Hybrid methods for land use and land cover classification using remote sensing and combined spectral feature extraction: A case study of Najran City, KSA

1 Introduction

Land use and land cover (LULC) classification helps identify and track changes in land resources, such as urban areas, agriculture, forests, and water bodies. Such an understanding will provide a basis for creating sustainable cities [1], protecting agricultural land, maintaining ecological balance, and planning for land-use efficiencies in response to present and future demands. Accurate land use maps would help the government and planners manage natural resources such as water and forests more efficiently [2]. Understanding land use enables authorities to allocate resources more efficiently, preventing over-exploitation while ensuring the preservation of critical ecosystems.

The use of LULC classification techniques in Najran, Saudi Arabia, has diverse implications. This method categorizes LULC types and provides insights into decision-making and sustainable planning [3]. This aids in understanding the spatial distribution of urban, residential, agricultural, industrial, and natural land categories, thereby guiding resource utilization strategies [4]. Tracking changes includes urban growth and altered landscapes, informing urbanization trends, policy efficacy, and future trajectories. Integrating remote sensing and GIS enhances precision and identifies nuanced variations [5]. The results aid in addressing environmental issues, vegetation loss, and water body changes, thereby informing sustainable practices [6]. LULC classification involves categorizing different land use types and land cover from remote sensing or geospatial data. Feature extraction is a crucial step in this process, as it transforms raw data into meaningful representations that are used by machine learning algorithms for classification [7]. Remote sensing plays a crucial role in extracting spectral features for LULC classification. Traditional approaches to LULC classification utilize handcrafted features that require domain expertise to identify and design relevant feature-extraction methodologies [8]. Manual methods are error-prone, subjective, time-consuming, and unscalable for large datasets. This leads to low reproducibility and high variability among analysts. Issues of classification precision, difficulty in handling complex and dynamic landscapes, and limitations in traditional methods despite the advancements in Artificial intelligence (AI) technologies are still faced in the field [9]. AI methods, especially deep learning and hybrid approaches, automate feature extraction, reduce human errors, scale large datasets, and improve classification accuracy and consistency. There is a need for hybrid models that exploit the strengths of both deep learning and classical techniques to enhance performance. However, the emergence of deep learning has allowed for the automatic learning of features directly from raw data, thereby reducing the need for human intervention and improving classification accuracy [10]. Machine learning has been a key component of remote sensing classification for more than ten years and has performed well in LULC classification [11]. including the k-nearest neighbour, maximum likelihood estimator, SVM, random forest (RF), and decision tree [12]. Supervised learning models require training data and parameter adjustment [13]. Deep learning algorithms have recently attracted considerable attention for their potential to automatically classify remote sensing data [14]. This handles massive data issues and complicated land uses [15]. Several remote sensing applications have successfully used these techniques for image categorization and object segmentation [16]. Several studies have conducted a thorough analysis of remote sensing image classification based on Convolutional Neural Networks (CNNs) for land-use classification. The enhancements made to the CNN model describe how satellite image classification tasks have been carried out, and the findings demonstrated improved land use classification utilizing deep learning techniques [17]. In recent years, the application of deep learning techniques has witnessed remarkable progress in various fields, revolutionizing the extraction of complex patterns and information from raw data. Notably, environmental monitoring and analysis have greatly benefited from these advancements, enabling better comprehension and classification of the ever-evolving changes occurring within the planet [18]. Hybrid methods for spectral feature extraction using deep learning models have implications for advancing remote sensing technology, improving classification accuracy, supporting environmental monitoring, aiding disaster response, and promoting interdisciplinary collaborations [19]. These outcomes have led to more effective strategies for understanding and managing the Earth’s changes in an ever-changing world [20]. Multideep learning has demonstrated its ability to automatically learn complex patterns and representations from data. Still, it remains the hybrid method for spectral feature extraction using multi-deep learning models and using supervised classification of object-based classification has implications for advancing remote sensing technology, improving classification accuracy, supporting environmental monitoring, and aiding disaster response, where the outcomes lead to more effective strategies for understanding and managing the earth’s changes in an ever-changing world. Numerous challenges have arisen in the LULC classification. The advent of AI technologies has provided a partial resolution to some of these challenges. However, issues persist regarding the precision of LULC classification. Currently, hybrid methodologies have demonstrated efficacy in improving the accuracy of LULC classification. The hybrid approach in this study shows promising results in terms of the accuracy of LULC classification.

This article presents a proposed system of a hybrid method for spectral feature extraction using multi-deep learning models (AlexNet and GoogLeNet) and classification using an RF classifier to classify urban land changes in Najran City, the capital of the Najran Region in Saudi Arabia, using Landsat 8 satellite images from the year 2020. The main contributions of this study are as follows:

The main contributions of this study are as follows:

• Applying the methods to improve the Landsat 8 images of Najran City using the Stretch linear contrast (SLC) technique and the Laplacian filter;

• Applying a hybrid technique to classify Landsat 8 images of the city of Najran by combining the features of the AlexNet and GoogLeNet models; and

• A hybrid method of fusion of spatial features from the AlexNet and GoogLeNet models, followed by classification using RF algorithms.

This study is structured as follows. Section 2 reviews the existing literature concerning spectral feature extraction methods and the application of deep learning in earth change classification. Section 3 elaborates on the methodology employed, detailing the integration of spectral features.

2 Related work

The recent literature reflects significant advancements in remote sensing, deep learning, and hybrid methodologies for LULC classification. Studies have leveraged techniques and datasets ranging from Landsat and MODIS to high-resolution RGB imagery to enhance classification accuracy, temporal analysis, and spatial mapping. For instance, researchers have employed statistical projections, ensemble models, and deep learning frameworks, such as DenseNet, SqueezeNet, and Vision Transformers, to tackle LULC challenges across diverse geographies and periods. Despite these contributions, many existing approaches are grounded in single-model architectures, focusing narrowly on the spatial or spectral features.

Recent advancements in remote sensing, deep learning, and hybrid approaches have significantly enhanced LULC classification. Numerous studies have employed machine learning and deep learning techniques to achieve a high classification accuracy. For instance, Ray et al. [21] utilized Landsat data and achieved over 90% accuracy in a 30-year LULC assessment, whereas Nath et al. [22] employed maximum-likelihood classifiers to track long-term LULC changes. Similarly, Yang et al. [23] leveraged MODIS data to produce high-resolution global land cover maps, demonstrating the effectiveness of satellite-based classification.

Deep learning approaches have further improved LULC classification through automated feature extraction. Wang et al. [24] introduced OctSqueezeNet, a dual neural architecture that enhanced accuracy and efficiency, while Yaloveha et al. [25] demonstrated the superiority of DenseNet201 in land cover classification, achieving 92.01% test accuracy. Additionally, Manzanarez et al. [26] employed deep learning fusion models to automate large-scale land-cover labeling with minimal human intervention, further demonstrating the potential of these methods. Arfasa et al. [27] presented a study to forecast LULC change in the Vea catchment, Ghana, and its impact on irrigation water. They used the CA-Markov model for land-use predictions in 2038 and 2054 using Terrset software. The Relative Importance Index identifies the vital change drivers. The results indicated cropland growth from 181 m² (2038) to 183 m² (2054).

Despite these advancements, most existing studies have relied on single-model approaches, such as CNNs or vision transformers, without fully exploring the integration of spectral and spatial features. For example, while Meshram et al. [28] highlighted the benefits of bio-inspired and deep-learning models, their discussion did not address hybrid feature fusion. Similarly, Horry et al. [29] improved accuracy through ensemble models but did not investigate spectral feature extraction in depth. Furthermore, conventional methods, such as those used by Abd El Aal et al. [30] and Kalyan and Pathak [31], often overlook model biases and environmental variability, thereby limiting their generalizability. Alqahtany [32] analysed the 30-year urban growth of the AMA using GIS data and satellite image classification. The results showed that the urban area expanded from 199 sq.km in 1992 to 276 sq.km in 2022, with a 25.5% increase over the last decade. Statistical projections predict a further 26.8% growth by 2032. Land use analysis also revealed an 18% decline in vegetation over the past 10 years. Manikanta and Yaswanth [33] ambitiously applied the SWAT model to assess land use and climate change impacts on ecosystem vulnerability, relying on assumptions embedded in RCP 4.5 and RCP 8.5 scenarios without fully addressing model uncertainties or socio-economic dynamics. Historical data (1985–2023) and future predictions (up to 2100) were modelled using an ANN-Markov chain approach and validated at over 80% accuracy, but the limited validation timeframe may constrain predictive reliability.

A critical gap remains in the development of hybrid systems that systematically combine spectral feature extraction and ensemble classification. While some studies have employed multi-model frameworks, few have integrated pretrained CNNs with robust classifiers, such as RF, to enhance accuracy and interpretability. The proposed AlexNet-GoogLeNet-RF hybrid approach addresses this gap by:

• Leveraging complementary strengths: AlexNet extracts high-level spatial–spectral features, whereas GoogLeNet captures multi-scale spectral patterns through inception modules.

• Enhancing robustness: Feature fusion reduces the biases inherent in single-model approaches, and the RF classifier improves generalization through ensemble diversity.

• Improving interpretability: RF’s feature importance analysis provides clearer insights into model decisions, addressing a key limitation of purely deep-learning-based methods.

3 Materials and methods

3.1 Study area

This work studied changes in the capital city of Najran, Najran Region, in the Kingdom of Saudi Arabia, using Landsat8 satellite images from 2020 maps. Najran, a city in southwestern Saudi Arabia, is the provincial capital of Najran, as shown in Figure 1. One of the cities in the kingdom with the fastest population growth is Najran, which is designated as a new town. Najran is located at a latitude of 17.565604 and a longitude of 44.228944. The GPS coordinates for Najran in Saudi Arabia were 17°3356.1744′N, 44°13′44.1984E.

Figure 1

Geographical map of Najran City illustrating the study area case study from the 2020 Landsat-8 satellite imagery.

3.3 Preprocessing

The preprocessing approach entails analysing the precise position of the case study and accurately identifying the data once it has been retrieved from satellites using remote sensing technology.

Regarding land use dataset preparation, Landsat 8 imagery was clipped to be suitable for models and prepared to the appropriate size for data samples to be fed into the network architecture [34]. The sample set consisted of seven image bands (bands 1–7) and a labelled image. This annotates labelled images consisting of six semantic classes: background, forest, water, urban, miscellaneous, and agriculture. These classes are related to LULC Level 1 from the USGS. The dataset is shown in Figure 2.

Figure 2

Spectral reflectance variations across land use/land cover categories from Band 1 to Band 7 in the Najran City dataset.

The raw spectral bands of Red, Green, Blue, and Near-Infrared provide powerful tools for exploring and understanding the environment across various applications. Monitoring vegetation health and assessing land cover changes to capture stunning images and enable scientific research, raw spectral bands encompassing the Red, Green, Blue (RGB), and near-infrared (NIR) portions of the electromagnetic spectrum play a pivotal role in various fields, from remote sensing and environmental monitoring to photography and agriculture [35]. These spectral bands provide invaluable insights into the characteristics and composition of objects and landscapes, enabling researchers, scientists, and professionals to obtain critical information. Image Enhancement included noise removal and radiometric correction. These operations seek to improve satellite images for improved classification and repair of degraded images to depict the original scene accurately [36]. Owing to atmospheric effects and the sensing system’s limits, the likely range of pixel values is not fully utilized when acquiring data for creating remote sensing imagery. As a result, the received data are of low quality, with little contrast, a dark overlay, or much radiometric noise. The image was sharpened and smoothed using SLC, and its edges and textures were enhanced using a Laplacian.

The SLC and Laplacian filters were practically applied in this study and are not merely discussed in the literature. Their selection was made after carefully considering the specific characteristics of the Landsat 8 spectral bands and the requirements for an accurate LULC classification. The SLC technique was chosen for its simplicity, efficiency, and ability to expand the dynamic range of the raw satellite image, particularly in cases where the pixel values are compressed owing to atmospheric distortions. This enhancement ensures better visual separability of the spectral features, which is crucial for training robust classification models [37]. In contrast, the Laplacian filter was employed due to its strong ability to highlight rapid changes in pixel intensity, which corresponds to boundaries between different land classes. This makes it particularly useful for accentuating the edges and textures within the USGS map satellite images, primarily between urban and non-urban areas, thereby enhancing the accuracy of boundary detection in the classification task. The combination of SLC and Laplacian provided an optimal balance: SLC improved global contrast, and Laplacian refined local details.

These preprocessing steps directly contribute to the enhanced feature representation in the input data, ultimately improving the performance of the classification network.

3.4 Extraction of features

In the feature extraction stage, pixels or elements with similar characteristics are grouped into segments. This helps reduce noise and improve the accuracy of feature extraction. These features should capture spectral, textural, spatial, and contextual information indicative of different land-cover types [38]. Figure 3 shows the spectral feature extraction for LULC classification. Spectral feature extraction is crucial for LULC classification using remote sensing data such as satellite imagery. Spectral features capture information regarding the unique reflectance characteristics of different land cover types, which are used to distinguish and classify them [39]. The following step-by-step instructions on how to perform spectral feature extraction for LULC classification: Region of Interest (ROI) selection: Select representative regions corresponding to different land cover classes within the image. These ROIs were used to extract training samples for each class. Select diverse areas that capture the spectral variability of each land cover type. The spectral features were extracted from each pixel [40]. The standard spectral bands, indices, and transformations that are used as features include:

• Raw spectral bands (Red, Green, Blue, Near-Infrared);

• Vegetation indices (eNDVI, EVI); and

• Soil indices (NDMI, NDWI).

Figure 3

Visualization of spectral feature extraction for land use/land cover classification using AlexNet model.

3.5 Hybrid spectral feature fusion (AlexNet and GoogLeNet)

Spectral feature extraction is crucial in various image processing and computer vision tasks, particularly when dealing with data from different spectral bands and multispectral imagery. The goal of spectral feature extraction is to capture and represent the distinctive spectral characteristics of the input data that are utilized for classification, segmentation, and anomaly detection [41]. In spectral feature extraction, AlexNet and GoogLeNet improve results by learning hierarchical and abstract features from the data and leveraging deep learning architectures [42].

The hybridization approach integrates spectral features extracted from the outputs of the deep learning model. This study first extracts the experimental spectral features using AlexNet. The second step involved extracting spectral features using GoogLeNet.

AlexNet, GoogleNet, and the following hyperparameters were carefully tuned through empirical experimentation to achieve the best possible performance on the dataset. The final hyperparameters selected are listed in Table 2:

Table 2

Adjusting the hyperparameters of the AlexNet and GoogleNet models to classify the Najran city map

Hyperparameters	AlexNet	GoogleNet
Learning Rate	0.001	0.0001
Optimizer	Adam	Adam
Batch Size	30	35
Number of Epochs	50	60
Weight Decay	0.0005	0.0001

In this study, the hybrid fusion of spectral features extracted from AlexNet and GoogLeNet is carried out through a sequential feature extraction and fusion process, detailed as follows:

1. Spectral feature extraction with AlexNet:

   First, the multispectral satellite images, where each pixel contains information from several spectral bands, are input into a modified AlexNet model. AlexNet processes these inputs through its convolutional layers, extracting relevant spectral-spatial feature representations [41].

2. Spectral feature extraction with GoogLeNet:

   Second, the same multispectral images are processed separately using GoogLeNet. GoogLeNet employs its inception modules with different filter sizes to extract multi-scale spectral features, capturing both fine-grained and coarse spectral information [43].

3. Feature fusion process:

   After independently extracting feature vectors from AlexNet and GoogLeNet, the two sets of spectral features are combined sequentially. Specifically, the model concatenates the feature vectors output by both models into a single unified feature vector.

   This combined feature vector represents the complementary strengths of both AlexNet (strong spatial-spectral feature capturing) and GoogLeNet (strong multi-scale feature capturing).

4. Classification with RF:

The fused feature vectors are then input into an RF classifier, which uses the hybrid features to perform the final classification task [44].

This fusion strategy maximizes the spectral information extracted from multispectral imagery, leading to enhanced accuracy and robustness in LULC classification.

The major architectures, AlexNet and GoogLeNet, are meant for three-channel (RGB) inputs. However, multispectral satellite images usually carry more than three spectral bands (e.g. Red, Green, Blue, and Near-Infrared). To adapt this, the first convolutional layer of AlexNet and GoogLeNet was modified to accept the number of spectral bands in multispectral data analysis. That is, the input channels of the first convolutional layer are increased from three to the number of multispectral bands under consideration (e.g. four bands: Red, Green, Blue, NIR). Therefore, its convolutional layers were used to extract spatial and spectral features. In this way, having modified only the first layer of AlexNet and GoogLeNet, we preserved the ability of the network to effectively learn spectral-spatial representations without losing the advantages of being a pre-trained model. The features resulting from the modification were individually extracted from AlexNet and GoogLeNet, concatenating them into one feature vector combining attributes from hierarchical feature learning (AlexNet) and multi-scale feature extraction (GoogLeNet). Thus, the modifications resulted from careful design to accommodate multispectral input data while preserving the architectural integrity of both models.

Figure 4 illustrates a detailed spatial and statistical analysis of vegetation indices applied across the Najran geographic area.

Figure 4

Visualization of hybrid spectral feature extraction for land use/land cover classification using AlexNet model.

3.6 Proposed approaches

3.6.1 Training utilizing pre-trained methods

This study developed an initial methodology that leveraged the pre-trained models AlexNet and GoogLeNet to classify LULC attributes within the dataset. Diverse image-enhancement techniques have been implemented to refine the image quality [45]. These techniques encompass the application of a Gaussian filter to mitigate noise, precise cropping of land regions to emphasize pertinent areas, and data normalization for consistent standardization. The refined dataset resulting from the image enhancement procedures was channelled into the pre-trained models, wherein the convolutional layers of these models extracted pivotal land features from the satellite images [46]. This procedure entails the systematic movement of filters across satellite images, facilitating feature identification and the creation of corresponding feature maps. Convolutional and pooling layers were integrated to successfully encapsulate the extracted features by condensing the dimensionality of the feature maps. The outcomes originating from the pooling layers were then directed to the fully connected layers that functioned as a classifier. These layers meticulously scrutinized the ascertained features and generated predictions regarding the presence of LULC. These predictions were rooted in the patterns learned during the training phase.

3.6.2 Training of hybrid approaches

This approach involves analysing LULC characteristic satellite images of classes using a blend of techniques that combine features from various CNN models. Hybrid techniques in this study denote an approach that amalgamates multiple CNN models, AlexNet and GoogLeNet, to extract LULC features from input satellite images [47]. By employing this technique, the system strives to capitalize on the individual strengths of each model, thereby achieving more robust and precise feature representations. This method follows the sequence of steps outlined in Figure 5. Subsequently, the extracted combined features are stored in vectors. These combined features were fed into the RF classifier. The training set instructs the RF classifier to discern patterns and relationships between the LULC features and LULC categories. Following this, the testing set comes into play, gauging how effectively the trained models generalize to novel, unseen data [48].

Figure 5

A hybrid analytical framework combining spectral and deep learning features for land classification using an RF model.

3.6.3 RF classifier

RF is a popular machine learning algorithm for classification tasks. LULC classification involves categorizing different regions on the Earth’s surface into LULC classes, such as land area, forests, urban areas, water bodies, agricultural fields, etc. RF is particularly well-suited for this task due to its ability to handle complex and high-dimensional data and its robustness against overfitting [49]. RF prevents overfitting and provides reliable predictions through ensemble decision trees [20]. The idea is to introduce variability and diversity into the training process, which helps reduce overfitting and increases the model’s generalization ability. In addition to using random subsets of data, each decision tree in the RF uses a random subset of features when making decisions at each node. Each tree focuses on different aspects of the data, reducing the risk of individual trees dominating the decision-making process [50]. Each decision tree in the RF is constructed by recursively splitting the data into subsets based on the selected features. This process continues until a stopping criterion is met, such as a maximum tree depth or a minimum number of samples in a leaf node. Once all the trees are built, the RF algorithm aggregates the predictions of each tree to make a final prediction. For LULC classification, the most common approach is to use majority voting. Each tree is for a particular class, and the class with the most votes becomes the predicted class for the input data point. The categories have been seen in land change in Najran City. This study selected the following RF parameters through empirical experiments and cross-validation to optimize classification accuracy and reduce overfitting, as in Table 3. The QGIS software also produces the confusion matrix, which displays the percentage of pixels changing from one kind to the next [51]. Unclassified, Water Area, Built-up Area, Agriculture, Barren Land, and Other Land are the six samples for the six parameters that make up the input layers for model processing. These parameters are explained in Table 4. The samples based on the RGB colour composites of the Landsat8 images are the class Vegetation (red pixels in the RGB colour composite RGB = 432), which shows detailed changes in the area.

Table 3

Adjusting the parameters of the RF classifier to classify the terrain map of Najran city

Parameter	Value	Justification
Number of Trees (n_estimators)	100–300	A number stabilizes prediction; tested and validated
Maximum Depth (max_depth)	20–40	Controls overfitting; 30 was optimal in our experiments
Minimum Samples Split (min_samples_split)	2	Standard practice; allows maximum tree growth initially
Minimum Samples Leaf (min_samples_leaf)	1	Preserves pure class separation at the leaves
Maximum Features (max_features)	sqrt	Encourages randomness; recommended for classification tasks
Bootstrap Samples (bootstrap)	TRUE	Maintains randomness and reduces overfitting risk

Table 4

Description of LULC classes in the study area

Water Area	Rivers, dams, and seasonal water bodies (wadis)
Agriculture	Cultivated lands, farms, and irrigated fields located mainly around valleys and water sources
Built-up Area	Urban areas including residential, commercial, industrial structures, roads, and related infrastructure
Barren Land	Exposed soil, rocks, and desert areas without vegetation, characteristic of the region’s arid climate
Other Land	Unclassified or unknown land categories

3.6.4 Performance evaluation metrics

The evaluation metrics used in this study are provided in [1 2 3 4 5 6] equations. Accuracy measures the overall accuracy of the model in classifying all the categories. Sensitivity evaluates the ability of a model to correctly identify positive samples. Specificity measures the ability of a model to correctly identify negative samples. Precision reflects the proportion of correctly identified positive predictions among all positive predictions. AUC quantifies the entire two-dimensional area under the Receiver Operating Characteristic (ROC) curve. The AUC indicates a better model’s ability to distinguish between classes. The Kappa coefficient statistic measures the level of agreement between the predicted and true classes, adjusted for agreement occurring by chance [52].

(1) Accuracy = TN + TP TN + TP + FN + FP ⁎ 100 % ,

(2) Sensitivity = TP TP + FN ⁎ 100 % ,

(3) Specificity = TN TN + FP ⁎ 100 ,

(4) Precision = TP TP + FP ⁎ 100 % ,

(5) AUC = TP Rate FP Rate ,

(6) kappa = p o −   p e 1 − p e ,

where p o is the observed agreement (equivalent to Accuracy) and p e is the expected agreement by random chance, calculated as:

(7) p e = row total i × column total i N 2 ,

where k is the number of classes and N is the total number of samples.

These metrics were computed for each landform class (Water, Agriculture, Built-up, Barren Land, Other Land, and Unclassified) and overall performance, as summarized in Table 6.

4 Results

A comprehensive presentation of outcomes of the LULC classification system proposed for Najran City using diverse deep-learning models is found in Table 5. This study has implemented LULC classification of the Najran city map for Landsat 8 satellite images in 2020 using AlexNet, GoogLeNet, AlexNet with RF, GoogLeNet with RF, and combined spectral features of AlexNet-GoogLeNet with RF, as in Figure 6. The core of the principal component transformation relies on the global covariance matrix. This method often functions as a means of reducing the dimensionality of features, particularly when classes exhibit similar distributions. Spectral feature extraction involves deriving vectors that represent observations while reducing dimensionality. In the study of LULC classification, the goal is to extract features that effectively differentiate between classes.

Figure 6

Representative classified satellite image samples of Najran City generated by the proposed hybrid CNN-RF systems.

Table 5 clearly illustrates that hybrid feature extraction and classification approaches consistently outperform individual CNN models across most land classes, particularly those combining AlexNet and GoogLeNet with RF. For example, in the water area class, the highest area percentage (13.54%) is achieved when combining features from AlexNet and GoogLeNet with RF, significantly exceeding the results from AlexNet (4.96%) and GoogLeNet (7.25%) alone. A similar enhancement is observed in the agriculture class, where the hybrid model achieves 23.20%, significantly surpassing the individual CNN outputs.

The built‑up area shows relatively comparable values among all methods; however, the hybrid models, especially GoogLeNet with RF (14.81%), still exhibit strong performance. In contrast, for barren land, although GoogLeNet alone achieves the highest percentage (51.15%), the hybrid model combining features from AlexNet and GoogLeNet with RF also attains a robust 48.42%, demonstrating the general reliability and consistency of the hybrid strategy.

Moreover, the hybrid models again show superior area percentages for the other land and unclassified classes, with the combination of AlexNet-GoogLeNet and RF reaching 14.19% for the unclassified class.

In summary, the hybrid approach provides more effective feature representation and classification accuracy, as it captures diverse and complementary spatial features from both CNN architectures and leverages the RF’s robustness in classification.

5 Discussion and comparison

This study examines the outcomes and implications of applying the LULC classification approach, utilizing deep learning architectures and hybrid methods, for satellite image analysis in Najran City, Kingdom of Saudi Arabia. The Study evaluated the accuracy of the classification, the suitability of the chosen models (AlexNet, GoogLeNet, and hybrid models with RF networks), and the potential applications of the findings for urban planning and environmental management. The study reveals varying levels of accuracy in LULC classification, influenced by factors such as the complexity of land cover classes, image quality, and model performance. Notably, hybrid models that combine deep learning and RF networks demonstrate improved accuracy, underscoring the benefits of this approach. The selection of AlexNet and GoogLeNet as base models was significant due to their effectiveness in image classification. These models extract complex features, making them ideal for LULC classification. Adding RF networks enhances their adaptability and pattern recognition capabilities.

Complex classes (built-up, barren): the hybrid model’s precision and AUC for built-up (96.97 and 98.41%) and barren land (96.96 and 98.40%) exceed those of both single-model hybrids by 2–3 percentage points. This suggests that the complementary spatial-spectral representations learned by AlexNet (broader spatial patterns) and GoogLeNet (fine-scale features) are particularly effective for textured classes. Ambiguous classes of agriculture, other): for agriculture, the hybrid model’s accuracy (96.94%) closely matches GoogLeNet-RF  94.22%, reflecting persistent spectral overlap between sparse vegetation and barren soils in Landsat 8’s multispectral bands. These experimental results demonstrate that the hybrid model, which combines deep learning architectures (AlexNet, GoogLeNet) with an RF classifier, yields superior performance in LULC classification across all metrics. The hybrid model achieves an overall accuracy of 96.95% and a kappa coefficient of 0.9638, outperforming all standalone CNN models and other hybrid configurations. These improvements are particularly evident in classifying complex classes, such as built-up and barren land, where precision and AUC exceed 95%.

In the proposed models, as shown in Table 13, AlexNet with RF performed the best for water area classification, achieving 14750.85 km² (22.43%) of the total area. The hybrid AlexNet-GoogLeNet model (combine features) was second best at 8901.86 km² (13.54%). Using the GoogLeNet RF model, 15249.73 km² (23.20%) of the land has been classified as agricultural land, whereas the hybrid approach (combining features) could only classify 7829.80 km² (11.91%). For the built-up class, GoogLeNet with an RF once again obtained the highest area of 9733.26 km² (14.81%), and the hybrid model (combining features) got the second highest area of 8165.52 km² (12.42%). For barren land classification, the hybrid model (combining features) achieved the highest area of 31837.21 km² (48.42%), followed by AlexNet with RF, which achieved 28466.00 km² (43.30%). For other land classifications, the AlexNet with RF achieved the highest area of 4799.13 km² (7.30%), while the GoogLeNet with RF achieved 4489.81 km² (6.83%). For unclassified regions, the hybrid model (combine features) performed the best at 9331.09 km² (14.19%), and AlexNet with RF achieved 5476.47 km² (8.33%).

Most studies, such as those of Ratnadeep et al., Anindita et al., and Aixia et al., applied conventional or single deep learning models (e.g. maximum likelihood classifiers, basic CNNs) for LULC classification. However, they did not implement a hybrid feature fusion combining multiple CNN models and an RF classifier. Studies like those of Vladyslav et al. focus on standalone CNN models or ensemble CNN-Transformer models. Still, they lack the integration of pre-trained CNN feature maps with classical machine learning classifiers. Sergio et al. worked on deep learning label assignment but did not explore hybridization with spectral feature extraction or RF. Michael et al. applied ensemble strategies (e.g. CNN-ViT combinations, feature fusion from satellite modalities); a closer examination reveals specific gaps: Most existing ensemble methods primarily focus on either fusing data from multiple sources (e.g. Landsat, ALOS-2) or combining predictions from different deep learning models, without deeply addressing early-stage feature-level hybridization, especially spectral–spatial integration.

Spectral feature extraction using hybrid feature maps has mainly been overlooked in the literature. Traditional deep learning models rely on spatial patterns, while spectral richness remains underexploited. In contrast, our proposed AlexNet-GoogLeNet-RF hybrid system uniquely addresses the gap by extracting a richer set of spectral and spatial features from AlexNet and GoogLeNet and fusing multi-scale, multi-representation feature maps to improve land cover classification. RF as a final classifier enhances generalization and reduces overfitting, unlike most pure CNN-based or MLC-based approaches. Thus, compared to existing works, our hybrid methodology systematically combines spectral-spatial deep features with classical ensemble learning, offering better robustness, interpretability (through feature importance), and accuracy.

Our methodology achieved the highest overall accuracy (96.95%), significantly outperforming standalone CNN models. Moreover, the hybrid approach consistently demonstrated the best sensitivity, specificity, and precision across all land cover classes. The integration of principal component transformation further enhanced feature representation, improving the separability of similar classes.

Therefore, based on comparative analysis, the results presented in Table 13 validate the hybrid model’s strong capability in multiclass LULC classification, setting a new benchmark for satellite-based urban land analysis. These outcomes affirm the choice of AlexNet and GoogLeNet as base models for feature extraction and underscore the value of integrating RF networks to enhance classification reliability and reduce overfitting. The successful LULC classification in Najran City has important urban planning and environmental management implications. Accurate land cover maps support informed decision-making regarding land use regulations, infrastructure development, and ecosystem preservation. Additionally, it aids in monitoring urban sprawl and identifying environmental risks. However, there are Limitations of the proposed hybrid system: despite achieving high accuracy (96.95%), the study has several limitations that warrant discussion.

Spectral resolution constraints: The model relies on Landsat 8’s 11-band multispectral data, which may lack the granularity of hyperspectral sensors for distinguishing spectrally similar classes (e.g. barren land vs built-up area). Misclassifications occur where reflectance signatures overlap (e.g. dry soil and sparse urban structures).

Temporal and atmospheric effects: Cloud cover and seasonal variations (e.g. agricultural phenology) introduce noise. The study uses single-time point imagery (2020), limiting robustness to temporal changes.

Implications of misclassifications: Misclassifications have practical consequences for urban planning and environmental management. Mislabelling barren land as built up (e.g. due to similar textures) could overestimate urbanization, leading to unnecessary infrastructure investments. Under-detecting croplands (e.g. confusion with other land) may hinder food security assessments.

The proposed hybrid approach primarily relies on spectral feature extraction and principal component transformations grounded in the global covariance matrix. This global statistical basis ensures that the learned features capture underlying land cover patterns and relationships not exclusively tied to a specific geographic location.

Several factors contribute to the performance differences in the hybrid approach across various classes, such as agriculture, compared to others, due to the interplay between the dataset and model configurations. When compared to one or the other cnn models, such as AlexNet or GoogLeNet, the hybrid model’s performance strongly ameliorated some land cover classes that included built up, barren land, etc. for instance, the agriculture class is probably most affected by spectral signatures that are less distinct in multispectral imagery (like landsat 8), where vegetation is not dense, or the crop types have similar reflectance characteristics. The hybrid model effectively improves classification performance; however, separating agricultural areas from barren or built-up areas may be challenging due to the significant spectral overlap between these classes. Particularly, the feature extraction phase of the model, where AlexNet and GoogLeNet were combined, might not always highlight the subtle spectral differences necessary for correct agriculture classification. On the other hand, the hybrid model excels in more distinctive classes, such as barren land and built-up areas, where the spectral and spatial patterns are more separable.

6 Conclusions

This study introduced a novel hybrid approach proposed to effectively classify land changes within the urban landscape of Najran, located in the Kingdom of Saudi Arabia. By leveraging the power of satellite imagery from Landsat 8 in 2020, this research combines the distinctive features extracted from the AlexNet and GoogLeNet models. These integrated features are employed using the RF to achieve accurate and reliable land change classification. The hybrid system’s proposed methodology demonstrates a notable advancement in land change detection, offering a promising solution for urban planning, environmental monitoring, and resource management in Najran. The advantage of using a hybrid approach for LULC classification, which combines features of CNN models (AlexNet and GoogLeNet), is that it achieves improved accuracy, robustness, and flexibility compared to other methods. Furthermore, its outcomes underscore the importance of integrating machine learning and remote sensing techniques for comprehensive and efficient land change analysis. As demonstrated by its results, the hybrid system’s performance underscores its potential for application in a broader range of contexts, other regions, and periods, aiding decision-makers and researchers in understanding urban dynamics and facilitating informed decision-making.

Nonetheless, while the hybrid approach exhibits promising results, there are avenues for further enhancement and exploration. Future research will focus on refining feature extraction methods, exploring alternative machine learning algorithms, and incorporating multi-temporal data to improve land change classification. The findings, combined with future advancements, contribute to a growing body of knowledge that aids in the sustainable development and management of urban environments.