Neutrosophic c-means and pre-trained AlexNet FC8 deep learning for classification of carbon emissions levels in Forest Biomass

Sameh H. Basha, Mohammed Mostafa Ahmed, Noha MM. Abdelnapi, Rania Ahmed, Ashraf Darwish, Aboul Ella Hassanien,
Neutrosophic c-means and pre-trained AlexNet FC8 deep learning for classification of carbon emissions levels in Forest Biomass,
Applied Soft Computing,
Volume 183,
2025,
113613,
ISSN 1568-4946,
https://doi.org/10.1016/j.asoc.2025.113613.
(https://www.sciencedirect.com/science/article/pii/S156849462500924X)
Abstract: This paper proposes a novel hybrid model that integrates Neutrosophic c-Means (NCMs) and Convolutional Neural Networks (CNNs) to effectively classify the emission levels derived from Sentinel-2 images of forest biomass. Sentinel-2 data is primarily optimized for European and African regions, which may not capture the full variability of global forest ecosystems. So, Landsat-8 and Hyperion hyperspectral image data sets were implemented to provide additional spectral and temporal coverage across other regions. The model categorizes segmented regions into six distinct emission levels including no carbon, low, moderate, high, and very high. This paper presents a novel approach to image classification by combining Neutrosophic c-Means clustering with the FC8 architecture of AlexNet, a deep learning model. This innovative integration demonstrates superior performance compared to other combinations, such as k-means clustering with AlexNet and Neutrosophic c-Means clustering with VGGNet. The proposed method improves classification accuracy while significantly enhancing robustness against uncertainty inherent in multispectral imagery. The proposed model has four key phases which are data augmentation, segmenting carbon-related features, feature extraction, and classification. In the initial phase, Generative Adversarial Networks (GANs) are employed for data augmentation to enhance the robustness of the dataset. In the second phase, NCMs is utilized to segment carbon-related features in Sentinel-2 images accurately. In the third phase, feature extraction is carried out using a pre-trained AlexNet FC8 deep learning architecture from which significant features were selected. The last phase is classification, where several machine learning algorithms were used to classify carbon emission levels in forest biomass. The complexity time, inference time, and statistical validation via paired t-tests have been calculated, and the results indicate that the proposed model markedly enhances classification accuracy with the AlexNet FC8 deep learning architecture, demonstrating better performance when paired with the Random Forest (RF) machine learning classifier. In addition, the results validate the model's efficacy, achieving an average accuracy rate of 97.8 %. It further indicates that the proposed model exhibits high specificity, sensitivity, and a substantial Youden Index, affirming its potential for real-time and automated detection of carbon emission levels.
Keywords: Carbon Emissions; forest biomass; Neutrosophic Sets; Neutrosophic C-Means; Convolutional Neural Networks; AlexNet