Improved Deep Learning-Based Vehicle Detection for Urban Applications Using Remote Sensing Imagery

Abstract

Remote sensing (RS) data can be attained from different sources, such as drones, satellites, aerial platforms, or street-level cameras. Each source has its own characteristics, including the spectral bands, spatial resolution, and temporal coverage, which may affect the performance of the vehicle detection algorithm. Vehicle detection for urban applications using remote sensing imagery (RSI) is a difficult but significant task with many real-time applications. Due to its potential in different sectors, including traffic management, urban planning, environmental monitoring, and defense, the detection of vehicles from RS data, such as aerial or satellite imagery, has received greater emphasis. Machine learning (ML), especially deep learning (DL), has proven to be effective in vehicle detection tasks. A convolutional neural network (CNN) is widely utilized to detect vehicles and automatically learn features from the input images. This study develops the Improved Deep Learning-Based Vehicle Detection for Urban Applications using Remote Sensing Imagery (IDLVD-UARSI) technique. The major aim of the IDLVD-UARSI method emphasizes the recognition and classification of vehicle targets on RSI using a hyperparameter-tuned DL model. To achieve this, the IDLVD-UARSI algorithm utilizes an improved RefineDet model for the vehicle detection and classification process. Once the vehicles are detected, the classification process takes place using the convolutional autoencoder (CAE) model. Finally, a Quantum-Based Dwarf Mongoose Optimization (QDMO) algorithm is applied to ensure an optimal hyperparameter tuning process, demonstrating the novelty of the work. The simulation results of the IDLVD-UARSI technique are obtained on a benchmark vehicle database. The simulation values indicate that the IDLVD-UARSI technique outperforms the other recent DL models, with maximum accuracy of 97.89% and 98.69% on the VEDAI and ISPRS Potsdam databases, respectively.

1. Introduction

Driven by the challenges of rapid urbanization, urban cities are determined to apply modern sociotechnical transformations and change into smart cities. However, the success of these transformations depends heavily on detailed knowledge of spatial–temporal dynamics [1]. There has been much research discussion in the of science and engineering communities about the application of machine learning (ML) and artificial intelligence (AI) towards big data analytics and digitization, owing to the expansion of computer vision (CV) techniques [2]. Computer technology, used to perform tasks such as segmentation, detection, and the classification of remote sensing imagery (RSI), has become one of the major research topics. Among them, vehicle detection in RSI plays a major role in defense, urban vehicle supervision, safety-assisted driving, traffic planning, and so on. Fast and accurate detection approaches are essential for such tasks [3]. Compared with other detection targets in RSI (aeroplanes, buildings, and ships), the tasks of target detection become increasingly complex because there exist fewer target pixels, complex background interference data, and an irregular distribution of vehicle targets [4]. The detection of vehicles in RSI has been investigated for many years and has attained promising outcomes.

The deep learning (DL) methods and the traditional ML approaches are the two classes of target detection techniques using aerial images. The low-level image features, including color, corner, edge, shape, and texture, are extracted for classification and training in the ML approaches [5]. Researchers have introduced a target detection architecture, which uses a local binary pattern (LBP) in combination with histogram of gradients (HOG), for the detection of vehicle targets. The color difference is used for the recognition of blob-like areas extracted from grayscale and color features [6]. Moreover, there are CV technologies that use the optical flow and frame difference for the detection of moving vehicles. Vehicle detection aims to identify each instance of the vehicle in RSI. In former approaches, researchers have developed and extracted features and later categorized them to accomplish target detection [7]. The basic concept is to extract vehicle features and utilize ML methods for classification. RSI has been extensively utilized in several domains, namely urban and rural planning, disaster monitoring, national defense construction, agricultural management, and so on [8]. There also exists a huge gap between the detection performance and accuracy compared to the tremendous growth of DL techniques. Network models based on DL approaches have been prominently employed for large computer vision domains and have shown effective results [9]. Two categories of detection networks have been formed and continuously enhanced, single-phase networks and two-phase networks, with the progression of hardware technology and big data [10]. This study develops the Improved Deep Learning-Based Vehicle Detection for Urban Applications using Remote Sensing Imagery (IDLVD-UARSI) technique. The major aim of the IDLVD-UARSI method emphasizes the recognition and classification of vehicles in RSI. To achieve this, the IDLVD-UARSI method utilizes an improved RefineDet model for the vehicle detection and classification process. Once the vehicles are detected, the classification process takes place using the convolutional autoencoder (CAE) model. Finally, a Quantum-Based Dwarf Mongoose Optimization (QDMO) algorithm is applied to ensure an optimal hyperparameter tuning process and it helps in achieving improved performance. The simulation results of the IDLVD-UARSI technique are obtained on benchmark databases. The major contributions of the paper are given as follows.

• The IDLVD-UARSI technique significantly advances vehicle detection in urban environments using an improved RefineDet model, CAE classification, and QDMO-based hyperparameter tuning. To the best of our knowledge, the IDLVD-UARSI technique has never existed in the literature.
• It introduces an improved RefineDet model tailored to the challenges of urban contexts, resulting in more accurate and robust vehicle detection. Thus, CAEs can effectively extract and improve the accuracy of vehicle classification.
• The application of the Quantum-Based Dwarf Mongoose Optimization (QDMO) algorithm for hyperparameter tuning is a unique and innovative contribution. QDMO fine-tunes the model’s parameters, optimizing the overall performance of the vehicle detection and classification process.

The rest of the paper is organized as follows. Section 2 provides the related works and Section 3 offers the proposed model. Then, Section 4 gives the result analysis and Section 5 concludes the paper.

2. Related Works

Tan et al. [11] utilized a DL method to discover vehicles in high-resolution satellite RSI. Initially, the images were categorized by employing the AlexNet model, and then the vehicle target identification capability of the Faster R-CNN technique was verified. In [12], the authors developed a backbone architecture with a context data component and attention module to detect vehicles from a real image, which allowed the extraction feature network to improve the deployment of context data and prominent areas. Rafique et al. [13] presented a new vehicle recognition and segmentation technique for the monitoring of smart traffic, which employed a CNN for real image classification. Later, the identified vehicles were grouped into distinct subgroups. Eventually, detected vehicles were monitored by kernelized filter and Kalman filter (KF)-based methods.

In [14], the authors selected a single-phase DL-based target detection approach for research depending on the model’s real-time processing necessities. Moreover, the study improved the YOLOv4 architecture and introduced a novel technique. Primarily, a classification setting of the non-maximal suppression threshold was developed to improve the performance without impacting the speed. Secondarily, the authors analyzed the anchor frame allocation difficulty in YOLOv4 and developed two allocation systems. The authors [15] suggested an unsupervised domain adaptation algorithm that does not need trained data and can retain identification outcomes in the target domain at a minimal cost. The authors enhanced adversative domain adaptation by employing rehabilitation loss to enable the learning of semantic features. Xiao et al. [16] developed a bi-directional feature fusion module (BFFM) to construct a multiscale feature pyramid to detect targets at various measures. A feature decoupling model (FDM) was developed that uses the fusion of channel and spatial attention. In addition, a localization refinement module (LRM) was developed to automatically optimize the anchor box constraints to attain the spatial arrangement of anchor boxes and object regression features.

In [17], the authors recommended an innovative Multi-Scale and Occlusion Aware Network (MSOA-Net) for the classification of UAV-based vehicles. The MSFAF-Net with self-adaptive FFM was developed. In RATH-Net, a self-attention module is suggested to guide the location-sensitive sub-networks. Cao et al. [18] presented a novel OD method that makes improvements to the YOLOv5 network. The integration of RepConv, BiFPN, Transformer, and Encoder techniques into the unique YOLOv5 results in increased accuracy of detection. The C3GAM algorithm was developed by presenting the GAM attention module to overcome the issue of difficult contextual regions.

In [19], the authors focused on real-time smaller vehicle detection for UAVs from RSI and presented a depthwise-separable attention-guided network (DAGN) dependent on YOLOv3. Primarily, the authors integrated the feature concatenation and attention block to provide the model with a great capability to separate essential and irrelevant features. Afterwards, the authors enhanced the loss function and candidate integration approach from YOLOv3. Zheng et al. [20] presented a learning approach termed Vehicle Synthesis GANs (VS-GANs) for the generation of annotated vehicles in RSI. This approach rapidly creates higher-quality annotated vehicle data instances and significantly assists in the training of vehicle detectors. Yu et al. [21] examined a convolution CapsNet to detect vehicles in higher-resolution RSI. Primarily, a testing image was segmented into super-pixels to generate meaningful and non-redundant patches. Afterwards, these patches were input into a convolution CapsNet to label them as vehicles or background.

Although the employment of DL approaches for vehicle classification from RSI has demonstrated its ability, there is a lack of widespread research on automated hyperparameter tuning approaches tailored specifically to this application. Considering and developing hyperparameter optimization systems that are effective for RSI is important. Several researchers have focused on certain databases for the testing and training of DL approaches. However, there is a need for research that explores the generalized methods and hyperparameters across different RS databases. Emerging methods and hyperparameter settings, which are adjusted to distinct data sources and imaging conditions, are major problems. Resolving these research gaps will lead to more effective, correct, and adjustable DL-based vehicle classification approaches for RSI applications. It also supports the broader field of RSI by developing the combination of DL approaches and optimizer systems.

3. The Proposed Model

In this study, we focus on the design and development of the IDLVD-UARSI method using RSI. The major aim of the IDLVD-UARSI approach emphasizes the detection and classification of vehicles in RSI. In the IDLVD-UARSI technique, three major processes are involved, namely improved RefineDet vehicle detection, CAE-based classification, and QDMO-based hyperparameter tuning. Figure 1 describes the overall workflow of the IDLVD-UARSI method.

3.1. Vehicle Detector: Improved RefineDet Model

The improved RefineDet approach is applied for vehicle target detection. Improved RefineDet uses the VGG16 architecture as a backbone network, which constructs a set of anchors with an aspect ratio and scale from the feature map through the anchor generative model of RPN. After two classifications and regressions [22], it attains a certain amount of object bounding boxes, along with the existence probabilities of the various classes in the bounding box. The object detection module (ODM), anchor refinement module (ARM), and transfer connection block (TCB) are the three modules of the proposed algorithm. Lastly, the final regression and classification are achieved by non-maximum suppression (NMS).

(1) The ARM module primarily includes additional convolution layers and the backbone network, VGG16. Mainly, a negative anchor filter, feature extraction, anchor generation, and anchor refinement are implemented by the ARM module. The negative anchor filter can effectively extract the negative anchor box and mitigate the sample imbalance. Then, the fused feature is transformed into lower-level features via TCBs, such that the low-level mapping feature exploited for detection has a high-level semantic dataset and enhances the outcomes of floating objects.

(2) The TCB module interconnects ODM and ARM and then transfers the ARM data to ODM. Like the architecture of FPN, the nearby TCB is connected to increase the semantic dataset of lower-level features and realize the feature fusion of high- and low-level features.

(3) The ODM module primarily includes prediction layers (classification and regression layers, viz., 3 × 3 convolution kernels) and the output of TCBs. The outcome of the prediction layer is the specific category and the coordinate offset relative to the anchor refinement boxes. The anchor refinement boxes are exploited as input for the regression and classification, and, based on NMS, the final bounding box is selected.

3.2. Vehicle Classification: CAE Model

In this work, the CAE model is utilized for the identification and classification of vehicles. As a variant of AE, CAE integrates the capability of autoencoder (AE) to represent the input dataset and the capability of CNNs to effectively extract image features [23]. An encoder and decoder are two NN blocks used to recreate the input. The encoded block is used to encode the input $x$ into hidden outcome $l$, which is a compressed form of input. The dimension of the $l$ hidden outcome is smaller than that of the input $x$. The decoder relies on the $l$ hidden output and generates the output, $\tilde{x}$, to the input, $x$. AE is trained to minimize the reconstructed loss of the network for the proper regeneration of the original input. Figure 2 displays the infrastructure of CAE.

AE comprises an encoder $f\left(x, \theta \right)$ that compresses the information and decoder $g\left(l, \mathrm{\varphi }\right)$ that reconstructs the input as follows.

$$l=f\left(x, \theta \right)$$

(1)

$$\tilde{x}=g\left(l, \mathrm{\varphi }\right)=g\left(f\left(x, \theta \right), \mathrm{\varphi }\right)$$

(2)

where $\mathsf{\theta }$ and $\mathrm{\varphi }$ denote the trained parameters of the encoder and decoder, correspondingly. By training AE, this parameter is set to minimize the cost function.

$$\mathcal{C}\left(x,\tilde{x}\right)=\mathcal{L}\left(x,\tilde{x}\right)+regularizationterm$$

(3)

Here, $\mathcal{L}\left(x,\tilde{x}\right)$ indicates a reconstructed loss function, i.e., a binary cross-entropy $\left(BCE\right)$ or mean squared error $\left(MSE\right)$ averaged over the overall amount of samples, which are evaluated as follows:

$${\mathcal{L}}_{MSE}=\frac{1}{N}{\displaystyle \sum _i^N}{\left(x^i-{\tilde{x}}^i\right)}^2$$

(4)

$${\mathcal{L}}_{BCE}=\frac{1}{N}{\displaystyle \sum _i^N}-\left[x^i\log \left({\tilde{x}}^{\mathrm{i}}\right)+\left(1-{\tilde{x}}^{\mathrm{i}}\right) \log \left(1-{\tilde{x}}^{\mathrm{i}}\right)\right]$$

(5)

In Equation (3), to prevent the network from learning the identity function, the regularization term is essential for the DNN, thereby increasing the generalization ability. In CAE, the pooling layer is used to prevent the overfitting of the network and enhances the generalization capability of the method. Therefore, the reconstructed loss is exploited as a cost function. The encoded CAE comprises convolution and pooling layers for the fully connected (FC) and network layers. A hidden layer is signified as an FC layer (1D layer). The decoder mainly includes a sequence of transposed convolutional layers with FC layers. However, transposed convolution tends to produce checkerboard artifacts and is replaced with the alternative layer of upsampling and convolution. Upsampling, convolution, and pooling are three major operations of CAE.

3.3. Hyperparameter Tuning: QDMO Algorithm

Lastly, the QDMO method is exploited for the optimum selection of the parameters of the CAE architecture. DMO was developed in [24] as an MH method that mimics the behavior of a DM looking for rich food. The scout group stage takes place when the new sleeping mound (SM) (i.e., food source) is found depending on the old SM. The DMO method initializes the population of mongoose ($X$) and begins with the generation of the matrix dimension $(n$x$d)$, where the problem dimension ($d$) is presented in a column, and the size of the population ($n$) is signified in rows.

$$X=\left[\begin{array}{llll}{\chi }_{1,1}\hfill & x_{1,2}\hfill & \dots \hfill & x_{1,d}\hfill \\ x_{2,1}\hfill & x_{2,2}\hfill & \dots \hfill & x_{2,d}\hfill \\ ⋮\hfill & ⋮\hfill & {\chi }_{i,j}\hfill & ⋮\hfill \\ x_{n,1}\hfill & x_{n,2}\hfill & \dots \hfill & x_{n,d}\hfill \end{array}\right]$$

(6)

where ($x$) represents the component of $\mathrm{j}\mathrm{th}$ problem dimensions at $i\mathrm{th}$ population solution.

$${\chi }_j=unifrnd\left(LB, UB, D\right)$$

(7)

The value is based on Equation (7) as a uniformly distributed random integer constrained by $\left(UB\right)$ and $\left(LB\right)$, the lower and upper limitations of the problem. Next, the fitness value of the solution estimates its probability, and this procedure is expressed as follows:

$$\alpha =\frac{fi{t}_i}{{\displaystyle {\Sigma }_{i=1}^n}fi{t}_i}$$

(8)

where $n$, the number of mongoose, is upgraded to

$$n=n-bs$$

(9)

In Equation (9), the number of babysitters can be represented as $bs$, and the female alpha exploits different vocalization ($peep$) to communicate with others. Therefore, the DMO exploits Equation (10) to update the $X_i$ performance.

$$X_{i+1}=X_i+phi\times peep$$

(10)

where $phi$ represents a random integer in $\left[-1, 1\right]$ for each iteration. Moreover, the SM can be upgraded using Equation (11).

$$s{m}_i=\frac{fi{t}_{i+1}-fi{t}_i}{ \max \left\{\left|fi{t}_{i+1},fi{t}_i\right|\right\}}$$

(11)

After calculating the average SM $\left(\varphi \right)$, the formula is represented as follows:

$$\varphi =\frac{{\displaystyle {\Sigma }_{i=1}^n}s{m}_i}{n}$$

(12)

In the scouting phase, the new location of the candidate for food or SM is scouted, while the existing SM is ignored by the nomadic custom [25]. The movement regulating parameter $\left(CF\right)$, a movement vector ($M$), a stochastic multiplier ($rand$)$,$ and the existing position of the $SM$ are used to navigate the scouted location for food and SMs:

$$X_{i+1}=f\left(x\right)=\left\{\begin{array}{ll}{X}_i-CP\times phi\times rand\times \left[X_i-M\right], \hfill & {\varphi }_{i+1}>{\varphi }_i\hfill \\ X_i+CF\times phi\times rand\times \left[X_i-M\right], \hfill & otherwise\hfill \end{array}\right.$$

(13)

The new scouted location (X) and the future movement are simulated by the success and failure of the growth of a complete solution based on the mongoose group. The mongoose movement (M) determined in Equation (14) and handled by $CF$ is evaluated in Equation (15). At first, the collective–volitive parameters allow rapid exploration during the search process, but, utilizing each iteration of the work, we gradually shift towards defining a new region to exploit an effective one.

$$M={\displaystyle \sum _{i=1}^n}\frac{X_i\times s{m}_i}{X_i}$$

(14)

$$CF={(1-\frac{iter}{{\mathrm{Max}}_{iier}})}^{\left(\frac{2\times iter}{{\mathrm{Max}}_{iter}}\right)}$$

(15)

The alpha group shifts its position with a babysitter in the evening or after midday, which provides the possibility to explore the younger colony. The population size may affect the ratio of caretakers to foraging mongoose. Equation (16) is modeled to imitate the exchange process in the early evening and late afternoon.

$$phase=\left\{\begin{array}{ll}Scout, \hfill & C<L\hfill \\ Babysitting,\hfill & C\ge L\hfill \end{array}\right.$$

(16)

The initial weight is set to 0 to ensure that the average weight of the alpha group is minimized. This procedure ends when the iteration count is reached and returns the best solution. In QBO, the features are considered as a quantum bit, where $\mathrm{q}$ defines the superposition of a binary value in the range of [0, 1]. The QDMO technique is based on the quantum-based optimization (QBO) approach. A binary number represents the feature selection (1) or removed (0) [25]. The following equation is used to generate the arithmetical modelling of the $Q-bit$.

$$\mathrm{q}=\mathsf{\alpha }+\mathrm{i}\mathsf{\beta }={\mathrm{e}}^{\mathrm{i}\mathsf{\theta }}, \left|\mathsf{\alpha }{|}^2+\right|\mathsf{\beta }{|}^2$$

(17)

In Equation (17), $\mathsf{\alpha }$ and $\mathsf{\beta }$ characterize the probability values of the $\mathrm{Q}$-bit as 0 and 1. The parameter $\mathsf{\theta }$ indicates the angle of $\mathrm{q}$, and it can be upgraded through ${\tan }^{-1}\left(\mathsf{\alpha }/\mathsf{\beta }\right)$:

$$\mathrm{q}\left(\mathrm{t}+1\right)=\mathrm{q}\left(\mathrm{t}\right)\times \mathrm{R}\left(\Delta \mathsf{\theta }\right)=\left[\mathsf{\alpha }\left(\mathrm{t}\right)\mathsf{\beta }\left(\mathrm{t}\right)\right]\times \mathrm{R}\left(\Delta \mathsf{\theta }\right)$$

(18)

$$\mathrm{R}\left(\Delta \mathsf{\theta }\right)=\left[\begin{array}{ll}\cos \left(\Delta \mathsf{\theta }\right)\hfill & -\sin \left(\Delta \mathsf{\theta }\right)\hfill \\ \sin \left(\Delta \mathsf{\theta }\right)\hfill & \cos \left(\Delta \mathsf{\theta }\right)\hfill \end{array}\right]$$

(19)

Now, the rotation angle of $\mathrm{i}\mathrm{th}$ $\mathrm{Q}$-bit of $\mathrm{j}\mathrm{th}$ $\mathrm{Q}$-solution is represented as $\Delta \mathsf{\theta }$. QBO is used to optimize the ability to balance among the exploitation and exploration of DMA while searching for better performance. At first, $\mathrm{N}$ agents represent the population generation. Each performance has $\mathrm{D}$ features and $\mathrm{Q}$-bits:

$${\mathrm{X}}_{\mathrm{i}}=\left[{\mathrm{q}}_{\mathrm{i}1}\left|{\mathrm{q}}_{\mathrm{i}2}\right|\dots {|\mathrm{q}}_{\mathrm{iD}}\right]=\left[{\mathsf{\theta }}_{\mathrm{i}1}\left|{\mathsf{\theta }}_{\mathrm{i}2}\right|\dots {|\mathsf{\theta }}_{\mathrm{iD}}\right],\mathrm{i}=1,2,\dots ,\mathrm{N}$$

(20)

The superposition of probability for the features selected or not is characterized by ${\mathrm{X}}_{\mathrm{i}}$. The primary goal of QDMO is to upgrade the agents until they meet the final conditions:

$${\mathrm{BX}}_{,\mathrm{j}}=\left\{\begin{array}{ll}1\hfill & \mathrm{if} \mathrm{rand}< |\mathsf{\beta }{|}^2\hfill \\ 0\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(21)

Here, $\mathrm{rand}\in \left[0, 1\right]$ shows the random integer. The fitness selection is an important element in the QDMO method. An encoder solution is applied to measure the goodness (aptitude) of the candidate results. Here, the accuracy values are the primary condition employed in developing an FF.

$$\mathrm{Fitness} = \max \left(\mathrm{P}\right)$$

(22)

$$\mathrm{P}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$

(23)

Here, $\mathrm{TP}$ and $\mathrm{FP}$ signify the true and false positive values.

4. Results and Discussion

The performance of the IDLVD-UARSI method was tested using two databases: the VEDAI database [26] and the ISPRS Potsdam Database [27]. The VEDAI dataset comprises RGB tiles (1024 × 1024 px) and the associated infrared (IR) images at a 12.5 cm spatial resolution. For each tile, annotations are provided with the vehicle class, the coordinates of the center, and the four corners of the bounding polygons for all the vehicles in the image. VEDAI is used to train a CNN for vehicle classification. The ISPRS Potsdam Semantic labeling dataset is composed of 38 ortho-rectified aerial IRRGB images (6000 × 6000 px) at a 5 cm spatial resolution, taken over the city of Potsdam (Germany). A comprehensive pixel-level ground truth is provided for 24 tiles. The VEDAI database includes 3687 samples and the ISPRS Potsdam database includes 2244 samples, as defined in

Table 1. Description of the VEDAI database.

VEDAI Database
Class	No. of Samples
Car	1340
Truck	300
Van	100
Pickup Car	950
Boat	170
Camping Car	390
Other	200
Plane	47
Tractor	190
Total No. of Samples	3687

and

Table 2. Description of ISPRS Potsdam database.

ISPRS Potsdam Database
Class	No. of Samples
Car	1990
Truck	33
Van	181
Pickup Car	40
Total No. of Samples	2244

. Figure 3 demonstrates the sample visualization results offered by the proposed model. The figure illustrates that the proposed model properly classifies different classes of vehicles.

Figure 4 exhibits the classifier outcomes of the IDLVD-UARSI method on the VEDAI database. Figure 4a,b describe the confusion matrix presented by the IDLVD-UARSI technique at 70:30 of the TR set/TS set. The outcome indicates that the IDLVD-UARSI method has detected and classified all nine class labels accurately. Similarly, Figure 4c exhibits the PR examination of the IDLVD-UARSI system. The figure shows that the IDLVD-UARSI technique has obtained the highest PR outcomes in all nine classes. Finally, Figure 4d shows the ROC analysis of the IDLVD-UARSI method. It shows that the IDLVD-UARSI system has resulted in promising outcomes with high ROC performance on all nine class labels.

The vehicle recognition results of the IDLVD-UARSI technique are examined on the VEDAI database in

Table 3. Vehicle recognition outcomes of IDLVD-UARSI technique on VEDAI database.

VEDAI Database
Class	${\mathit{A}\mathit{c}\mathit{c}\mathit{u}}_{\mathit{y}}$	${\mathit{P}\mathit{r}\mathit{e}\mathit{c}}_{\mathit{n}}$	${\mathit{R}\mathit{e}\mathit{c}\mathit{a}}_{\mathit{l}}$	${\mathit{F}}_{\mathit{S}\mathit{c}\mathit{o}\mathit{r}\mathit{e}}$	MCC
Training Phase (70%)
Car	96.16	92.22	97.91	94.98	91.99
Truck	97.44	90.16	77.46	83.33	82.23
Van	98.18	72.09	46.97	56.88	57.34
Pickup Car	98.29	95.32	98.19	96.74	95.60
Boat	97.79	77.36	71.30	74.21	73.12
Camping Car	97.09	80.47	93.36	86.44	85.11
Other	97.40	79.66	68.61	73.73	72.59
Plane	98.76	100.00	13.51	23.81	36.53
Tractor	98.33	85.94	81.48	83.65	82.81
Average	97.72	85.91	72.09	74.86	75.26
Testing Phase (30%)
Car	96.84	93.28	97.91	95.54	93.16
Truck	97.47	85.54	81.61	83.53	82.19
Van	98.83	88.89	70.59	78.69	78.64
Pickup Car	97.20	91.26	98.60	94.79	93.01
Boat	98.55	95.35	74.55	83.67	83.62
Camping Car	98.01	88.36	96.27	92.14	91.12
Other	96.75	75.47	63.49	68.97	67.54
Plane	99.10	00.00	00.00	00.00	00.00
Tractor	98.28	90.91	72.73	80.81	80.47
Average	97.89	78.78	72.86	75.35	74.42

. The outcomes show that the IDLVD-UARSI system properly recognizes the vehicles. On 70% of the TR set, the IDLVD-UARSI method gives average ${accu}_y$, ${prec}_n$, ${reca}_l$, $F_{score}$, and MCC of 97.72%, 85.91%, 72.09%, 74.86%, and 75.26%, correspondingly. Moreover, on 30% of the TS set, the IDLVD-UARSI technique offers average ${accu}_y$, ${prec}_n$, ${reca}_l$, $F_{score}$, and MCC of 97.89%, 78.78%, 72.86%, 75.35%, and 74.42%, correspondingly.

Figure 5 displays the training accuracy ${TR\_accu}_y$ and ${VL\_accu}_y$ of the IDLVD-UARSI technique on the VEDAI database. The ${TL\_accu}_y$ is determined by the evaluation of the IDLVD-UARSI technique on the TR database, whereas the ${VL\_accu}_y$ is computed by evaluating the performance on a separate testing database. The results show that ${TR\_accu}_y$ and ${VL\_accu}_y$ increase with an increase in epochs. Accordingly, the performance of the IDLVD-UARSI technique is improved on the TR and TS databases with the maximum number of epochs.

In Figure 6, the $TR\_loss$ and $VR\_loss$ outcomes of the IDLVD-UARSI method on the VEDAI database are shown. The $TR\_loss$ defines the error among the predictive performance and original values on the TR data. The $VR\_loss$ represents a measure of the performance of the IDLVD-UARSI technique on individual validation data. The results show that the $TR\_loss$ and $VR\_loss$ tend to decrease with rising epochs. This indicates the enhanced performance of the IDLVD-UARSI technique and its capability to generate accurate classification. The reduced values of $TR\_loss$ and $VR\_loss$ reveal the enhanced performance of the IDLVD-UARSI technique in capturing patterns and relationships.

The comparison results of the IDLVD-UARSI technique on the VEDAI database are reported in

Table 4. ${Accu}_y$ outcomes of IDLVD-UARSI technique and other methods on VEDAI database.

VEDAI Database
Methodology	Accuracy
AlexNet Model	92.69
VGG16 Model	94.90
Inception Model	95.78
DenseNet-161 Model	96.01
EfficientNet Model	96.97
IDLVD-UARSI	97.89

and Figure 7 [28,29]. The experimental outcomes show that the AlexNet and VGG16 techniques have poor performance. In addition, the Inception, DenseNet-161, and EfficientNet models show moderately improved results. Nevertheless, the IDLVD-UARSI technique exhibits improved results, with maximum ${accu}_y$ of 97.89%.

Figure 8 shows the classifier outcomes of the IDLVD-UARSI method on the ISPRS Potsdam database. Figure 8a,b describe the confusion matrix presented by the IDLVD-UARSI method at 70:30 of the TR set/TS set. The outcome indicates that the IDLVD-UARSI method has detected and classified all four class labels accurately. Similarly, Figure 8c demonstrates the PR examination of the IDLVD-UARSI system. The figure shows that the IDLVD-UARSI method has gained the maximum PR performance on all four classes. Finally, Figure 8d shows the ROC analysis of the IDLVD-UARSI method. The outcome reveals that the IDLVD-UARSI system has resulted in promising outcomes, with high ROC values on all four class labels.

The vehicle recognition results of the IDLVD-UARSI technique are examined on the ISPRS Potsdam database in

Table 5. Vehicle recognition outcomes of IDLVD-UARSI technique on ISPRS Potsdam database.

ISPRS Potsdam Database
Class	${\mathit{A}\mathit{c}\mathit{c}\mathit{u}}_{\mathit{y}}$	${\mathit{P}\mathit{r}\mathit{e}\mathit{c}}_{\mathit{n}}$	${\mathit{R}\mathit{e}\mathit{c}\mathit{a}}_{\mathit{l}}$	${\mathit{F}}_{\mathit{S}\mathit{c}\mathit{o}\mathit{r}\mathit{e}}$	MCC
Training Phase (70%)
Car	97.71	98.09	99.35	98.72	88.23
Truck	99.62	100.00	76.00	86.36	87.01
Van	98.28	93.75	84.00	88.61	87.84
Pickup Car	99.17	75.86	78.57	77.19	76.78
Average	98.69	91.92	84.48	87.72	84.96
Testing Phase (30%)
Car	96.74	96.75	99.67	98.19	82.73
Truck	99.41	83.33	62.50	71.43	71.89
Van	97.63	95.45	75.00	84.00	83.44
Pickup Car	99.41	100.00	66.67	80.00	81.40
Average	98.29	93.89	75.96	83.40	79.86

. The experimental outcomes show that the IDLVD-UARSI method properly detected the vehicles. On 70% of the TR set, the IDLVD-UARSI system shows average ${accu}_y$, ${prec}_n$, ${reca}_l$, $F_{score}$, and MCC of 98.69%, 91.92%, 84.48%, 87.72%, and 84.96%, respectively. Moreover, on 30% of the TS set, the IDLVD-UARSI technique offers average ${accu}_y$, ${prec}_n$, ${reca}_l$, $F_{score}$, and MCC of 98.29%, 93.89%, 75.96%, 83.40%, and 79.86%, correspondingly.

Figure 9 presents the training accuracy ${TR\_accu}_y$ and ${VL\_accu}_y$ of the IDLVD-UARSI technique on the ISPRS Potsdam database. The ${TL\_accu}_y$ is determined by the evaluation of the IDLVD-UARSI technique on the TR database, whereas the ${VL\_accu}_y$ is computed by evaluating the performance on a separate testing database. The results show that ${TR\_accu}_y$ and ${VL\_accu}_y$ increase with an increase in epochs. Accordingly, the performance of the IDLVD-UARSI technique is improved on the TR and TS databases with a rise in the number of epochs.

In Figure 10, the $TR\_loss$ and $VR\_loss$ results of the IDLVD-UARSI technique on the ISPRS Potsdam database are shown. The $TR\_loss$ defines the error among the predictive performance and original values on the TR data. The $VR\_loss$ represents a measure of the performance of the IDLVD-UARSI method on individual validation data. The results indicate that the $TR\_loss$ and $VR\_loss$ tend to decrease with rising epochs. This highlights the enhanced performance of the IDLVD-UARSI technique and its capability to generate accurate classification. The reduced values of $TR\_loss$ and $VR\_loss$ reveal the enhanced performance of the IDLVD-UARSI technique in capturing patterns and relationships.

The comparison results of the IDLVD-UARSI technique on the ISPRS Potsdam database are reported in

Table 6. ${Accu}_y$ outcomes of IDLVD-UARSI technique and other methods on ISPRS Potsdam database.

ISPRS Potsdam Database
Methodology	Accuracy
AlexNet Model	94.38
VGG16 Model	94.99
Inception Model	95.76
DenseNet-161 Model	96.98
EfficientNet Model	97.15
IDLVD-UARSI	98.69

and Figure 11. The experimental outcomes show that the AlexNet and VGG16 approaches have poor performance. In addition, the Inception, DenseNet-161, and EfficientNet models show moderately improved results. Nevertheless, the IDLVD-UARSI technique exhibits improved results, with maximum ${accu}_y$ of 98.69%.

These performance results highlight the superior outcomes of the IDLVD-UARSI methodology compared with other methods. The improved solution of the IDLVD-UARSI algorithm is due to the employment of QDMO-based hyperparameter tuning, which appropriately chooses the optimum values for the hyperparameters of the provided CAE approach. Hyperparameters are settings that cannot be learned in the training but are supposed to be set earlier in training. They have a major impact on the model performance, and choosing an optimum solution leads to the best accuracy. By integrating QDMO-based hyperparameter tuning, the IDLVD-UARSI system obtains optimum solutions by focusing on better settings for the method. These outcomes ensure the greater solution of the IDLVD-UARSI methodology compared with other existing systems.

5. Conclusions

In this study, we focus on the design and development of the IDLVD-UARSI method using RSI. The major aim of the IDLVD-UARSI method focuses on the detection and classification of vehicle targets in RSI. In the IDLVD-UARSI technique, three major processes are involved, namely improved RefineDet vehicle detection, CAE-based classification, and QDMO-based hyperparameter tuning. The design of the QDMO technique helps in the optimal hyperparameter tuning process and it aids in achieving improved performance. The simulation results of the IDLVD-UARSI technique are obtained on a benchmark vehicle database. The simulation values indicate that the IDLVD-UARSI algorithm outperforms other recent DL models under various metrics. In the future, we aim to extend the IDLVD-UARSI technique by the use of feature fusion approaches. Moreover, the combination of multi-modal data sources such as multi-spectral images and LiDAR data offers richer data for classification. Examining systems to efficiently fuse and exploit these several data types from DL approaches will lead to better accuracy. Furthermore, the development of lightweight and effective DL techniques appropriate for utilization in resource-constrained remote sensing environments like satellites or drones is critical for real-time and on-device classification.