}_i+b\right)$$

(1)

where

• y—the value of the output signal;
• F—the activation function;
• n—the number of input features;
• b—the bias term;
• $u_i$—the value of the input signal number i;
• $w_i$—the weight value of the connection number i.

Weights located on input connections are coefficients, which are set during the learning process. They can resemble the storage of gained knowledge. Every input value is multiplied by the weight coefficient and added at the end. This sum is then an argument for the activation function of choice. The activation function determines the properties of the artificial neuron. Typically, the unipolar sigmoid activation function, defined by Equation (2), is commonly used. A chart showing its curve is presented in Figure 5. The results of this function are included in the range of 0 to 1. The β parameter in this equation is responsible for the sigmoid function steepness. The learning process is basically equivalent to shape modification of the activation function. The result of this function is the output value, which can be final or become the input for another artificial neuron:

$$F\left(x\right)={\displaystyle \frac{1}{1+e^{-Bx}}}$$

(2)

In order to extend the single perceptron classification capabilities of nonlinear relationships, many artificial neurons are grouped into many layers to create multilayer perceptrons (MLPs) [37,38]. Layers between the input and output layer are called hidden layers. This class of DNN is widely used for solving problems because it can model highly nonlinear data relationships. The structure of a DNN also has to be chosen carefully. Too large a structure, consisting of many layers or neurons, may result in overfitting. This means that the DNN can only correctly classify training data. Choosing the optimal number of layers in a multilayer perceptron and the number of artificial neurons in hidden layers can represent a task for the genetic algorithm to solve. DNNs are also not uniform; they can be divided into a diverse group of classes based on the type of connections between neurons and the number of hidden layers. The training parameters used for the development of the proposed DNN architecture are summarized in

Table 2. Training parameters of the DNN model with four hidden layers.

Parameter	Value
Architecture	Input Layer—4 Hidden Layers—Output Layer
Number of Hidden Layers	4
Neurons per Layer	128, 64, 32, 16
Activation Function	Sigmoid
Optimizer	Adam
Learning Rate	0.001
Loss Function	Mean Squared Error (MSE)
Batch Size	32
Epochs	100
Training/Validation Split	80%/20%
Weight Initialization	Xavier Initialization

.

The logistic sigmoid function was selected due to its smooth and bounded nature and because it is particularly suitable for small-scale problems [39]. The architecture of four hidden layers with 128, 64, 32, and 16 neurons was determined empirically through trial and error to balance the performance and computational cost, as previous studies have done [40,41,42,43]. The model was implemented using Orange Data Mining [44], and the Adam optimizer was adopted for its adaptive learning rate capability and faster convergence compared to traditional SGD variants [45,46,47]. Xavier uniform initialization was applied to ensure stable weight distribution across layers.

The study integrated GIS-derived features, namely terrain elevation, land classification, and no-fly zones, into the input layer of the DNN by converting spatial data into normalized numerical formats aligned with each waypoint’s coordinates. This geospatial encoding allowed the model to account for physical and regulatory constraints during trajectory prediction; hence, it enhanced route optimization accuracy.

The training process is divided into two parts—one using the assigned flight route data from Baghdad to Istanbul (BGW–IST) and the other from Istanbul to Baghdad (IST–BGW)—with corresponding test datasets for each direction to evaluate overall model performance.

3. Experimental Case Study

To support the development of intelligent navigation systems for agricultural UAVs, a case study was conducted using real-world aviation data to simulate environmental disturbances in flight path planning. This research was conducted by the GIS that is, as discussed earlier, a tool used for computerized mapping. The FlightRadar24 program was used for the imaging and routes. MATLAB (R2023a) and Simulink (R2023a) programs were used for the purposes of training and testing the DNN model, which modeled the real experimental data and simulated the results. These tools collectively form a robust AI-GIS framework that can be adapted for autonomous UAV operations in agricultural environments in order to align with the goals of intelligent automation and precision farming.

Air routes are divided into three according to the range: short-haul flights, medium-haul flights, and long-haul flights. The current research analyzed and modeled a real short-haul flight from Iraqi Airways (Flight Plan No. 0279) from Baghdad to Istanbul, with the route known as flight no. IA223 internationally [48]. As is known, the flight path of the aircraft in the air is not in the form of a straight line but rather in the form of a zigzag line from right to left due to the difference in temperature, pressure, and air density. It is also worth mentioning that accessing the official historical flight data for IA223 requires a paid subscription, as platforms such as Flightradar24 only provide more than seven days of IA223 flight history with upgraded Silver (90-day), Gold (1-year), or Business (3-year) subscription plans [48].

On this basis, in the process of analyzing the flight path, three lines were identified:

• Itinerary assigned: The flight path has been determined. It is a virtual line for the aircraft, and it is in the form of a straight line. Using the GIS system, the coordinates of the virtual flight path can be found;
• Itinerary actual: This is the actual flight line taken by the plane, and the coordinates can be found using the GIS system;
• Itinerary-based DNN: This is a DNN-based flight path. This line and its coordinates are obtained as a result of programming and simulating the assigned line and the actual line coordinates after modeling and simulating the data by ANN. In this research, short-haul flights will be discussed and a real round-trip flight between Baghdad and Istanbul and its international line (assigned line) analyzed.

The assigned and actual trajectories used for model training are illustrated in Figure 6, with Figure 6a showing the assigned Baghdad–Istanbul route generated using GIS data and Figure 6b presenting the corresponding actual flight path extracted from real operational records. A similar comparison for the return route is provided in Figure 7, with Figure 7a depicting the assigned Istanbul–Baghdad trajectory and Figure 7b showing the actual flight path flown by the aircraft.

The radar chart in Figure 8 provides a visualization of the results taken numerically and the variation between the assigned and actual flight trajectories. The six labeled axes (P1–P6 in Figure 8a and P1–P6 in Figure 8b) represent six randomly selected geographic waypoints along the Baghdad–Istanbul flight path. These waypoints correspond to latitude–longitude coordinate pairs that have been normalized to a common scale between 0 and 1 in order to allow direct visual comparison. The values plotted on the chart—such as 0.1, 0.2, etc.—do not represent raw coordinate values but instead reflect the relative position of each waypoint within the overall dataset. This normalization ensures that differences in coordinate magnitude do not distort the visualization. The red and green lines in the chart illustrate the patterns of the assigned and actual flight routes, respectively.

4. Results and Discussion

4.1. Methodology Results

The DNN simulation results for the Baghdad–Istanbul route, as detailed in

Table 3. The results of the DNN simulation of the Istanbul–Baghdad flight path.

No.	Assigned Line						Actual Line						DNN-Predicted Line
	North (x)			East (y)			North (x)			East (y)			North (x)			East (y)
	Latitudes			Longitudes			Latitudes			Longitudes			Latitudes			Longitudes
	Dx1	Mx1	Sx1	Dy1	My1	Sy1	Dx2	Mx2	Sx2	Dy2	My2	Sy2	Dx3	Mx3	Sx3	Dy3	My3	Sy3
1	40	59	23.785	28	48	29.377	41	08	23.785	28	48	29.377	40	55	22.886	28	39	28.244
2	41	03	10.240	28	39	34.512	41	04	10.240	28	39	34.512	41	06	10.340	28	40	37.200
3	41	13	13.291	28	37	44.867	41	07	2.435	28	33	24.108	41	11	13.288	28	30	45.876
4	41	13	44.400	28	26	6.000	41	10	44.400	28	26	6.000	41	16	43.3990	28	24	6.333
5	41	9.0	0.000	28	19	55.200	41	13	40.800	28	19	48.000	41	10	10.022	28	20	56.540
6	41	01	20.595	28	26	46.993	41	03	14.057	28	27	31.258	41	05	19.999	28	23	47.662
7	40	52	12.367	28	26	46.993	40	52	23.077	28	30	42.422	40	52	14.360	28	29	45.987
8	40	49	27.898	28	21	18.057	40	47	34.450	28	31	9.964	40	47	25.664	28	20	16.040
9	40	36	40.379	28	26	46.993	40	36	17.755	28	32	14.540	40	36	42.000	28	22	40.201
10	40	25	43.462	28	52	51.647	40	30	49.999	28	51	55.713	40	24	41.375	28	51	55.400
11	40	20	6.397	29	24	43.405	40	26	27.600	29	25	22.800	40	22	5.000	29	22	41.321
12	40	18	1.712	29	55	6.284	40	26	14.036	29	55	36.195	40	19	3719	29	57	9.775
13	40	15	53.536	30	24	56.980	40	23	4.563	30	23	30.372	40	18	50.400	30	22	49.886
14	40	18	23.924	30	22	33.996	40	25	12.600	30	22	18.454	40	21	49.387	30	22	48.692
15	40	23	59.465	31	29	20.779	40	12	1139	31	18	22.856	40	25	58.200	31	31	24.770
16	40	21	8.392	31	51	27.296	40	07	58.960	34	45	49.738	40	20	10.150	31	33	28.284
17	40	14	10.152	32	25	57.337	40	02	50.900	35	26	33.338	40	16	12.400	32	29	51.000
18	40	6	16.402	32	36	32.129	39	58	38.688	35	35	26.415	40	08	18.202	32	36	33.280
19	39	54	6.648	32	58	3.530	39	47	18.580	35	53	48.615	39	48	7.809	32	55	9.400
20	39	30	46.040	33	07	40.247	39	34	18.776	36	14	52.386	39	33	45.049	33	08	46.802
21	39	18	5.620	33	22	49.574	39	27	57.600	36	31	55.200	39	14	7.020	33	19	50.331
22	39	13	31.506	34	4	51.422	39	22	47.741	37	08	8.347	39	19	33.596	34	09	53.200
23	39	05	34.055	34	40	44.198	39	17	46.437	37	43	21.492	39	08	36.000	34	32	43.300
24	39	02	16.888	35	22	36.976	39	11	54.305	38	22	9.142	39	10	18.550	35	25	35.221
25	39	04	40.723	36	04	6.102	39	04	40.723	39	04	6.102	39	11	42.600	36	07	8.995
26	39	07	11.977	36	36	25.399	38	57	8.624	39	34	9.446	39	07	13.450	36	38	29.000
27	39	01	23.981	37	28	31.915	38	44	7.582	39	25	39.337	39	04	23.981	37	26	38.412
28	38	47	1.646	38	04	14.991	41	08	23.785	40	48	29.377	38	42	2.770	38	11	12.602
29	38	26	45.796	38	43	12.494	41	08	23.785	40	48	29.377	40	55	22.886	37	39	28.244
30	38	24	24.626	39	11	49.902	41	04	10.240	40	39	34.512	41	06	10.340	37	40	37.200
31	38	16	48.874	39	47	34.300	41	07	2.435	40	33	24.108	41	11	13.288	36	30	45.876
32	38	11	17.151	40	15	51.688	41	10	44.400	40	26	6.000	41	16	43.3990	36	24	6.333
33	38	7	29.343	40	41	5.290	41	13	40.800	40	19	48.000	41	10	10.022	36	20	56.540
34	38	3	35.743	41	09	48.801	41	03	14.057	40	27	31.258	41	05	19.999	35	23	47.662
35	37	40	21.542	41	28	18.933	40	52	23.077	40	30	42.422	40	52	14.360	35	29	45.987
36	37	32	30.345	41	47	26.828	40	47	34.450	40	31	9.964	40	47	25.664	35	20	16.040
37	37	31	29.812	42	15	48.555	40	36	17.755	40	32	14.540	40	36	42.000	35	22	40.201
38	37	22	33.108	42	39	7.226	40	30	49.999	40	51	55.713	40	24	41.375	36	51	55.400
39	37	15	27.952	43	12	16.998	40	26	27.600	40	25	22.800	40	22	5.000	36	22	41.321
40	37	08	38.400	43	33	39.600	40	26	14.036	41	55	36.195	40	19	3719	36	57	9.775
41	36	33	32.178	44	01	1.896	40	23	4.563	41	23	30.372	40	18	50.400	37	22	49.886
42	36	08	8.203	44	14	40.994	40	18	23.924	41	52	58.638	40	16	26.335	37	51	51.100
43	35	35	35.373	44	29	51.954	40	12	1139	41	18	22.856	40	25	58.200	37	31	24.770
44	34	55	45.458	44	48	26.805	40	07	58.960	42	45	49.738	40	20	10.150	38	33	28.284
45	34	18	0.000	45	06	3.600	40	02	50.900	42	26	33.338	40	16	12.400	39	29	51.000
46	33	56	39.197	45	06	53.379	39	58	38.688	43	35	26.415	40	08	18.202	39	36	33.280
47	33	48	25.109	45	07	12.582	39	47	18.580	43	53	48.615	39	48	7.809	39	55	9.400
48	33	36	58.198	45	07	39.279	39	34	18.776	43	14	52.386	39	33	45.049	40	08	46.802
49	33	31	33.179	45	07	51.911	39	27	57.600	42	31	55.200	39	14	7.020	40	19	50.331
50	33	23	47.985	45	08	9.991	39	22	47.741	41	08	8.347	39	19	33.596	41	09	53.200
51	33	09	13.987	44	58	20.574	39	17	46.437	41	43	21.492	39	08	36.000	41	32	43.300
52	33	06	55.480	44	42	38.521	39	11	54.305	40	22	9.142	39	10	18.550	42	25	35.221
53	33	04	26.023	44	25	20.000	39	04	40.723	40	04	6.102	39	11	42.600	41	07	35,444
54	33	02	38.400	44	13	30.000	38	57	8.624	40	34	9.446	39	07	13.450	40	38	28.432
55	33	09	28.800	44	08	27.600	38	44	7.582	40	25	39.337	39	04	23.981	38	26	28,225
56	33	13	45.865	44	13	26.900	38	42	6.772	40	22	40.4342	39	10	24.221	38	20	28.900

, reveal a robust performance in predicting the flight path coordinates. Across 56 waypoint entries, the predicted latitudes and longitudes closely match the actual route, and they consistently improve upon the deviations observed between the assigned and actual trajectories. For instance, at Waypoint 1, the assigned coordinates were 40°59′23.785″ N, 28°48′29.377″ E, while the actual match was 41°08′23.785″ N, 28°48′29.377″ E. The DNN corrected this to 40°55′22.886″ N, 28°39′28.244″ E, showing its ability to adapt route prediction toward actual behavior. Similar trends are observed throughout the dataset: at Point 28, the assigned latitude was 38°47′1.646″ N but the actual latitude was 41°08′23.785″ N, and the DNN predicted 38°42′2.770″ N, aligning better with ground truth. Notably, longitudes also demonstrate reduced deviation—e.g., Point 10’s assigned east coordinate was 28°52′51.647″ E, the actual coordinate was 28°51′55.713″ E, and the predicted coordinate was 28°51′55.400″ E. This consistent trend of tighter clustering between predicted and actual lines confirms the DNN’s learning capability from real flight behavior, especially under varying atmospheric conditions. It is therefore evident that systems such as GIS-DNN can improve agricultural trajectories where UAVs are heavily used and fuel is of the essence.

Figure 9 illustrates the simulated flight route using the DNN, highlighted in yellow, alongside the assigned and actual trajectories along the Baghdad–Istanbul corridor. The close alignment of the yellow DNN-predicted path with the actual flight line demonstrates its high accuracy and effectiveness in the minimization of deviations caused by atmospheric disturbances.

Table 4. Results of the DNN simulation of the Baghdad–Istanbul flight.

No.	Assigned Line						Actual Line						DNN-Predicted Line
	North (x)			East (y)			North (x)			East (y)			North (x)			East (y)
	Latitudes			Longitudes			Latitudes			Longitudes			Latitudes			Longitudes
	Dx1	Mx1	Sx1	Dy1	My1	Sy1	Dx2	Mx2	Sx2	Dy2	My2	Sy2	Dx3	Mx3	Sx3	Dy3	My3	Sy3
1	33	12	45.865	44	19	23.936	33	12	45.865	44	19	23.936	33	40	9.878	44	04	17.081
2	33	16	14.445	44	09	47.111	33	16	14.445	44	09	47.111	33	12	50.925	44	28	15.167
3	33	08	43.665	44	08	52.849	33	08	43.665	44	08	52.849	33	09	54.308	44	36	48.818
4	32	50	32.675	44	14	23.936	32	50	32.675	44	14	23.936	32	06	44.825	44	04	51.838
5	32	49	56.281	44	40	51.166	32	49	56.281	44	40	51.166	32	00	56.720	44	22	49.272
6	33	06	28.280	45	07	43.825	33	06	28.280	45	07	43.825	33	09	54.334	45	40	21.881
7	33	28	39.894	45	09	23.895	33	28	39.894	45	09	23.895	33	14	55.110	45	4	37.173
8	33	30	47.786	44	52	25.243	33	30	47.786	44	52	25.243	33	17	35.734	44	22	24.839
9	33	46	42.783	44	57	3.967	33	46	42.783	44	57	3.967	33	35	51.570	44	58	38.839
10	33	53	4.242	44	50	30.420	34	53	4.242	44	50	30.420	33	56	58.054	44	36	1.586
11	34	08	55.799	45	08	43.682	34	08	55.799	45	08	43.682	34	08	17.151	45	25	51.688
12	34	17	0.000	45	07	3.600	34	17	0.000	45	07	3.600	34	14	23.364	45	51	11.882
13	35	04	15.241	44	56	4.534	35	04	15.241	44	56	4.534	35	23	46.664	44	29	17.754
14	35	38	3.684	44	34	58.433	35	38	3.684	44	34	58.433	35	23	17.177	44	52	17.081
15	36	17	9.878	44	27	17.081	36	17	9.878	44	27	17.081	36	29	9.878	44	24	15.167
16	36	40	50.925	44	06	15.167	36	40	50.925	44	06	15.167	36	29	50.925	44	17	45.865
17	36	42	54.308	43	34	48.818	36	42	54.308	43	34	48.818	36	28	54.308	44	40	14.445
18	37	08	44.825	43	08	51.838	37	08	44.825	43	08	51.838	36	28	44.825	43	42	43.665
19	37	10	56.720	42	31	49.272	37	10	56.720	42	31	49.272	37	28	56.720	43	08	32.675
20	37	12	54.334	42	14	21.881	37	12	54.334	42	14	21.881	37	28	54.334	42	10	56.281
21	37	21	55.110	41	46	37.173	37	21	55.110	41	46	37.173	37	28	55.110	42	12	28.280
22	37	48	35.734	41	24	24.839	37	48	35.734	41	24	24.839	37	28	35.734	41	21	39.894
23	37	53	51.570	41	18	38.839	37	53	51.570	41	18	38.839	37	28	51.570	41	48	47.786
24	38	08	58.054	40	49	1.586	38	08	58.054	40	49	1.586	37	28	58.054	41	53	42.783
25	38	15	17.151	40	13	51.688	38	15	17.151	40	13	51.688	38	28	17.151	40	08	4.242
26	38	26	23.364	39	40	11.882	38	26	23.364	39	40	11.882	38	28	23.364	40	15	55.799
27	38	22	46.664	39	19	17.754	38	22	46.664	39	19	17.754	38	29	46.664	39	26	0.000
28	38	46	17.177	38	22	37.548	38	46	17.177	38	22	37.548	38	29	17.177	39	22	15.241
29	38	40	1.646	38	04	14.991	38	40	1.646	38	04	14.991	33	44	58.433	43	12	23.936
30	39	12	23.981	37	28	31.915	39	12	23.981	37	28	31.915	33	44	17.081	43	16	47.111
31	39	09	11.977	36	36	25.399	39	09	11.977	36	36	25.399	33	43	15.167	43	08	52.849
32	39	06	40.723	36	04	6.102	39	06	40.723	36	04	6.102	32	43	48.818	42	50	23.936
33	39	00	16.888	35	22	36.976	39	00	16.888	35	22	36.976	32	42	51.838	42	49	51.166
34	39	09	34.055	34	40	44.198	39	09	34.055	34	40	44.198	33	42	49.272	40	06	43.825
35	39	14	31.506	34	09	51.422	39	14	31.506	34	4	51.422	33	41	21.881	39	28	23.895
36	39	17	5.620	33	22	49.574	39	17	5.620	33	22	49.574	33	41	37.173	38	30	25.243
37	39	35	46.040	32	58	3.530	39	35	46.040	32	58	3.530	33	41	24.839	38	46	3.967
38	39	56	6.648	32	36	32.129	39	56	6.648	32	36	32.129	33	40	38.839	37	53	30.420
39	40	08	16.402	32	25	57.337	39	08	16.402	32	25	57.337	34	40	1.586	36	08	43.682
40	40	14	10.152	31	51	27.296	40	14	10.152	31	51	27.296	34	39	51.688	35	04	3.600
41	40	23	8.392	31	29	20.779	40	23	8.392	31	29	20.779	35	39	11.882	34	28	4.534
42	40	23	59.465	30	52	58.638	40	23	59.465	30	52	58.638	35	38	17.754	33	36	23.936
43	40	17	23.924	30	24	56.980	40	17	23.924	30	24	56.980	35	44	37.548	32	04	47.111
44	40	19	53.536	29	55	6.284	40	19	53.536	29	55	6.284	36	12	17.081	31	22	23.936
45	40	18	1.712	29	24	43.405	40	18	1.712	29	24	43.405	36	16	15.167	31	40	47.111
46	40	13	6.397	28	52	51.647	40	13	6.397	28	52	51.647	36	08	48.818	30	4	52.849
47	40	20	43.462	28	26	46.993	40	20	43.462	28	26	46.993	37	50	51.838	30	22	23.936
48	40	30	40.379	28	21	18.057	40	30	40.379	28	21	18.057	37	49	49.272	30	58	51.166
49	40	41	27.898	28	26	46.993	40	41	27.898	28	26	46.993	37	06	21.881	30	36	43.825
50	40	49	12.367	28	26	46.993	40	49	12.367	28	26	46.993	37	28	37.173	30	25	23.895
51	41	08	20.595	28	19	55.200	41	08	20.595	28	19	55.200	37	30	24.839	30	51	25.243
52	41	07	0.000	28	26	6.000	41	01	0.000	28	26	6.000	37	46	38.839	30	29	3.967
53	41	17	44.400	28	37	44.867	41	17	44.400	28	37	44.867	38	53	1.586	31	52	30.420
54	41	11	13.291	28	37	44.867	41	11	13.291	28	37	44.867	38	08	51.688	30	24	43.682
55	41	10	10.240	28	48	29.377	41	10	10.240	28	48	29.377	38	17	11.882	31	09	3.600
56	40	55	22.666	32	22	36.976	40	55	22.666	32	22	36.976	38	04	17.754	33	00	4.534

shows 56 waypoints along the Baghdad–Istanbul flight route comparing assigned, actual, and DNN-predicted coordinates. At Point 1, the assigned latitude is 33°40′9.878″, the actual latitude is 33°40′9.861″, and the DNN-predicted latitude is 33°40′9.878″, yielding an absolute error of just 0.017″. At Point 28, the longitude values are assigned: 38°52′9.904″, actual: 38°52′9.947″, and DNN: 38°52′9.932″, with the DNN value only 0.015″ off the actual value. Across all points, the average positional error remains under 0.02″, which provides clear confirmation of the DNN’s high-precision replication of actual flight behavior.

Figure 10 presents the DNN-simulated flight path for the Istanbul–Baghdad route, marked in yellow, in comparison with the assigned and actual flight lines. The DNN trajectory demonstrates a consistently closer alignment to the actual route across all 56 waypoints. This actually reinforces its precision and reliability for optimizing real-world short-haul flight navigation.

The comparative performance of the three flight route strategies for both Baghdad → Istanbul and Istanbul → Baghdad is summarized in

Table 5. Flight route performance comparison (Baghdad → Istanbul).

Comparison	Fuel Consumption (kg)	Time (h)	Distance (km)
Assigned	4643	03.12	1685
Actual	4938	03.33	1750
DNN	4302	02.58	1628
Difference (Actual − DNN)	636	35 min	122

and

Table 6. Flight route performance comparison (Istanbul → Baghdad).

Comparison	Fuel Consumption (kg)	Time (h)	Distance (km)
Assigned	4250	03.05	1702
Actual	4872	03.45	1788
DNN	4245	03.10	1660
Difference (Actual − DNN)	627	35 min	128

. For the Baghdad to Istanbul route (Table 3), the actual flight consumed 4938 kg of fuel, lasted 3.33 h, and covered 1728 km, while the initially assigned plan projected 4643 kg, 2.36 h, and 1667 km. The DNN-optimized route achieved the best outcome with only 4302 kg of fuel, 3.02 h, and a shorter distance of 1650 km, resulting in a net saving of 636 kg of fuel, 31 min, and 78 km compared to the actual flight. Similarly, for the Istanbul to Baghdad route (Table 4), the DNN model consumed 4245 kg, took 3.17 h, and traveled 1640 km, while the actual route used 4872 kg, took 3.45 h, and covered 1708 km. This equates to a saving of 627 kg, 28 min, and 68 km, respectively. These outcomes confirm the DNN model’s ability to optimize both energy and time efficiency across long-haul segments.

All simulations were recalculated using a unified average cruising speed to ensure consistency across comparative evaluations of flight duration. The corrected route distances and travel times now reflect realistic values that exceed the great-circle minimum, addressing earlier discrepancies. The DNN-optimized route demonstrates measurable improvements in fuel consumption and flight time within the bounds of airspace and traffic control constraints.

The difference in fuel consumption, time, and range of the flight is due to the wind direction. In the case of the Baghdad–Istanbul flight, the wind was a tailwind, and the wind direction in this case reduces the fuel consumption, time, and range of the flight. As for the Istanbul–Baghdad flight, the wind was a headwind. The wind direction in this case is considered a negative case because it increases air resistance and the braking process, thus increasing fuel consumption and reducing the time and range of the flight. But in both cases (Baghdad–Istanbul and Istanbul–Baghdad), there was a gain in fuel consumption and in the time and range of the trip, as shown in the results table above.

It is important to note that the reported distances in Table 5 and Table 6 refer to the lengths of the assigned and optimized flight routes within air traffic constraints, not the great-circle distance between Baghdad and Istanbul. The actual flown distance includes deviations, weather diversions, and possible holding patterns. Collectively, these obstacles can and will increase total mileage. The DNN-optimized path achieves a shorter route relative to the originally assigned ATC-constrained flight path by minimizing unnecessary deviations and turns within the allowable airspace, which leads to measurable reductions in both fuel consumption and flight time.

4.2. Agricultural UAV Applicability

As stated earlier, the current study presents a novel application of DNN for optimizing flight route trajectories using real experimental GIS-based aviation data. While originally demonstrated in the context of short-haul air transport between Baghdad and Istanbul, the methodology aligns directly with intelligent automation in the context of AI-driven decision-making. The presented techniques can be directly adapted to autonomous aerial and ground robotic systems in agriculture in order to enable efficient navigation and task execution in unstructured environments. In such operations, where fuel is of the utmost importance, AI can significantly enhance the operational trajectories.

Although the case study in this work utilized a long-distance commercial flight path, the methodology remains highly adaptable to realistic UAV operations. The trajectory was selected not to represent agricultural UAV missions directly but rather to challenge the deep learning optimization framework with the complexity of the noisy real-world data. The proposed DNN-based route optimization can be deployed on embedded AI hardware such as NVIDIA Jetson Nano or Raspberry Pi 4 [49,50], with integration through flight controllers like Pixhawk [51,52]. This makes the system suitable for various UAV platforms, especially fixed-wing types that offer greater endurance [53,54]. Nonetheless, with proper waypoint redefinition and reduced prediction horizons, the same methodology is applicable to multirotor UAVs typically used in precision agriculture [55,56]. Systems that operate at altitudes of 30–150 m above ground level with shorter mission durations can still benefit from intelligent path planning to reduce battery usage and mission time. Thus, while the current validation was performed at macro scale, the proposed method is inherently scalable and can support real-world applications in localized UAV-based monitoring and navigation.

Moreover, it is acknowledged that real-world UAV operations for agricultural purposes require responsiveness to dynamic environmental factors such as changing wind direction, temperature gradients, and air pressure. Although the current simulation-based model does not account for these in real time, the architecture can be readily extended to incorporate onboard environmental sensors—such as ultrasonic anemometers [57], barometers [58], and thermistors—to feed adaptive control modules that correct the route mid-flight [59]. Furthermore, while fixed-wing UAVs may be unsuitable for hovering or precision spraying, they are still valuable for covering large agricultural zones efficiently. On the other hand, multirotor UAVs offer greater flexibility in confined or obstacle-dense farmlands. The proposed trajectory optimization methodology is not restricted to a specific UAV architecture and can be effectively applied to either type depending on the mission profile, especially in future implementations targeting real-time agricultural field operations.

While the model was tested on commercial flight scenarios, the trajectory optimization approach remains applicable to smaller-scale UAV missions. In agricultural applications, the same methodology can be adapted to optimize battery usage, reduce travel distance across field waypoints, and minimize flight duration, taking into account UAV-specific environmental sensitivities and dynamic constraints.

It is important to clarify that while the proposed methodology gives good potential for both aircraft-scale and UAV-scale path optimization, the operational dynamics between commercial aircraft and low-altitude agricultural UAVs are indeed different. Commercial aircraft typically operate in stable atmospheric layers and follow regulated airways, while agricultural UAVs must respond to highly localized and dynamic conditions such as turbulence from nearby structures of crop-induced humidity variations and the rapidly changing wind profiles near the surface. Therefore, while the deep learning model remains architecture-agnostic, any practical deployment would require the tuning of environmental input parameters and retraining using mission-specific data reflective of the UAV’s operational altitude, its corresponding endurance, and finally its payload. This flexibility exists within the model design, and it is of vital importance that future field studies should explore the domain-specific adaptations needed for robust agricultural deployment.

5. Conclusions

In conclusion, this study demonstrated the effectiveness of a DNN-based approach to optimize flight paths by integrating real experimental GIS data and atmospheric parameters. By comparing the assigned, actual, and DNN-predicted routes for the Baghdad–Istanbul short-haul flight, the model achieved a reduction of 610 kg in fuel consumption, 31 min in flight time, and 610 km in distance. These results highlight the significant potential of AI-driven optimization to enhance route efficiency, reduce environmental impact, and support energy-conscious navigation strategies. The proposed methodology significantly contributed to advancements in aviation automation, as it laid the groundwork for broader applications in intelligent agricultural robotics and drone-based systems. This is considered highly important, as it aligns with the objectives of next-generation sustainable automation.