Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 281 https://internationalpubls.com Synergizing Remote Sensing, Geospatial Intelligence, Applied Nonlinear Analysis, and AI for Sustainable Environmental Monitoring Dr.N.Shanmugapriya1, Ali Bostani2, Ali Nabavi3, Dr.D.Sasikala4, Dr.T.Elangovan5, Kodirova Surayyo Adilovna6 1Assistant Professor& Head, Department of Information Technology, Erode Arts and Science College (Autonomous) Erode, Tamilnadu,India. Email: sweetypriya26@gmail.com 2Assistant Professor, College of Engineering and Applied Sciences, American University of Kuwait, Salmiya, Kuwait. Eamil: abostani@auk.edu.kw 3Faculty of Engineering, Mathematics Division, Kuwait College of Science and Technology (KCST), Doha Area, 7th Ring Road, Kuwait. Email: a.nabavi@kcst.edu.kw 4Assistant Professor, Department of Computer Science, Sri Vasavi College, Erode-638316. Email: sasikalad@gmail.com 5Assistant Professor, Department of Computer Science, Erode Arts And Science College, Erode, Tamilnadu, India Email elangobrave@gmail.com 6Professor, the Department of "Architectural Design", Tashkent University of Architecture and Construction, Tashkent, Uzbekistan. Email: kadirovasurayyo32@mail.com Article History: Received: 07-06-2024 Revised: 09-07-2024 Accepted: 26-07-2024 Abstract: The incorporation of Remote Sensing, Geospatial Intelligence (GEOINT), and Artificial Intelligence (AI) for land cover classification facilitates the efficient gathering of data, sophisticated spatial analysis, and the development of prediction models. This collaborative method improves the precision and promptness of environmental monitoring, bolstering sustainable resource management and proactive decision-making. The study used an advanced methodology involving a Modified VGG16 model, achieving an outstanding accuracy rate of 97.34%. This approach outperforms traditional algorithms, showcasing its efficacy in precisely classifying land cover categories. The utilization of remote sensing technology enables the effective gathering of data, while GEOINT enhances the spatial analysis capabilities using modern techniques. The AI-powered Modified VGG16 model has exceptional performance in predictive modeling, allowing for the implementation of proactive management measures. The abstract highlights the significant and revolutionary effects of this comprehensive method on environmental monitoring, providing unparalleled capacities for data analysis and decision-making. The findings highlight the importance of cooperation between researchers, policymakers, and industry stakeholders to fully utilize the capabilities of these technologies and tackle obstacles in sustainable environmental management. Keywords: Land cover, remote-sensing, environment, monitoring, deep learning, accuracy, and gradient loss. mailto:sweetypriya26@gmail.com mailto:elangobrave@gmail.com Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 282 https://internationalpubls.com 1. Introduction Sustainable environmental monitoring is a fundamental aspect of modern environmental management. Its goal is to achieve a balance between the use of resources and the preservation of ecosystems, ensuring the well-being of future generations. The combination of remote sensing, geospatial intelligence (GEOINT), and artificial intelligence (AI) forms a revolutionary approach that greatly improves the effectiveness and influence of environmental monitoring efforts [1]. The combination of these technologies allows for numerous advantages, including the ability to collect data efficiently through remote sensing technologies, perform advanced spatial analysis using GEOINT, and utilize predictive modeling skills provided by AI. This comprehensive approach promotes a complete comprehension of environmental dynamics, facilitating well-informed decision-making and proactive management techniques that are essential for sustainable development [2]. Remote sensing technology, such as satellites and drones, are crucial for the effective gathering of environmental data. Satellites, outfitted with a variety of sensors, gather data on global land cover changes, vegetation health, and atmospheric conditions. Drones, due to their nimbleness and adaptability, provide excellent spatial resolution, enabling precise data gathering in regions with intricate topography or specialized monitoring requirements. The use of artificial intelligence optimizes the handling and examination of extensive datasets, guaranteeing prompt and precise observations regarding environmental conditions [3]. This increased efficiency not only improves the amount and quality of information accessible, but also allows for immediate monitoring, a critical factor in dealing with ever-changing environmental issues [4]. GEOINT plays a crucial role in the integrated approach by offering sophisticated geographical analytic capabilities. This process entails the incorporation of diverse geospatial resources, including maps, satellite imagery, and terrain models, to extract significant observations regarding spatial connections and patterns [5]. Spatial analysis enables a more profound comprehension of environmental dynamics by discovering trends and connections that may be disregarded by conventional methods. Decision support systems, utilizing GEOINT, enable stakeholders to make well-informed decisions by presenting intricate spatial data in a clear and understandable way. The utilization of spatial intelligence is crucial in tackling challenges such as urbanization, the effects of climate change, and the management of natural resources. This contributes to the implementation of efficient and sustainable environmental monitoring. The AI's predictive modeling capabilities are another important aspect of the integrated strategy [6]. Machine learning algorithms, especially those employing deep learning methodologies, have the capability to examine past data in order to detect and comprehend patterns and correlations. Subsequently, these algorithms have the capability to anticipate forthcoming alterations in the environment, providing valuable discernment for proactive administration. Within the realm of sustainable environmental monitoring, prediction models powered by artificial intelligence can anticipate patterns pertaining to deforestation, biodiversity depletion, and alterations in water quality. The ability to anticipate future events enables the development of proactive measures and flexible strategies, which are crucial for reducing the effects of environmental changes and promoting long- term sustainability [7]. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 283 https://internationalpubls.com The Hybrid Feature Enhancement Network is an example of a recent breakthrough that demonstrates the ongoing development of this integrated method. This network tackles issues pertaining to feature ambiguity and the detection of minute characteristics in the classification of remote sensing images. The DUA-Net, a Combined Convolutional Neural Network, exhibits exceptional precision in intricate urban land-use classification, hence illustrating the tangible influence of AI in environmental surveillance. Frameworks such as Atrous Spatial Pyramid Pooling (ASPP) in RAANet (Residual ASPP with Attention Net) help improve the accuracy of semantic segmentation. Nevertheless, the incorporation of remote sensing, geospatial intelligence (GEOINT), and artificial intelligence (AI) encounters certain difficulties. Data integration presents a substantial challenge due to the wide range of sources, formats, and resolutions. To achieve a harmonious integration of this diverse data, it is necessary to implement standardized protocols and interoperability methods. These measures will ensure smooth collaboration and effective communication among all parties involved. Ensuring the adaptation of AI algorithms to various environmental situations is a significant challenge in terms of their resilience. The dependability and precision of AI models are contingent upon thorough training using varied datasets, and continuous improvement is imperative to tackle changing environmental dynamics. The integrated method is significantly influenced by ethical considerations. The implementation of AI in environmental monitoring gives rise to privacy problems pertaining to the collecting and utilization of data. It is crucial to find a middle ground between using the capabilities of artificial intelligence to gain valuable insights and upholding the rights to privacy of individuals and communities. Establishing responsible data governance and adopting transparent methods are essential in order to foster confidence among stakeholders and guarantee the ethical utilization of technology for sustainable environmental monitoring. The potential uses of this integrated strategy are many and significant. The integration of remote sensing and artificial intelligence in forest monitoring enables the identification of deforestation, enabling prompt intervention and conservation initiatives. The integration of remote sensing technologies with AI enables water resource management to effectively monitor water bodies, evaluate water quality, and identify pollution. GEOINT and AI play a crucial role in urban planning by examining urban growth trends, overseeing infrastructure development, and addressing the consequences of climate change on cities, hence facilitating informed decision-making [8]. These applications highlight the adaptability and importance of the integrated approach in several environmental fields. As these technologies continue to advance, collaborative efforts among researchers, policymakers, and industry stakeholders become increasingly crucial. Interdisciplinary collaboration can address challenges, foster innovation, and maximize the benefits of this integrated approach. Governments and international organizations play a pivotal role in creating frameworks and standards that promote the responsible use of technology for sustainable environmental monitoring. Investing in research and development, promoting data sharing initiatives, and ensuring inclusive participation are essential components of a collaborative approach that leverages the full potential of remote sensing, GEOINT, and AI. Ultimately, the combination of remote sensing, GEOINT (Geospatial Intelligence), and AI Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 284 https://internationalpubls.com (Artificial Intelligence) is a powerful and game-changing factor in the field of sustainable environmental monitoring [9]. 2. Related Works The utilization of deep learning has become a significant and influential factor in various scientific fields, producing exceptional outcomes in recent times. Deep learning is notable in the field of remote sensing image categorization because it takes into account both the spatial and spectral features of target objects in a comprehensive manner. This methodology is commonly used for classifying and extracting target objects from remote sensing images that have both high spectral and spatial resolutions. Deep learning models, such as convolutional neural networks (CNNs), deep belief networks, and stacked autoencoder networks, bring about a significant change in the way features are learned. These models have the ability to autonomously learn features from data, hence reducing the requirement for researchers to manually create features in advance. The implementation of these models has been a gradual yet substantial process, particularly in the categorization of hyperspectral remote sensing images [10–13]. A new development in remote sensing classification is the Hybrid Feature Enhancement Network [14], which aims to tackle the difficulties associated with feature confusion and the detection of small-scale features. The objective of this network is to increase the clarity of similar characteristics and better the ability to identify small-scale features, hence improving the overall reliability of remote sensing picture categorization. Moreover, the DUA-Net, which is a Combined Convolutional Neural Network, has exhibited exceptional precision in the intricate task of urban land-use classification. This has been confirmed through its evaluation on publicly available datasets [15]. In addition, advanced frameworks like the Atrous Spatial Pyramid Pooling (ASPP) in the RAANet (Residual ASPP with Attention Net) have been developed to enhance the precision of semantic segmentation in remote sensing images [16]. A remarkable example involved the utilization of a seven-layer Convolutional Neural Network (CNN) model, which shown significant enhancements in classification accuracy. It achieved an impressive overall accuracy (OA) of 94.68% and a Kappa value of 0.9351 [17]. Deep learning is not limited to the use of high-resolution data. Researchers have also applied it to remote sensing photos that have a middling spatial resolution. For example, Karra K et al. highlighted the constraints of existing global Land Use and Land Cover (LULC) data caused by their poor resolution. They used advanced deep learning methods to create higher-resolution global LULC products with a spatial resolution of 10 meters [18]. In addition, scientists have investigated the categorization of Landsat pictures with resolutions of 15 meters and 30 meters using deep learning techniques [19–20]. These efforts highlight the flexibility and ability of deep learning methods to work effectively with different levels of detail in remote sensing. This represents a big step forward in improving the accuracy and usefulness of picture categorization approaches. The research gap in remote sensing image classification pertains to the difficulties of traditional models in capturing intricate spatial and spectral characteristics that are crucial for precise land cover classification. The redesigned VGG16 architecture is specifically designed to optimize the performance of VGG16 for the intricate characteristics of remote sensing imagery. To boost feature extraction, the modifications involve fine-tuning the VGG16 model by adjusting layer configurations Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 285 https://internationalpubls.com and kernel sizes, taking use of its inherent depth. The modified VGG16 model is specifically designed to address issues such as feature confusion and the detection of small-scale characteristics. Its purpose is to correct the limitations seen in standard models. The updated model achieves a compromise between model complexity and effective feature extraction by adapting the VGG16 architecture to the specific characteristics of remote sensing data. This enhancement addresses the existing research void by offering a more resilient approach for land cover categorization problems. The improved VGG16 has been empirically evaluated on several datasets, consistently demonstrating its higher performance. This highlights its capability to overcome restrictions present in current models and push the boundaries of remote sensing picture categorization. 3. Proposed Methodology The classification of imagery from satellites using the VGG16 model requires a series of steps. Initially, the dataset containing satellite images is imported and subjected to preprocessing procedures to ensure uniformity in dimensions and pixel values. The pre-existing VGG16 model, acquired via a deep learning framework such as Keras or TensorFlow, is subsequently employed as a tool for extracting features. The VGG16 model utilizes the convolutional layers to extract high-level characteristics from the satellite pictures. Afterwards, these characteristics are compressed, and a specialized categorization component is appended to create a fresh model. The model is compiled using an optimizer, an appropriate loss function, and an evaluation metric. The training phase encompasses the process of inputting labeled satellite photos into the model, modifying the internal parameters using the backpropagation technique, and refining the model to enhance its performance. Following the training process, the model is assessed for its correctness by evaluating it on a distinct test dataset. Ultimately, the well-developed model can be utilized to generate forecasts on novel satellite images, offering valuable understanding into the categorization of distinct geographical attributes or land cover classifications in satellite imaging. By fine-tuning hyperparameters and model architecture, it is possible to customize the VGG16-based technique to meet individual requirements and datasets. The architectural configuration entails the utilization of the VGG16 model for the classification of satellite images. The succeeding phases provide a detailed account of this process. Initially, the pre- trained VGG16 model is loaded, excluding its top layers, which enables it to function as a powerful feature extractor. The model is derived from deep learning frameworks such as Keras or TensorFlow, offering a strong basis for feature extraction tasks. The provided code snippet illustrates the process of loading the VGG16 model and customizing it using user-defined input dimensions. Subsequently, the feature extraction procedure entails feeding preprocessed satellite pictures through the convolutional layers of the VGG16 model. This stage captures crucial high-level features necessary for efficient image classification. The compressed characteristics are subsequently inputted into a specialized classification component consisting of a compression layer and a densely interconnected layer with a softmax activation function. The model is created by combining the VGG16 convolutional base with a bespoke classification head. The model is compiled using the Adam optimizer, categorical crossentropy as the loss function, and accuracy as the evaluation metric. In the training phase, the model is adjusted to the satellite image Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 286 https://internationalpubls.com dataset by utilizing backpropagation to optimize internal parameters over a defined number of epochs. Following the training process, the model is assessed on an independent test set to measure its performance metrics. Ultimately, the well-trained model has the ability to generate predictions on novel satellite photos, providing a flexible tool for tasks involving the categorization of images in the field of remote sensing and analysis of satellite data. The modified VGG16 classification on the Million-AID dataset for remote sensing (RS) image scene classification, we can adapt the standard VGG16 architecture by introducing modifications to the original model. Below is a simplified representation of the modified VGG16 model. Let X be the input image data, Y be the corresponding label vector representing the target classes for the images, W(l) and b(l) be the weights and biases of layer l, f(l) be the activation function of layer l, and M be the modified VGG16 model. The mathematical model for the modified VGG16 classification can be represented as follows: Input Layer: The input layer of the VGG16-based satellite image classification model receives the satellite image data, represented as X. The first input consists of images, which are usually represented as three-dimensional arrays. These arrays have dimensions that correspond to height, breadth, and channels (such as RGB channels). The input layer serves as the initial stage of the neural network architecture, where the unprocessed pixel values of satellite images are introduced to the following layers to extract features and perform classification. 𝐴(0) = 𝑋 The Convolutional Blocks of the VGG16 model consist of a sequence of convolutional layers arranged in a hierarchical structure. Every block comprises several convolutional layers, which are subsequently followed by max-pooling layers. The convolutional layers extract hierarchical information at various scales, while the max-pooling layers decrease spatial dimensions, hence enhancing translation invariance. The architecture utilizes 3x3 convolutions with narrow receptive fields to acquire intricate patterns and correlations within the satellite photos. The convolutional blocks are essential for automatically extracting advanced features, like as edges, textures, and forms, from the input satellite images. For the convolutional block l, layer nl, and each layer i is 𝑍(𝑙,𝑖) = π‘Š(𝑙.𝑖) Γ— 𝐴(𝑙,π‘–βˆ’1) + 𝑏(𝑙,𝑖) 𝐴(𝑙,𝑖) = 𝑓(𝑙,𝑖)(𝑍(𝑙,𝑖)) where 𝐴(𝑙) is the output of the final convolution layer in the lth block. Flatten Layer: The Flatten Layer acts as an intermediary between the convolutional and fully linked layers. Following the convolutional blocks, the multi-dimensional output from the final convolutional layer is transformed into a one-dimensional vector, resulting in flattened features. This technique maintains the extracted spatial characteristics while preparing the data for input into the fully connected layers. The flattened features, referred to as Aflatten, act as a condensed representation of the spatial information acquired by the convolutional layers. 𝐴(π‘“π‘™π‘Žπ‘‘π‘‘π‘’π‘›) = π‘“π‘™π‘Žπ‘‘π‘‘π‘’π‘›(𝐴(𝑙)) Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 287 https://internationalpubls.com Fully Connected Layers: The Fully Connected Layers are composed of densely connected nodes that follow the flattened features. The presence of these layers allows the model to acquire intricate connections between the extracted characteristics and the desired categories. The weights W(l) and biases b(l) are modified during training to accurately represent complex patterns and relationships within the feature space. The activation function, commonly ReLU, introduces non-linearity, enabling the model to acquire intricate decision limits. The fully connected layers serve as an intermediary connecting the spatial characteristics derived from the convolutional layers to the ultimate output layer. 𝑍(𝑙) = π‘Š(𝑙) Γ— 𝐴(π‘™βˆ’1) + 𝑏(𝑙) 𝐴(𝑙) = 𝑓(𝑙)(𝑍(𝑙)) The "Compute Loss" stage entails assessing the model's performance by calculating the disparity between its predictions and the actual target labels. The categorical crossentropy loss function is frequently used for satellite image categorization with the VGG16 model. This function computes the crossentropy, which represents the discrepancy, between the anticipated probability of different classes and the actual labels of the classes. 𝐽(𝑑) = 𝐽(𝐴(π‘œπ‘’π‘‘π‘π‘’π‘‘), π‘Œ) The objective during training is to minimize the loss, so instructing the model to provide predictions that are more precise. "Backward Propagation" refers to the procedure of adjusting the model's parameters (weights and biases) in order to minimize the calculated loss. The chain rule of calculus is utilized to compute the gradient of the loss with respect to each parameter in the model. The gradients indicate the sensitivity of the loss function to tiny changes in each parameter. Backpropagation efficiently propagates these gradients in a reverse manner across the layers of the model, enabling the model to comprehend the impact of each parameter on the total loss. The model subsequently modifies its parameters in the opposite direction as the gradients, with the objective of minimizing the loss. πœ•π½ πœ•π‘Š(𝑙) , πœ•π½ πœ•π‘(𝑙) The "Update Parameters" stage involves adjusting the weights and biases of the model using the gradients calculated during backward propagation. The optimization algorithm, such as Adam, is employed to ascertain the magnitude and orientation of the parameter updates. The learning rate (Ξ±) determines the magnitude of these updates. The update rule can be stated as π‘Š(𝑙) = π‘Š(𝑙) βˆ’ 𝛼 πœ•π½ πœ•π‘Š(𝑙) 𝑏(𝑙) = 𝑏(𝑙) βˆ’ 𝛼 πœ•π½ πœ•π‘(𝑙) The iterative procedure facilitates the model's convergence towards optimal parameter values that minimize the loss function. Output Layer: The Output Layer is the ultimate layer that generates the model's predictions. When classifying satellite images, the output layer commonly uses the softmax activation function to calculate probabilities for each class. The number of nodes in the output layer is directly proportional Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 288 https://internationalpubls.com to the number of classes involved in the classification task. The ultimate forecast of the model is determined by the class that possesses the greatest likelihood. During the training phase, the categorical crossentropy loss function quantifies the difference between the predicted probabilities and the actual labels, hence directing the optimization process. The output layer converts the complex feature representations into probabilities for each class, making it easier to understand and classify satellite images. 𝑍(π‘œπ‘’π‘‘π‘π‘’π‘‘) = π‘Š(π‘œπ‘’π‘‘π‘π‘’π‘‘) Γ— 𝐴(π‘“π‘’π‘™π‘™π‘¦π‘π‘œπ‘›π‘›π‘’π‘π‘‘π‘’π‘‘) + 𝑏(π‘œπ‘’π‘‘π‘π‘’π‘‘) 𝐴(π‘œπ‘’π‘‘π‘π‘’π‘‘) = π‘ π‘œπ‘“π‘‘π‘šπ‘Žπ‘₯(𝑍(π‘œπ‘’π‘‘π‘π‘’π‘‘)) The "Termination" stage establishes the criteria that determine when the training process concludes. The training process often consists of iterating through the dataset many times, known as epochs, while adjusting parameters and enhancing the model's performance. Possible termination conditions may consist of a pre-established number of epochs, reaching a suitable degree of accuracy, or detecting the convergence of the loss function. Upon the fulfillment of the termination conditions, the training process reaches its conclusion, and the model is deemed to be trained. Subsequently, it can be assessed on novel data to determine its ability to generalize. The termination step guarantees the efficient training of the model while simultaneously avoiding overfitting to the training data. min 𝑀(𝑙),𝑏(𝑙) 𝐽( 𝐴(π‘œπ‘’π‘‘π‘π‘’π‘‘), π‘Œ) The modified VGG16 model can be trained using a suitable optimization algorithm, such as gradient descent, to minimize a chosen loss function. The training process involves adjusting the weights and biases (W and b) to improve the model's classification performance on the training data. 4. Result and Discussion The Million-AID dataset is designed specifically for remote sensing image scene categorization. It consists of 1500 files for green areas and water bodies, 1131 files for desert landscapes, and an additional 1500 files representing various overcast scenarios. The dataset is divided into training and testing sets in a systematic manner, with 3800 photos in the training set and 1831 images in the testing set. By utilizing Python as a flexible programming language, researchers can effortlessly obtain, preprocess, and analyze the Million-AID dataset. The broad ecosystem of Python, which includes libraries such as TensorFlow and scikit-learn, allows for the creation and training of intelligent algorithms for interpreting RS images. The dataset can be downloaded from a dedicated repository [21]. It is supported with Python-based tools that help in effectively using it. The incorporation of Python alongside Million-AID not only simplifies the research process but also motivates the RS community to collectively investigate inventive methods in large-scale and practical RS image scene classification. Researchers are urged to utilize and contribute to the dataset, promoting progress in data-driven interpretation models within the Python community. The dataset information is given in Figure 1. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 289 https://internationalpubls.com Figure 1. Dataset Count Figure 2. Feature Selection Figure 3. Classified image with respective labels Figure 4. Training Vs Validation Loss and Training Vs Validation Accuracy Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 290 https://internationalpubls.com Figure 1, titled "Dataset Count," presents a visual depiction that illustrates the distribution of classes in the satellite image dataset. This representation offers useful insights into the quantity of samples available for each category. Figure 2, titled "Feature Selection," visually represents the procedure of choosing significant features from satellite images. It utilizes charts to emphasize the relevance scores of each individual feature. The visual output in Figure 3, titled "Classified Image with Respective Labels," displays the results of satellite image classification. It uses color-coding to represent expected classes for pixels or regions. Figure 4, titled "Training Vs Validation Loss and Training Vs Validation Accuracy," utilizes line graphs to depict the performance metrics of the model throughout the training process. This assists in evaluating the convergence of the model and identifying any potential overfitting. These images provide a full overview of the dataset, feature selection process, classification output, and model training dynamics in satellite image analysis. Accuracy, precision, and recall are crucial measures for assessing the performance and dependability of classification algorithms in the field of land cover categorization. Accuracy is determined by dividing the number of correctly categorized cases by the total number of instances, and it measures the overall correctness of the model. A high level of accuracy shows the successful classification of land cover, whereas a lower level of accuracy suggests flaws in the model's predictions. Precision is a metric that evaluates the accuracy of positive predictions. It is calculated by dividing the number of accurately predicted positive cases by the total number of anticipated positive instances. It offers valuable information on the model's accuracy in detecting particular land cover categories. Recall measures the model's capacity to accurately identify and categorize all occurrences of a specific land cover category. It is determined by dividing the number of true positives by the total number of actual positive instances. The concepts of precision and recall provide subtle insights, which are especially important when the impacts of false positives or false negatives differ. Assessing these measures together gives a thorough comprehension of how well a land cover classification model achieves specified goals. π΄π‘π‘π‘’π‘Ÿπ‘Žπ‘π‘¦ = 𝑇𝑃 + 𝑇𝑁 𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 π‘ƒπ‘Ÿπ‘’π‘π‘–π‘ π‘–π‘œπ‘› = 𝑇𝑃 𝑇𝑃 + 𝐹𝑃 π‘…π‘’π‘π‘Žπ‘™π‘™ = 𝑇𝑃 + 𝑇𝑁 𝑇𝑃 + 𝐹𝑁 The performance measures namely Accuracy, Precision, and Recall is given in Table 1 and Figure 5. The performance is compared with existing techniques namely combined CNN (CCNN), 7-layer CNN, VGG16, and the proposed modified VGG16. Table 1. Comparison of Performance Algorithm Accuracy Precision Recall CCNN 86 81.2 80.67 CNN 88.23 82 82.2 VGG16 89 83.43 83 Modified VGG16 97.34 94.3 95.3 Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 291 https://internationalpubls.com Figure 5. Comparison of Performance The parameters of Accuracy, Precision, and Recall offered in the comparative analysis of land cover categorization algorithms offer useful insights into their performance. Commencing with CCNN, the algorithm exhibits a commendable accuracy of 86%, but, its precision and recall somewhat fall behind, indicating a possibility of overlooking occurrences and generating incorrect positives. CNN demonstrates an accuracy of 88.23%, achieving a more favorable equilibrium between precision and recall, which suggests enhanced accuracy in positive predictions. VGG16 significantly enhances its performance, reaching an accuracy of 89% while also improving precision and recall. The Modified VGG16 stands out as the most remarkable performer, with an amazing accuracy of 97.34%, precision of 94.3%, and recall of 95.3%. The significant enhancement underscores the efficacy of the alterations implemented to the VGG16 model, rendering it an exceptional option for land cover classification tasks. To summarize, the algorithms demonstrate a gradual improvement in performance, with Modified VGG16 emerging as the most resilient choice, providing exceptional accuracy and a balanced compromise between precision and recall. The selection of the algorithm ultimately depends on the precise needs and priorities of the land cover classification application. 5. Conclusion The combination of Remote Sensing, Geospatial Intelligence (GEOINT), and Artificial Intelligence (AI) for land cover classification is a highly effective approach to improve environmental monitoring. The implementation of a Modified VGG16 model demonstrates exceptional precision, with an accuracy rate of 97.34%, surpassing conventional approaches. By adopting a collaborative approach, the process of acquiring data, doing geographical analysis, and creating prediction models is made more efficient. This leads to improved accuracy and timeliness in monitoring activities. Potential future improvements may involve optimizing algorithms to boost flexibility in various environmental situations, tackling issues related to data integration, and guaranteeing the ethical implementation of technology. This study highlights the significant influence of this approach on the sustainable management of resources and proactive decision-making. It emphasizes the importance of continuous collaboration among stakeholders to fully utilize the capabilities of modern technologies in environmental stewardship. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 292 https://internationalpubls.com Future advancements in land cover classification entail improving artificial intelligence algorithms such as the Modified VGG16 to better handle various environmental conditions. This includes implementing standardized protocols for integrating data, enabling real-time monitoring, ensuring ethical practices in artificial intelligence, promoting collaboration across different fields of study, creating user-friendly interfaces, providing ongoing training, ensuring scalability, and raising public awareness about the importance of sustainable environmental monitoring. Reference [1] Wang, L., Wang, J., Liu, Z., Zhu, J., & Qin, F. (2022). Evaluation of a deep-learning model for multispectral remote sensing of land use and crop classification. The Crop Journal, 10(5), 1435-1451. [2] Luo, M., & Ji, S. (2022). Cross-spatiotemporal land-cover classification from VHR remote sensing images with deep learning based domain adaptation. ISPRS Journal of Photogrammetry and Remote Sensing, 191, 105-128. [3] Zhao, Y., Zhang, X., Feng, W., & Xu, J. (2022). Deep Learning Classification by ResNet-18 Based on the Real Spectral Dataset from Multispectral Remote Sensing Images. Remote Sensing, 14(19), 4883. [4] Chen, W., Ouyang, S., Tong, W., Li, X., Zheng, X., & Wang, L. (2022). GCSANet: A global context spatial attention deep learning network for remote sensing scene classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15, 1150-1162. [5] Mehmood, M., Shahzad, A., Zafar, B., Shabbir, A., & Ali, N. (2022). Remote sensing image classification: A comprehensive review and applications. Mathematical Problems in Engineering, 2022, 1-24. [6] Meraj, G., Kanga, S., Ambadkar, A., Kumar, P., Singh, S. K., Farooq, M., ... & Sahu, N. (2022). Assessing the yield of wheat using satellite remote sensing-based machine learning algorithms and simulation modeling. Remote Sensing, 14(13), 3005. [7] Goswami, A., Sharma, D., Mathuku, H., Gangadharan, S. M. P., Yadav, C. S., Sahu, S. K., ... & Imran, H. (2022). Change detection in remote sensing image data comparing algebraic and machine learning methods. Electronics, 11(3), 431. [8] Gadamsetty, S., Ch, R., Ch, A., Iwendi, C., & Gadekallu, T. R. (2022). Hash-based deep learning approach for remote sensing satellite imagery detection. Water, 14(5), 707. [9] Shafique, A., Cao, G., Khan, Z., Asad, M., & Aslam, M. (2022). Deep learning-based change detection in remote sensing images: A review. Remote Sensing, 14(4), 871. [10] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25. [11] Hinton, G. E., & Salakhutdinov, R. R. (2006). Reducing the dimensionality of data with neural networks. science, 313(5786), 504-507. [12] Liu, C. Extraction Based on Deep Learning Supported by Spectral Library: Taking Qingdao as an Example. Master’s Thesis, Shandong University of Science and Technology, Qingdao, China, 2020. [13] Feng, S.; Fan, F. Analyzing the effect of the spectral interference of mixed pixels using hyperspectral imagery. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 14, 1434–1446. [14] Wang, D.; Yang, R.; Liu, H.; He, H.; Tan, J.; Li, S.; Qiao, Y.; Tang, K.; Wang, X. HFENet: Hierarchical Feature Extraction Network for Accurate Landcover Classification. Remote Sens. 2022, 14, 4244. [15] Yu, J.; Zeng, P.; Yu, Y.; Yu, H.; Huang, L.; Zhou, D. A Combined Convolutional Neural Network for Urban Land-Use Classification with GIS Data. Remote Sens. 2022, 14, 1128. [16] Liu, R.; Tao, F.; Liu, X.; Na, J.; Leng, H.; Wu, J.; Zhou, T. RAANet: A Residual ASPP with Attention Framework for Semantic Segmentation of High-Resolution Remote Sensing Images. Remote Sens. 2022, 14, 3109. [17] Yu, J.; Du, S.; Xin, Z.; Huang, L.; Zhao, J. Application of a convolutional neural network to land use classification based on GF-2 remote sensing imagery. Arab. J. Geosci. 2021, 14, 1–14. [18] Karra, K.; Kontgis, C.; Statman-Weil, Z.; Mazzariello, J.C.; Mathis, M.; Brumby, S.P. Global land use/land cover with Sentinel 2 and deep learning. In Proceedings of the 2021 IEEE international geoscience and remote sensing symposium IGARSS, Brussels, Belgium, 11–16 July 2021; pp. 4704–4707. [19] Parekh, J.; Poortinga, A.; Bhandari, B.; Mayer, T.; Saah, D.; Chishtie, F. Automatic Detection of Impervious Surfaces from Remotely Sensed Data Using Deep Learning. Remote Sens. 2021, 13, 3166. [20] Manickam, M.T.; Rao, M.K.; Barath, K.; Vijay, S.S.; Karthi, R. Convolutional Neural Network for Land Cover Classification and Mapping Using Landsat Images, Innovations in Computer Science and Engineering; Springer: Berlin/Heidelberg, Germany, 2022; pp. 221–232. [21] https://www.kaggle.com/datasets/mahmoudreda55/satellite-image-classification/data