ADVANCED NEURAL NETWORK APPROACHES FOR TEXTURE RECOGNITION IN REMOTE SENSING DATA

ABSTRACT

This study explores advanced neural network approaches for texture recognition in satellite imagery, focusing on the integration of traditional feature extraction techniques with deep learning models. A hybrid framework is proposed that combines convolutional neural networks (CNNs) with handcrafted features derived from the Gray-Level Co-occurrence Matrix (GLCM) and wavelet transform coefficients. This fusion strategy addresses the limitations of conventional methods in capturing complex spatial patterns and enhances the performance of CNNs in distinguishing visually similar textures.

To ensure robustness, preprocessing and data augmentation techniques were employed to mitigate noise and reduce data dependency. The model was evaluated using the EuroSAT benchmark dataset, achieving a classification accuracy of 94.7%, which represents an improvement of 12.4% over GLCM-only models and 5.2% over standalone CNN architectures. Notably, the proposed method excelled in distinguishing closely related land cover types, such as urban and agricultural areas, achieving over 93% accuracy in both categories.

The findings confirm the effectiveness of multi-scale texture representation and feature fusion in remote sensing applications. This work provides practical insights for the development of efficient, cost-effective hybrid models for use in domains such as precision agriculture, urban planning, and environmental monitoring.

АННОТАЦИЯ

В данной работе рассматривается проблема точного распознавания текстур на спутниковых изображениях с использованием гибридного подхода, объединяющего традиционные методы анализа текстур и глубокие сверточные нейронные сети (CNN). Предлагаемая модель сочетает архитектуру CNN с признаками, извлечёнными с помощью матрицы совместной встречаемости уровней серого (GLCM) и вейвлет-преобразования, что позволяет компенсировать недостатки традиционных методов при распознавании сложных пространственных структур.

Для повышения устойчивости модели были применены методы предварительной обработки данных и аугментации, направленные на устранение зависимости от искажённых данных и снижение уровня шума. Модель была протестирована на эталонном наборе данных EuroSAT и достигла точности классификации 94.7%, что на 12.4% выше по сравнению с моделями, основанными только на GLCM, и на 5.2% выше по сравнению с обычными CNN.

Особо стоит отметить улучшение качества классификации схожих по визуальным характеристикам классов землепользования, таких как урбанизированные зоны и сельскохозяйственные угодья — точность распознавания по этим категориям превысила 93%. На основе полученных результатов предложена оптимизированная методология для задач дистанционного зондирования, подтверждена эффективность многомасштабного представления текстур и обозначены практические направления использования гибридных моделей при ограниченных ресурсах. Результаты могут быть применены в таких областях, как точное земледелие, градостроительство и экологический мониторинг [15][16].

Keywords: texture recognition, satellite images, deep learning, CNNs, remote sensing, GLCM, wavelet transform, land cover classification, feature extraction, data augmentation.

Ключевые слова: распознавание текстур, спутниковые изображения, глубокое обучение, CNN, GLCM, вейвлет-преобразование, классификация землепользования, извлечение признаков, аугментация данных.

Introduction

Accurate texture recognition in satellite imagery is a critical component of land cover classification and numerous remote sensing applications. However, traditional texture analysis techniques—such as the Gray-Level Co-occurrence Matrix (GLCM)—often struggle with the complex and heterogeneous spatial patterns commonly found in satellite data [1][2]. While Convolutional Neural Networks (CNNs) have achieved state-of-the-art performance in image classification tasks, they frequently fail to effectively capture fine-grained texture details, particularly in scenarios involving limited datasets or high levels of noise [3][4].

As illustrated in Figure 1, conventional methods face considerable limitations when analyzing complex spatial structures, underscoring the need for a hybrid approach that leverages the advantages of both traditional and deep learning techniques.

This study addresses the pressing need for improved texture recognition accuracy, particularly in distinguishing visually similar land cover types such as urban and agricultural regions. To tackle the weaknesses inherent in both standalone GLCM and CNN models, we propose a hybrid model that integrates CNN architectures with features derived from GLCM and wavelet transform coefficients. The key characteristics of this model include enhanced multi-level feature fusion and tailored preprocessing techniques specifically designed for satellite image data.

The primary contributions of this research are as follows:

• A novel feature fusion strategy that combines deep learning with handcrafted texture descriptors;
• A customized preprocessing pipeline to address challenges such as data scarcity, noise, and inconsistency in satellite imagery;
• Empirical validation of the proposed model on the EuroSAT benchmark dataset, achieving a classification accuracy of 94.7%, outperforming both GLCM-only and CNN-only baselines.

The experimental results demonstrate significant improvements in classification performance, especially in distinguishing visually similar classes. These findings offer practical value for remote sensing tasks in domains such as environmental monitoring, precision agriculture, and urban planning.

The remainder of this paper is structured as follows:
Section 2 describes the proposed methodology;
Section 3 presents the experimental setup and results;
Section 4 concludes the paper and outlines directions for future work.

Figure 1. Limitations of traditional methods in texture recognition

Methodology

In this research, a new hybrid approach to texture recognition in satellite images is proposed that merges convolutional neural networks (CNNs) with standard texture analysis methods. The methodology is developed for computational efficiency and high classification accuracy and is able to run on standard consumer hardware.

A. Dataset

A subset of 300 images (64x64 pixels each) was extracted from the EuroSAT dataset including three major land use classes with distinct texture signatures: urban (100 images), agricultural (100 images), and forests (100 images) [5]. Each land cover class was split into training (60%), validation (15%), and test (25%) as shown in Table 1.

Table 1.

Dataset distribution across training, validation, and testing sets

The dataset was deliberately scaled down in size to facilitate fast experimentation while still adhering to a statistical significance (p < 0.05). With 60 training samples per class, this configuration is conducive to effective transfer learning. The overall dataset size is limited to under 500MB, and total processing time is less than three hours on a standard laptop with a GPU.

B. Data Preprocessing

The first preprocessing step was min-max normalization, where pixel values were scaled to the [0, 1] range. Each class had filtering methods applied, including urban areas: A 3 × 3 median filter to preserve edge information. Agricultural fields: An unsharp mask filter to enhance crop patterns. Forests: A Gaussian blur (σ = 0.5) to reduce leaf noise for high-frequency information. Filters were chosen on an empirical basis to enhance class-specific texture features for more efficient model learning.

C. Feature Extraction

To achieve adequate texture encoding, we employed a dual-path feature extraction approach. Deep learning path: MobileNetV2 was used to extract 320-dimensional feature vectors from RGB patches. This model is a suitable choice for resource-constrained systems because it strikes a good balance between speed and accuracy, leading to a real-time model. Traditional path: Texture features were computed using a simplified GLCM (0° orientation, 1-pixel distance) methodology, where we derived features only from contrast and homogeneity.

Additionally, we employed single-level Haar wavelet decomposition (LH-band energy) to measure energy in different orientations of texture [10]. Originally, we considered EfficientNet-B0 because of its increased accuracy; however, at the end of the evaluation process, we went with MobileNetV2, as it was easier to implement and more efficient with computation (but also performed comparably) [7]. The entire extraction pipeline used less than 3GB of RAM and just above 200MB of storage.

Figure 2. Sample satellite image dataset

D. Model Training and Implementation

The models were trained in TensorFlow and PyTorch [6]. The CNNs used the Adam optimizer (learning rate = 0.001) and early stopped to limit overfitting. Performance was measured using the following metrics: accuracy, precision, recall, F1-score, and confusion matrix [11]. The results were compared between the GLCM SVM and the CNNs as independent models. The final prediction score was calculated as a weighted average of the CNN and GLCM classifiers:

Class Score=0.7× +0.3×                                      (1)

In this context,  indicates the probability output from the last depthwise layer ofMobileNetV2 , whereas ​ indicates confidence based on texture features. These weights were selected empirically, given the stated greater discriminative power of CNN features but also to use the complementary power of the texture descriptors.

Figure 3. Accuracy Comparison of different models

Results

The suggested hybrid model was tested on a curated subset of the EuroSAT dataset and assessed against both conventional texture analysis techniques and standalone deep learning architectures. The findings support that the hybrid model demonstrates clear superiority by effectively combining CNN-based features with statistical texture descriptors [17][18].

A. Performance Metrics

The hybrid model reached a total classification accuracy of 94.7%, higher than that of the traditional classifiers based on GLCM (82.3%) and standalone CNN models (89.5%) [8][12]. A summary of performance metrics for each class of land cover is in Table 2.

Table 2.

 Performance Metrics by Class

The high F1-scores for all classes are strong indications of model generalization and its robustness in differentiating visually similar land cover classes. Model improvements arise from the combined advantages of both deep features and artisanal texture descriptors. The F1-score applied to this study is expanded from its traditional formulation to account for the influence of features that were extracted, defined as:

  (2)

 where α and β weighting coefficients for taking into consideration additional features, Featureᵢ based on deep learning features, ExtraFeatureⱼ conventional texture descriptors (e.g., GLCM contrast/homogeneity, wavelet energy), 𝑛  and 𝑚 total features in each individual set[19].

D. Computational Efficiency

Even with extra processing for texture analysis, the hybrid model stayed feasible on the computer. Training time increased about 15% compared to only a CNN, which is primarily due to the GLCM and wavelet processing being added. Inference speed remained respectable, confirming its acceptability for satellite image pipeline applications [20]. Average training epoch time: 45 seconds on NVIDIA V100 GPU Total time to train: Less than 3 hours for the complete dataset Inference time per image: ~50 ms on consumer hardware [13].

These results show the hybrid method is able to balance performance vs. effort, making it an efficient method for remote sensing applications which are resource-constrained.

Conclusion

The experimental study demonstrates that the proposed hybrid model, combining convolutional neural networks (CNNs) with conventional texture analysis approaches, achieves a significant boost in classification performance for satellite image texture recognition. The study achieves a robust representation of land cover textures by integrating high-level semantic features from CNNs with low-level statistical descriptors, like the gray-level co-occurrence matrix (GLCM) and energy descriptors from wavelet-based analysis [8][9].

In particular, the hybrid approach provides 12.4% greater accuracy than the solely CNN model, and 15.2% greater accuracy than the GLCM classification method. The experiment indicated the efficacy of fusing features to improve discriminative power for classifying satellite imagery.

As well, the model maintains processing efficiency, allowing for use on consumer-grade hardware and reasonable processing times for a practical application.

References:

1. Haralick, R. M., Shanmugam, K., & Dinstein, I. (1973). Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, SMC-3(6), 610–621.
2. Manjunath, B. S., & Ma, W. Y. (1996). Texture features for browsing and retrieval of image data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 18(8), 837–842.
3. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 1097–1105.
4. Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint, arXiv:1409.1556.
5. Helber, P., Bischke, B., Dengel, A., & Borth, D. (2019). EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7), 2217–2226.
6. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770–778.
7. Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking model scaling for convolutional neural networks. Proceedings of the 36th International Conference on Machine Learning (ICML), 6105–6114.
8. Zhang, L., Zhang, L., & Tao, D. (2018). Deep learning for remote sensing data: A technical tutorial on the state of the art. IEEE Geoscience and Remote Sensing Magazine, 6(2), 22–40.
9. Chen, Y., Jiang, H., Li, C., Jia, X., & Ghamisi, P. (2016). Deep feature extraction and classification of hyperspectral images based on convolutional neural networks. IEEE Transactions on Geoscience and Remote Sensing, 54(10), 6232–6251. https://doi.org/10.1109/TGRS.2016.2584107
10. Ma, L., Liu, Y., Zhang, X., Ye, Y., Yin, G., & Johnson, B. A. (2019). Deep learning in remote sensing applications: A meta-analysis and review. ISPRS Journal of Photogrammetry and Remote Sensing, 152, 166–177. https://doi.org/10.1016/j.isprsjprs.2019.04.015
11. Aptoula, E., Ozdemir, M. C., & Yanikoglu, B. (2016). Deep learning with attribute profiles for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 13(12), 1970–1974. https://doi.org/10.1109/LGRS.2016.2619354
12. Yue, J., Zhao, W., Mao, S., & Liu, H. (2015). Spectral–spatial classification of hyperspectral images using deep convolutional neural networks. Remote Sensing Letters, 6(6), 468–477. https://doi.org/10.1080/2150704X.2015.1047045
13. Li, Y., Zhang, H., & Shen, Q. (2017). Spectral–spatial classification of hyperspectral imagery with 3D convolutional neural network. Remote Sensing, 9(1), 67. https://doi.org/10.3390/rs9010067
14. Ienco, D., Gaetano, R., Dupaquier, C., & Maurel, P. (2017). Land cover classification via multitemporal spatial data by deep recurrent neural networks. IEEE Geoscience and Remote Sensing Letters, 14(10), 1685–1689.
15. Chen, Y., Lin, Z., Zhao, X., Wang, G., & Gu, Y. (2014). Deep learning-based classification of hyperspectral data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7(6), 2094–2107.
16. Hu, F., Xia, G. S., Hu, J., & Zhang, L. (2015). Transferring deep convolutional neural networks for the scene classification of high-resolution remote sensing imagery. Remote Sensing, 7(11), 14680–14707. https://doi.org/10.3390/rs71114680
17. Roy, S. K., Krishna, G., & Bandyopadhyay, S. (2021). Lightweight CNN for remote sensing image classification using ensemble learning. Remote Sensing Applications: Society and Environment, 22, 100487. https://doi.org/10.1016/j.rsase.2021.100487
18. Maggiori, E., Tarabalka, Y., Charpiat, G., & Alliez, P. (2017). Convolutional neural networks for large-scale remote sensing image classification. IEEE Transactions on Geoscience and Remote Sensing, 55(2), 645–657.
19. Ball, J. E., Anderson, D. T., & Chan, C. S. (2017). Comprehensive survey of deep learning in remote sensing: Theories, tools, and challenges for the community. Journal of Applied Remote Sensing, 11(4), 042609. https://doi.org/10.1117/1.JRS.11.042609
20. Bulletin of the Karaganda University. (2025). Physics series. Karaganda: Karaganda State University. https://phys.ksu.kz/