Weather Image Classification Using a Modified CNN with EfficientNetB3

Currently, more and more focus is put on the application of artificial intelligence and machine learning for environmental monitoring and managing natural disasters. These works emphasize the crucial role of data-oriented methods in not only enhancing the accuracy of predictions and the quality of decision-making but also in dealing with real-world complexities effectively. Tang et al. [1] put forward a machine learning-guided system for predicting medium- and long-term precipitation using data augmentation techniques. Their method showed that supplementing datasets can greatly enhance prediction accuracy. Though, it is mainly about the rainfall forecasting part and its direct use for weather image classification is limited.

Albahri et al. [2] conducted a systematic review of trustworthy artificial intelligence applications in natural disasters. Their study emphasized the importance of reliability, transparency, and robustness in AI systems used for disaster management. Although the review provides valuable insights into AI safety and trustworthiness, it does not specifically address weather image classification or intelligent learning architectures for visual data analysis. Similarly, Al Shafian and Hu [3] reviewed the integration of machine learning and remote sensing for post-disaster building damage assessment. Their work highlights the effectiveness of combining ML with remote sensing data for disaster analysis. However, the focus is primarily on structural damage detection rather than weather image classification, and it does not explore CNN-based intelligent learning models in detail.

Islam et al. [4] introduced a domain-adaptive machine learning approach for solar power prediction using remote sensing data. Their model showed an amazing ability to adapt to different locations without using source data. Though, the study did not consider weather classification from images.

Islam et al. [5] developed a location-agnostic Neural learning model for rainfall prediction. Their method aims at enhancing geographical adaptability in different regions without the necessity of local training data. Despite the fact that the model exhibits excellent performance for precipitation forecasting, it is mainly concentrated on numerical prediction and does not deal with image-based weather classification issues.

Rahman et al. [6] performed a comparative analysis of various machine learning models for weather forecasting. They found that the success of machine learning algorithms is mostly influenced by the nature of the dataset and the features chosen. Nevertheless, the research points out that conventional ML methods are not only incapable of dealing with complicated nonlinear weather patterns, but they are also less robust compared to computational learning approaches.

Singh et al. [7] proposed a computational learning approach for the classification of biomedical signals and proved the potential of CNN-based architectures for feature extraction from complicated data. As a result to this study, it has been scientifically demonstrated that AI-driven methods are extremely capable of identifying very complex patterns, even though this analysis was not focused on weather-related problems., which is a key component of image-based weather classification. Papadimitriou et al. [8] created a CNN-based model for classifying weather images, which led to a high accuracy by using very deep convolutional features. Nevertheless, the accuracy of their model is largely reliant on the quality of the dataset, and it does not consider efficiency optimization for real-time applications. Mittal and Sangwan [9] presented a deep neural network system for classifying weather images on a large scale. Although they managed to design a highly scalable and accurate method, this was also reliant on the availability of huge datasets and expensive computational resources. Gad and Hosahalli [10] compared different machine learning algorithms on weather data sets and found that although traditional ML methods may deliver acceptable results, they are less capable than intelligent learning methods when it comes to recognizing complex patterns. Jaseena and Kovoor [11] provided a comprehensive survey of intelligent weather forecasting models, highlighting that although many techniques achieve high predictive accuracy, most lack stability analysis and real-world robustness evaluation. Shrivastava et al. [12], Nagaraj and Kumar [13], and Lee et al. [14] explored computational learning models such as DNN, CNN-LSTM, and multi-scale architectures for temperature prediction. These studies have confirmed the effectiveness of using artificial intelligence-powered methods to grasp tricky and nonlinear weather patterns., but they also revealed challenges such as high computational cost and limited generalization across different regions. Shen et al. [15] proposed a multi-scale CNN-LSTM-attention model that improved prediction accuracy significantly. However, the complexity of hybrid architectures makes them less suitable for lightweight or real-time systems. Gupta and Goel [16] utilized deep reinforcement learning for weather image classification by employing pre-trained models, and managed to achieve a high level of accuracy. Similarly, Li and Luo [17] presented a ViT-based technique coupled with attention mechanisms, which further increased the classification performance. These tests demonstrate the effectiveness of techniques based on artificial intelligence in classifying complex and non-linear weather patterns. Moreover, Salim et al. [18], Ding et al. [19], and Riad et al. [20] have been working on CNN improvements with kernel optimization, stride learning, and structural design modifications, among other topics. In essence, these research works reveal the significance of CNN architectural optimization in enhancing feature extraction and efficiency, an aspect that is indeed helping to refine weather image classification models Classification of weather and their forecasting have gained more methods recently. Exact features of those methods, i.e., both pros and cons, can be reviewed by small tabulation in Table 1. Besides that, it serves as nice visual support for placing current methods on a map clearly. The comparison covers the area, the problem tackled, the main findings, the advantages, and the disadvantages of the methods.

Table 1. Comparative analysis of existing methods for weather image classification and forecasting

This comparison highlights that existing methods suffer from a trade-off between accuracy, computational cost, and model complexity, motivating the need for more efficient architectures such as EfficientNetB. According to some latest investigations, architecturally advanced networks like Vision Transformers (ViT), ResNet, and DenseNet are capable of delivering excellent accuracy in classifying weather images. Nevertheless, these models usually consume a large amount of computation and demand extensive training datasets. EfficientNetB3, on the other hand, offers a better balance between accuracy and efficiency. It matches the performance level of other models; however, it has far fewer parameters and requires less computation. Therefore, it is a more appropriate option for real-time applications and resource-limited environments.

4.1 Convolutional Neural Network layer architecture

Figure 1 illustrates the architecture of a CNN. It primarily comprises several types of layers, including convolutional layers, pooling layers, and fully connected layers, which are arranged sequentially as depicted in the figure.

Figure 1. Architecture of Convolutional Neural Network (CNN)

4.2 Convolutional layer operation

When a computer processes an image, it essentially interprets a series of numbers organized in matrices. The smallest component of an image is called a pixel, and each pixel is assigned a numerical value that indicates its brightness level [8]. Figure 2 illustrates the pixels of an image, as explained by Adam Geitgey on Medium.

Figure 2. Image pixels

The kernel filter in convolutional layers plays a crucial role in enabling a network to identify and extract the most significant features from a given image [7]. As the filter traverses the entire image, it calculates the dot product between its elements and the corresponding image pixels at each position. This process generates the activation map, which a CNN utilizes to learn image features. The convolution operation is mathematically represented by Eq. (1), while Figure 3 illustrates the convolutional operation and the resulting activation map.

$\begin{gathered}\text {Activation Map}=\text {Input × Filter} \\ \operatorname{Activation} \operatorname{Map}(i, j) =\sum \_p \sum \_q \text {Input}(i-p, j-q) \cdot \operatorname{Filter}(p, q)\end{gathered}$                (1)

Figure 3. Convolutional operation

A convolutional layer is defined by three primary parameters: the kernel size, stride length, and padding [9]. The kernel size determines the dimensions of the filter applied to the input feature map. The stride length specifies the distance the filter moves with each step as it scans the input image, affecting both the size and resolution of the resulting output feature map [10].

Padding adds zero values around the borders of the input feature map, ensuring the output feature map retains essential image details while maintaining consistent dimensions [11].

4.3 The pooling layer

The primary purpose of incorporating a pooling layer after a convolutional layer is to down sample the feature maps, thereby reducing computational complexity. Specifically, max-pooling is employed to retain significant features by selecting the maximum value within each localized region. Additionally, this process helps mitigate the risk of overfitting, making the model more robust. An illustration of max-pooling can be found in Figure 4 [12, 13].

Figure 4. Pooling layer

4.4 Fully connected layer

The last part of the CNN design is the fully connected layer. That layer shapes to 1D vector the feature maps extracted from the previous layers and it also performs the classification task.

4.5 ReLU activation function

The most efficient and straightforward way to improve model training is by utilizing the ReLU activation function. ReLU is by far the most popular activation function for CNNs, largely due to its simplicity and speed of computation. Simply, it retains positive values and changes negative values to zero. So, training speeds up and the quality of outcomes is most of the time better too. ReLU operates by preserving positive values as they are, while converting all negative values to zero, making it particularly favorable for handling positive inputs. Moreover, it not only minimizes computational complexity but also expedites the overall training process of the model.

This part of the study reveals the test results from the proposed technique. The collected data are examined and the explanation of findings is aided by graphs and various performance metrics. The training and validation loss curves are shown in Figure 10. Here the red line is for training loss and green line is the validation loss. The training and validation accuracy curves are provided in Figure 11. The red line is for training accuracy and the green line is for validation accuracy. The confusion matrix in Figure 12 evaluates the proposed model comprehensively across the eleven weather classes: Dew, Fog/Smog Frost Glaze, Hail Lightning Rain, Rainbow Rime Sandstorm, and Snow. The diagonal elements of the confusion matrix stand for correctly classified weather conditions, revealing generally strong performance, with particularly high accuracy for classes like dew (70 correct out of 71), fog smog (60 perfect classifications), and rime (71 correct out of 72). However, the off-diagonal entries highlight areas where the modified CNN architecture exhibits confusion. For instance, frost is occasionally misclassified as glaze and more frequently as rime, suggesting potential similarities between these conditions. Similarly, rain instances are sometimes mistaken for dew or snow, and sandstorm predictions show some overlap with snow and rime. The model that was proposed showed high classification performance for the different categories of weather. Still, a few classes had virtually indistinguishable features which in some cases led to misclassification. Enhancements in distinguishing between these visually similar weather conditions will be the focus of the future work.

Figure 10. Training and validation loss

Figure 11. Training and validation accuracy

Figure 12. Confusion matrix

Table 4 is a per-class report that shows the Precision, Recall, F1-score, and Support for each category. Overall, the model demonstrates strong performance across most classes, with several achieving perfect or near-perfect scores. Notably, hail, lightning, and rain exhibit exceptional results with 1.00 Precision, Recall, and F1-score, indicating that the model flawlessly identifies these classes when they are present and rarely misclassifies other phenomena as these. Furthermore, dew also performs remarkably well with 0.99 for all three metrics, suggesting a highly reliable classification. While most other classes maintain high F1-scores above 0.90, frost, glaze, rain, rime, and snow show slightly lower, though still respectable, performance, indicating some room for minor improvement in distinguishing these particular categories.

Table 3 presents the loss and accuracy performance across various stages of the newly modified CNN model. The network demonstrates a remarkable training accuracy of 99.84% alongside a minimal loss value of 0.1814, highlighting its strong learning capability. Furthermore, the test and validation accuracies have improved to 93.45% and 93.75%, respectively, indicating that the proposed model is both robust and well-suited for generalizing to unseen data, ensuring reliable performance.

Table 3. Loss and accuracy across training, testing, and validation sets

Table 4. Represents precision, recall, F1-score and support

Table 4 provides a detailed breakdown of how effectively various weather categories were classified using metrics such as precision, recall, F1-score, and support. In analyzing the results, the stationary model demonstrated strong performance, particularly in accurately classifying categories like lightning and rainbow, both achieving a perfect precision score of 1.00. Meanwhile, Fog/Smog and glaze received a precision score of 0.91, which, although slightly lower, still reflects a very high level of accuracy.

Table 5 offers an aggregated view of the model's performance, presenting overall Accuracy, Macro Average (avg), and Weighted Average (avg) for the entire dataset of 550 samples. The model achieves an impressive overall Accuracy of 0.93, signifying that 93% of all predictions were correct. Both the Macro Average and Weighted Average for Precision, Recall, and F1-score are consistently high, ranging from 0.93 to 0.94. The Macro Average, which treats all classes equally, reinforces the notion of balanced performance across the various categories, while the Weighted Average, which considers the support (number of instances) for each class, further confirms the robust overall effectiveness of the proposed model. These strong aggregate metrics, combined with the detailed per-class scores, suggest that the model is well-suited for its classification task, showing both high overall correctness and commendable performance across individual categories.

Table 5. Accuracy, macro average and weighted average