(PDF) Meteorology-Net: A Deep Learning Framework for Multi-class Weather Phenomenon Detection, Classification and Forecasting Us

National Academy Science Letters https://doi.org/10.1007/s40009-026-02124-8 SHORT COMMUNICATION Meteorology-Net: A Deep Learning Framework for Multi-class Weather Phenomenon Detection, Classification and Forecasting Using Remote Sensing Data Yogesh H. Bhosale1 Received: 26 March 2025 / Revised: 6 September 2025 / Accepted: 16 March 2026 © The Author(s), under exclusive licence to The National Academy of Sciences, India 2026 Abstract The recent Artificial Intelligence (AI) boom has reignited interest in utilizing effective Deep Learning (DL) techniques in various domains. There exists proof to suggest that by integrating Machine Learning (ML) based data analysis with prediction and Convolutional Neural Networks (CNN’s) within the weather detection, forecasting pipeline, and weather recommendations can be improved. This paper tried to examine whether DL techniques can complement traditional numerical weather simulations and data aggregation platforms by enhancing image-based classification and forecasting accuracy. The proposed Meteorology-Net utilizes the DL-based 4-CNNs variants (VGG19, MobileNet, ResNet152V2, InceptionV3) for feature extraction using weather images. The dataset comprised 6862 weather images across 11 classes (dew, fogsmog, frost, glaze, hail, lightning, rain, rainbow, rime, sandstorm, snow). Augmentation techniques such as rotation, scaling, and flipping were applied to balance class distribution (~ 1050 - 1100 samples per class) for weather prediction. Based on the attained results the base DCNNs reached the accepted range of accuracy. Hence to improve the performance of 11-class weather prediction the proposed Meteorology-Net integrates decision-level ensemble concatenation with a multi-task learning framework, enabling simultaneous classification of weather phenomena and forecasting of weather cues. Our proposed Meteorology-Net decision-level ensemble outperformed base DCNNs by achieving higher accuracy (0.9798 vs. 0.9773), improved precision and recall, and reduced error rates, demonstrating superior robustness and generalization for multi-class weather classification. We hope that this study will be helpful for the meteorology department of various nations in detecting and predicting real-time weather using remote sensing data. Keywords Weather forecasting · Weather detection · Remote sensing data · Machine learning · Convolutional neural networks (CNNs) · Ensemble learning · Multi-class classification · Artificial intelligence in meteorology Introduction The weather and climate science community is increasingly adopting contemporary deep learning technologies to address challenges in data analysis, mathematical simulation, and post-processing tasks related to weather forecasting. The steps of classical weather predicting include observations, evaluation, and predictions [1]. The above Yogesh H. Bhosale

1 Department of Computer Science and Engineering, CSMSS Chh. Shahu College of Engineering, Chhatrapati Sambhajinagar, Maharashtra 431011, India calculations takes into consideration several variables, including air pressure, heat, moisture, air acceleration, and cloud coverage [2, 3]. Cloud coverage is the quantity of clouds that envelop and obstruct the horizon and is sometimes referred to as cloudage. Cloud amount/cloudiness is one of these characteristics that is crucial to consider while assessing the weather condition. In classical meteorology [1], it is usual practise to use cloud covering during the observations stage by deploying workers out into the fieldwork to collect images from the surface and during the evaluation stage by using experienced staff from the meteorological agency. Later, the expert will apply human knowledge to predict the upcoming weather. However, this leads to laborious, error-prone, and extra human effort. To overcome these issues we proposed a Meteorology-Net 13 Y. H. Bhosale architecture that deals with weather detection and classification using imagery data. In 2021, weather-related disasters cost the global economy $329 billion [4]. On collision, unexpected catastrophic events create destruction. Tornadoes, flooding, heat waves, cold waves, deserts, and wildfires are a few examples of severe meteorological and climatic disasters. Climaterelated dangers to public health can result in significant financial damages. The economic disruption brought on by catastrophes varies geographically and has an effect on biodiversity. The destruction of technology and intellectual resources, combined with the deaths, disturbs the countries economy. About $343 billion was lost to catastrophic disasters worldwide in 2021, with floods being the cause of the greatest damage [4]. Nearly 20 natural disasters costing $1 billion or more occurred in the US in 2021. The most expensive cyclone was Cyclone Ida, a Grade 4 cyclone that made landfall just on the Louisiana coastline in August, with an estimated financial damage of $75 billion USD. The long-distance tornados can also be detected using IoT and DL techniques using remote sensing data. In recent days the importance of deep machine learning is being used in various domains such as medical imaging [5 - 7], NLP with OCR [7], automobile [8, 9], virtual assistants [10], composing music [11], etc. Hence we look for additional studies that can be majorly utilized in weather detection and classification work specifically using ground imagery data. The major contribution of this study is as follows: (1) Development of state-of-the-art tools for weather prediction and detection from imaging datasets. (2) The optimized feature extraction using classical DL variants i.e. VGG19, MobileNet, ResNet152V2, InceptionV3. (3) Eleven different weather conditions were selected in the proposed study. (4) Finding the best suitable DCNN for weather prediction by evaluating the performance of individual base DCNN. (5) The proposed Meteorology-Net utilizes the technique of ensemble fusion at decision level for improving the performance. The remaining sections are organized as follows. Literature Work section describes the state-of-the-art methods used in weather prediction. The proposed method for generating Meteorology-Net is described in Proposed Work & Methodology. Obtained results from DCNNs and ensemble models are described in Results and Discussion section. Lastly, the concluding remarks are added in Conclusion. 13 Literature Work In recent years, artificial intelligence technology has started to be used in many areas. This part includes a discussion of cutting-edge machine learning theories and how they relate to meteorological data and its relevant statistical features. Bin Zhao et. al. [12] explained the limitations of current weather recognition works for image classification purposes, and proposed a multi-task weather classification task followed by a segmentation process. For the work, two types of datasets were used as 2-class & 5-class; in parallel two more datasets have been used to verify the generality of the study. The experimentation has been performed on all these four datasets and simulated. Muruvvet Kalkan et. al. [13] aimed to classify cloud samples captured from the ground to estimate their cloudage by using DL methods. The model first acts as a base with preprocessing and output, then the design is trained using the fixed method of some method in the architecture. This study shows that among the proposed methods, the best results are calculated by the standard VGG16. The main reason for the deficiencies is insufÏcient images in the data set. Wei-Ta Chu et al. [14] contributed to building a large-scale image collection from multiple sources and providing the relationships between weather attributes and image-capturing behavior. The random forest (RF) algorithm is used with two parameters, temperature, and humidity, as a regression issue and is solved by RF-regression. Sanju Kuril et al. [15] projected a novel method of cloud detection utilizing wavelet transform and radial basis function networks. Continuous classification of weather from satellite images is an important method in many climate and environmental studies, such as climate observation and forecasting. Detection and categorization of clouds in satellite images is the most effective method of identifying individual clouds for climate analysis and forecasting. It includes two main stages of extraction and classification, feature extraction is done by Haar DWT. Noureldin Laban et al. [16] proposed a magnification technique for satellite image classification using CNNs. Satellite imagery has two features that cause significant operational problems; first, high data in satellite imagery, and second, high demands placed on CNNs. The development process depends on a good selection of suitable images. According to this measure, high distribution and low computational cost are obtained. They has processed three datasets: the WHU-RS, the UCMerced, and the Brazilian coffee scene. The implementation process improved performance over the direct use of the original array. Xunshi Yan et al. [17] presented a weather detection method based on samples collected by resident vision. They have provided three groups of features, including gradient Meteorology-Net: A Deep Learning Framework for Multi-class Weather Phenomenon Detection, Classification… Fig. 1 Before augmentation class distribution for weather dataset with class imbalance size histogram, HSV color histogram, and path description, and adopted Ada Boost-based algorithm to use the category model to perform the classification task. Tests confirm that the best of data were collected from images captured by vision. For weather descriptions, Xuelong Li et al. [18] recommend semantic segmentation of weather cues such as blue skies and white clouds. In addition, a multitasking convolutional neural network (CNN) was developed to perform weather forecasting and weather beacon segmentation tasks simultaneously. Because of the relationship between these two tasks, a semantic segmentation program that looks for weather clues can also assist to gain discrimination for the classification task, providing greater accuracy. To check the efÏciency of the suggested method, additional weather clues segmentation masks are generated from the available weather data. A new simplified model called ResNet15 has been proposed (an extension of the ResNet50) by et al. [19]. The ResNet15 has a convolutional layer (CL) and it was used to extract cloud features Then the features extracted from the previous layer are a shortcut to another four layers. Finally, the weather samples are categorized and recognized by the linked algorithm and the Softmax classifier. In addition, it is created a central road weather dataset named WeatherDataset4, which contains 4983 weather images with 4 categories and covering most of the extreme weather conditions. Xia’s study uses ResNet15 to train and test WeatherDataset4 and obtained the best results. The time series model used by Xiaowei Xu [20] proposed model is domainbased. There are challenges in how to analyze such information in the remote sensing field. In particular, deep learning techniques have been proven in remote sensing, which is often used for location classification. An advanced classification method involving neural networks (RNN) and RF has also been proposed for soil classification using publicly available satellite images for various studies using timeseries satellite images. Qasem Abu Al-Haija [21] expounded a DL-based method to identify abnormal or normal air sources for non-motorized vehicles. The proposed system used the strength of DCNNs transfer learning models (SqueezeNet, ResNet50, and EfÏcientNet) along with NVIDIA GPUs. The model was evaluated on the last two weather records, DAWN2020 and MCWRD2018. The assembled data was used to provide six types of weather: cloud, rain, snow, sand, sunlight, and sunrise. Dahmane K [22] highlighted a technique of working in real-time and the ability to classify the entire spectrum of weather and evaluate it by severity. The road can distinguish five types of weather: normal (no precipitation), hurricane, heavy rain, and light fog. The training and testing phase of the deep learning method was done using the Cerema Adverse Weather Highway database. While prior works [12, 13] used VGG16 or ResNet for weather classification, they did not incorporate ensemble fusion or decisionlevel optimization. In contrast, Meteorology-Net leverages decision-level ensemble learning to enhance robustness and accuracy. Despite advances, existing studies are limited by small or imbalanced datasets, restricted class coverage, lack of ensemble approaches, and weak integration with forecasting cues. This study addresses these gaps by proposing Meteorology-Net, which combines decision-level ensemble 13 13 Table 1 Weather dataset after preprocessing Labels Criteria Dew Sample size Actual 698 Eliminated 50 Remaining 648 Balanced class Real 648 Augmented 402 Fogsmog 851 60 791 791 259 Frost 475 30 445 445 555 Glaze 639 40 599 599 401 Hail 591 35 556 556 444 Lightning 377 27 350 350 650 Rain 526 36 490 490 510 Rainbow 232 12 220 220 780 Rime 1160 80 1080 1080 20 Sandstorm 692 45 647 647 353 Snow 621 31 590 590 410 Y. H. Bhosale Meteorology-Net: A Deep Learning Framework for Multi-class Weather Phenomenon Detection, Classification… Fig. 2 Meteorology-net overall flowchart Fig. 3 Training, validation accuracy, and loss of each DCNNs fusion with multi-task classification and forecasting using remote sensing imagery. Proposed Work & Methodology Dataset Dataset plays a crucial role in prediction, classification, and 13 Y. H. Bhosale Table 2 DCNN architecture details Table 3 Base DCNNs avg. performance for weather classification Bold values are optimal results Sr. no. 1 2 3 4 Sr. no. 1 2 3 4 5 6 7 8 9 Base model VGG19 MobileNet ResNet152V2 InceptionV3 Total params 20,024,384 3,228,864 58,331,648 21,802,784 Model/ performance metrics Accuracy (macro) Precision (macro avg.) Recall/sensitivity (TPR-Macro avg.) F1-score (macro) Cross entropy FNR (macro) FPR (macro) Specificity (TNR-macro) Overall accuracy regression. ML or DL models learn necessary features from the given preprocessed inputs. For our weather classification, we have used the Kaggles Weather Image Recognition dataset [23]. The Kaggle weather dataset is a global collection of surface and remote sensing images spanning multiple regions. The collected dataset contains 11-different weather conditions images. The data partitioning was done using 80:20% from training: testing sets respectively. Initially, the publicly downloaded dataset was skewed and does not contain an equal amount of samples per class. Hence, augmentation was applied only on the training set to prevent overlap with test data and avoid accuracy inflation. The augmentation properties are set to ‘horizontal_flip: True’ and ‘rotation_range:20’. After maintaining an equal amount of samples in each class in the train set, and remaining samples are allocated to validation subset of data. Figure 1 shows labelwise original and before augmentation used samples in the weather dataset. During preprocessing, a small portion of 5 - 12% of images was eliminated to remove corrupted or duplicate samples. The remaining dataset was obtained by subtracting the eliminated samples from the actual count. To address class imbalance, the balanced dataset was created by augmenting samples such that all classes contained 1100 images. In this process, minority classes such as Rainbow, Lightning, and Frost were heavily augmented, whereas majority classes such as Rime required only minimal augmentation as shown in Table 1. 13 Epoch (halted automatically) 19 11 9 16 Batch size 32 32 32 32 Train (loss, acc), Val (loss, acc) T: (0.6497,0.7668),V: (0.6497,0.7668) T: (0.2528,0.9086), V: (0.4324,0.8641) T: (0.3231,0.9339), V: (0.8472,0.7727) T: (0.4494,0.8464), V: (0.4494,0.8464) VGG19 MobileNet ResNet152V2 InceptionV3 0.9588 0.7865 0.7693 0.7759 3.3659 0.2307 0.0230 0.9769 0.7736 0.9773 0.8815 0.8856 0.8827 3.3623 0.1144 0.0126 0.9874 0.8751 0.9749 0.8742 0.8654 0.8669 3.3694 0.1346 0.0141 0.9859 0.8619 0.9701 0.8474 0.8424 0.8421 3.3701 0.1576 0.0168 0.9832 0.8356 Methodology The weather phenomenon image classification in Meteorology-Net consists of six steps. Beginning with the utilization of the publicly available Kaggle dataset [23], followed by preprocessing. The The deep features are extracted using four separate DCNNs. Once all the features are extracted from the DL architectures the test data samples are used to evaluate the performance using a confusion matrix. Deep learning has played a crucial role in image classification and regression. Our proposed work is twofold (Fig. 2). The first fold was implemented using individual DCNNs for feature extraction and the second fold deals with testing. Meteorology-Net utilized the 4 DL models are: VGG19, lightweight MobileNet, large ResNet152V2, and InceptionV3. The pretrained DL models had already trained large datasets called ImageNet and gained the highest performance. This is the main reason to choose these 4 DCNN models in our weather classification. The VGG19 built by the Visual Geometry Group contains 19-layers [24]. VGG19 network can classify 1000 object categories. VGG19 is a stacked connection of the Convolutional Layer (CL) followed Pooling Layer (PL). It accepts a default input size of 224 × 224 with 64-depth. The subsequent layers downsamples the features to 112 × 112, 56 × 56, 28 × 28, 14 × 14, and 7 × 7. The last layers (third and second) utilize the Fully Connected Layers (FCL) with a size of 4096. Finally, the modified last layers utilize the 11 Meteorology-Net: A Deep Learning Framework for Multi-class Weather Phenomenon Detection, Classification… Fig. 4 Confusion Matrix attained by each base DCNNs. Note: (0-10 labels indicate [‘dew’, ‘fogsmog’, ‘frost’, ‘glaze’, ‘hail’, ‘lightning’, ‘rain’, ‘rainbow’, ‘rime’, ‘sandstorm’, ‘snow’] in sequence neurons for the weather image classification. MobileNet is an open-sourced lightweight model developed by Google. MobileNet utilizes the depthwise convolution operation. The depthwise convolution rigorously reduces the parameter size to make it lightweight. It’s the first Tensorflow mobile developed by Google. MobileNet accepts 224 × 224 as the input size. The subsequent layer downsamples the convolution with depthwise+pointwise. This series of stacked were continued to 14 depth+pointwise convolutions. Finally, FCL had 1024 neurons. The last dense layer was modified for 11-neurons in our modified network. The ResNet152V2 has a total 152CLs. And its default input size 256 × 256 with 3-channels. The architecture is combined with batch_normalization, max-pooling (MP), and activation (ReLu). The Features are then donwsampled into 128, 64, 32, 16, and 8 sizes. The subsequent layer flatten (100352), dense (256), dense (1) finally softmax as an activation function to classify the results. InceptionV3 accepts 224 × 224 as the input size. The final dense layer was modified with 11-neurons using softmax. All these networks are trained with a 0.0001 learning rate. Early stopping and callback criteria were adopted to avoid model overfitting during training. All these model trainings were initially set to 45 epochs for training. But due to early halt training criteria, the VGG19 stopped at 19, MobileNet stopped at 11, ResNet152V2 stopped at 9 and InceptionV3 stopped at 16 epochs respectively. The optimizer: adam, and batch size were set to 32 during training. The accuracy loss plots are shown in Fig. 3 for the training and validation phases. And 13 Y. H. Bhosale Fig. 5 Confusion matrix attained by each ensemble DCNNs at decision levels with options of two, three and four models. Note: (0 to 10 labels indicate [‘dew’, ‘fogsmog’, ‘frost’, ‘glaze’, ‘hail’, ‘lightning’, ‘rain’, ‘rainbow’, ‘rime’, ‘sandstorm’, ‘snow’] in sequence their accuracy and loss values are also populated in Table 2. The total parameters accumulated by each DCNN are added in Table 2. The highest training accuracy of 0.9338 was attained by ResNet152v2. However, the lowest training accuracy attained by VGG19 with 0.7668. Decision-level 13 ensemble models integrate the outputs of multiple base DCNNs to enhance prediction reliability by leveraging complementary strengths of individual classifiers. While base models demonstrated strong standalone performance, the ensemble approach consistently improved accuracy, Meteorology-Net: A Deep Learning Framework for Multi-class Weather Phenomenon Detection, Classification… Fig. 5 (continued) precision, and robustness across weather classes. Decisionlevel ensemble fusion was applied, where each base DCNN outputs probabilities that are aggregated through weighted averaging to derive final predictions. The results confirm that decision-level fusion outperforms individual DCNNs, reducing misclassifications and ensuring more balanced generalization as per the results shown in Table 3. The proposed approach introduces Meteorology-Net, a multi-task DCNN framework that leverages decision-level ensemble fusion to simultaneously classify weather phenomena and forecast weather cues from remote sensing imagery. Results and Discussion Most useful practice to attain the imbalanced class prediction or classification performance using a confusion matrix (CM). The confusion matrix offers further information about a predicting algorithm’s effectiveness, as well as what categories are successfully and mistakenly projected, and what kinds of faults are being produced. The quantity of instances for which the expected label matches the actual label is represented by diagonal components. However, the incorrect labels assigned by the classifiers are represented by off-diagonal components. CM provides the TP, TN, FP, and FN as attributes. These attribute values help to measure the accuracy (Ac), precision (Pr.), recall/sensitivity (Re.) and F1score (F1), etc. performance metrics. The simulated base DCNNs for Meteorology-Nets confusion matrices are shown in Fig. 4. 13 Y. H. Bhosale Fig. 6 ROC curve attained on the weather test dataset by base DCNN’s Ac. = (TP + TN) / (TP + FP + FN + TN) (1) Pr. = TP/ (TP + FP) (2) Sensitivity/Re. = TP/ (TP + FN) (3) F1 = 2 ∗ (Re. ∗ Pr.) / (Re. + Pr.) (4) Specificity = TN/ (TN + FP) (5) Misclassifications were frequent between visually similar classes such as frost vs. snow and fog vs. smog, while distinct classes like rainbow and lightning were classified with near-perfect accuracy. Figure 5 shows the ensemble DCNN models with strong classification performance 13 across most weather categories, with a clear diagonal dominance in the confusion matrices. Classes such as rainbow, lightning, and hail achieved near-perfect recognition, reflecting high sensitivity and precision. However, certain overlaps are observed particularly between frost, snow, and glaze, which share similar visual characteristics, leading to occasional misclassifications. The four-model ensemble (VGG19 + MobileNet+ResNet152V2 + InceptionV3) demonstrates the most balanced performance, minimizing false negatives and achieving the highest overall consistency across all 11 classes. These results confirm that decision level ensemble fusion improves robustness and generalization compared to individual or two-model combinations. Table 3 shows the different efÏciencies attained by base DCCNs during weather classification. In comparison Meteorology-Net: A Deep Learning Framework for Multi-class Weather Phenomenon Detection, Classification… Fig. 7 ROC curve attained by ensemble at decision level models with two, three and four options on the weather test dataset to all 4 models, the MobileNet attained the best and most acceptable range of results. The highest 0.9793 accuracy, 0.8842 precision, 0.8856 recall, 0.8827 F1score, 3.36 crossentropy, 0.1144 FNR, 0.0126 FPR, and 0.9874 specificity by Meteorology-Net. The ROC curve is a chart that displays how well a sorting algorithm performs across all categorization levels. Two-factor percentage values are plotted on this curve: TPR and FPR. Figures 6 and 7 shows the ROC curve attained during the weather phenomenon testing phase. The ROC 13 Y. H. Bhosale Fig. 7 (continued) value > 0.5 is necessary for showing the better performance of the classifier DL network. The highest ROC values are from base DCNN’s are: 1. VGG19 attained ROC of 0.97 for lightning. 2. 0.99 ROC attained by MobileNet for rainbow and lightning respectively. 3. 1.00, 0.99, 0.097 ROC attained by ResNet152V2 for lightning, rainbow, and fogsmog respectively. 4. 1.00, 0.99, 0.95, 0.95 ROC values attained by InceptionV3 for lightening, rainbow, fogsmog, and dew respectively. In comparison to all four base DCNNs, the lowest ROC of 0.76 and 0.78 values are attained by VGG19 and InceptionV3 for the snow label. 13 The Fig. 7 shows the ROC -AUC curves of ensemble models confirm strong discriminative capability across all 11 weather classes. Models such as MobileNet+ResNet152V2 + InceptionV3 and VGG19 + MobileNet+ResNet152V2 achieved near-perfect AUC values (0.98 - 1.0) for critical classes like lightning, rainbow, hail, and fogsmog, indicating highly reliable detection. While relatively weaker performance was noted for snow (AUC 0.84 - 0.85) and frost (AUC 0.87 - 0.88), the overall AUC values remain consistently high (> 0.90) for most categories. These results highlight that decision-level ensemble fusion significantly enhances model robustness and ensures balanced classification across diverse weather phenomena. With respect to base DCNNs although ResNet152V2 achieved strong results, MobileNet offered higher efÏciency with lower computational cost, making it more practical for real-time deployment. The above Table 4 Meteorology-Net: A Deep Learning Framework for Multi-class Weather Phenomenon Detection, Classification… Table 4 Ensemble DCNNs at decision level avg. performance for weather classification (Bold values are optimal) Sr. no. Model/performance metrics VGG19 + ResNet152V2 + VGG19 InceptionV3 + VGG19 MobileNet 1 Accuracy (macro) 0.97609 0.9749 0.97384 2 Precision (macro avg.) 0.8779 0.8762 0.8692 3 Recall/sensitivity (TPR-Macro avg.) 0.8758 0.8649 0.8611 4 F1-Score (macro) 0.8759 0.8676 0.8614 5 Cross entropy 3.36229 3.36979 3.37325 6 FNR (macro) 0.12417 0.13508 0.13886 7 FPR (macro) 0.01331 0.01417 0.01471 8 Specificity (TNR-macro) 0.98685 0.98619 0.98529 9 TPR-macro 0.87583 0.86492 0.86114 10 Overall accuracy 0.8685 0.8619 0.8561 11 Error-rate 0.1314 0.1380 0.1439 12 Zero one loss 180 189 197 Sr. Model/perfor- MobileNet+InceptionV3 ResNet152V2 MobileNet+VGG19 MobileNet+ResNet152V2 no. mance metrics + InceptionV3 + ResNet152V2 + InceptionV3 1 2 3 4 5 6 7 8 9 10 11 12 Accuracy 0.97875 (macro) Precision 0.8891 (macro avg.) Recall/sensitiv- 0.8893 ity (TPRMacro avg.) F1-Score 0.8884 (macro) Cross entropy 3.36153 FNR (macro) 0.11066 FPR (macro) 0.01186 Specificity 0.98814 (TNR-macro) TPR-macro 0.88934 Overall 0.8831 accuracy Error-rate 0.1168 Zero one loss 160 MobileNet+ResNet152V2 0.97596 0.97875 0.9798 0.97835 0.8866 0.8861 0.8855 3.36145 0.11394 0.01216 0.98784 0.88606 0.8809 0.1190 163 VGG19 + MobileNet+ResNet152V2 + InceptionV3 0.97888 0.8805 0.8929 0.8991 0.8962 0.8708 0.8861 0.8934 0.8880 0.8729 0.8884 0.8942 0.8909 3.36906 0.12917 0.01358 0.98642 3.36249 0.11394 0.01194 0.98806 3.36568 0.10663 0.01137 0.98863 3.36305 0.11202 0.01188 0.98812 0.87083 0.86779 0.88606 0.8831 0.89337 0.88897 0.88798 0.8839 0.1322 181 0.1168 160 0.111 152 0.1161 159 shows the ensemble DCNN at decision level models demonstrated consistently superior performance compared to individual architectures. Among the combinations, MobileNet+ResNet152V2 and MobileNet+InceptionV3 achieved the highest macro-accuracy (0.97875 and 0.97875, respectively) with low error rates (~ 0.116). The triple ensemble (MobileNet+ResNet152V2 + InceptionV3) further improved precision (0.8991) and recall (0.8934), indicating a robust balance between sensitivity and specificity. These results confirm that ensemble fusion enhances classification reliability over single-model baselines. The above Fig. 8 shows the sample classification results by Meteorology-Net demonstrate that the model accurately predicts most weather categories such as fog/smog, frost, hail, lightning, and rain, while occasional confusion occurs between visually similar phenomena like frost and glaze. Among the base models, MobileNet achieved the highest accuracy (0.9773). However, the decision-level ensemble of MobileNet+ResNet152V2 + InceptionV3 outperformed all with 0.9798 accuracy, 0.8991 precision, and 0.8934 recall, while reducing the error rate to 0.111. Confusion matrix analysis shows frequent misclassification between visually similar classes such as frost and snow, while robust detection is achieved for distinct phenomena like rainbow and lightning. Development Environment The Meteorology-Net are trained on i7-12700, with 16GB RAM, and 2GB graphics as hardware environment. Tensorflow 2.7, Keras, Sklearn, matplotlib, pandas, numpy, and seaborn are the tools/libraries utilized to perform the simulation. 13 Y. H. Bhosale Fig. 8 Meteorology—net predicted and actual labels 13 Meteorology-Net: A Deep Learning Framework for Multi-class Weather Phenomenon Detection, Classification… Conclusion With the growing accessibility of weather forecasts on mobile devices, meteorological inaccuracies are now more visible to the public, increasing scrutiny and raising doubts about their reliability. The rate of errors will be considered as long as weather predictions are made by humans. Only current AI technology can end dependence on people. As a first step toward this goal, the study employs deep learning algorithms to classify surface-acquired cloud images as clear or overcast and to estimate their precipitation potential. The proposed approach introduces Meteorology-Net, a multi-task DCNN framework that leverages ensemble fusion at the decision level to concurrently perform multiclass weather phenomenon classification and forecast weather cues from remote sensing imagery. We used a publicly available dataset of weather image phenomena to compensate for the absence of training data. Meteorology-Net utilized eleven labels for weather classification. The base models, particularly MobileNet, achieved strong standalone performance with an accuracy of 0.9773, precision of 0.8815, and recall of 0.8856. However, Meteorology-Net (decision-level ensemble fusion) further boosted performance, with the best ensemble achieving 0.9798 accuracy, 0.8991 precision, and 0.8934 recall. Compared to the base models, ensemble learning not only improved accuracy by ~ 0.5 - 1% but also reduced error rate and FNR, demonstrating its robustness and superior generalization for multi-class weather classification. In future work, we plan to take the trafÏc-scene weather recognition and apply it to autonomous driving assistance, farming, remote location weather detection, etc. systems. Meteorology-Net demonstrates that DL-based ensembles can significantly improve multi-class weather classification and forecasting cues. Rather than replacing physical models, this framework complements traditional simulations by providing fast, accurate, and image-driven insights. Future work will focus on hybrid integration with temporal models and validation on broader real-world datasets. Declarations 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. Conflict of interest The author declares no conflict of interest. 15. References 1. 2. Mürüvvet Kalkan G, Bostancı E et al (2022) Cloudy/clear weather classification using deep learning techniques with cloud images. Comput Electr Eng 102:0045 - 7906 Manandhar S, Dev S, Lee YH, Winkler S, Meng YS (2018) Systematic Study of weather variables for rainfall detection, In: IGARSS 20182018 IEEE international geoscience and remote 16. 17. sensing symposium, Valencia, Spain, pp. 30273030. https://doi. org/10.1109/IGARSS.2018.8517667 Kondapally M, Kumar KN, Krishna Mohan C (2024) Object detection in transitional weather conditions for autonomous vehicles. In: 2024 international joint conference on neural networks (IJCNN), Yokohama, Japan, 2024, pp. 18. https://doi.org/10.1109 /IJCNN60899.2024.10651445 Aon (2022) Weather catastrophe economic losses globally from 2007 to 2021 (in billion U.S. dollars) [Graph]. In Statista. Retrieved 16 March 2023. https://www.statista.com/statistics/818 411/weather-catastrophes-causing-economic-losses-globally/ Bhosale YH, Patnaik KS (2023) Bio-medical imaging (X-ray, CT, ultrasound, ECG), genome sequences applications of deep neural network and machine learning in diagnosis, detection, classification, and segmentation of COVID-19: a Meta-analysis & systematic review. Multimed Tools Appl. https://doi.org/10.1007/s1104 2-023-15029-1 Bhosale YH, Patnaik KS (2022) Application of deep learning techniques in diagnosis of covid-19 (coronavirus): a systematic review. Neural Process Lett. https://doi.org/10.1007/s11063-02211023-0 Bhosale YH, Patnaik KS, Zanwar SR et al (2025) Thoracic-net: explainable artificial intelligence (XAI) based few shots learning feature fusion technique for multi-classifying thoracic diseases using medical imaging. Multimed Tools Appl 84:5397 - 5433. htt ps://doi.org/10.1007/s11042-024-20327-3 Jaganathan A, Mathesan K (2022) Risk prediction model and classification of various hazards in automobile industry using HAO based deep CNN. Sādhanā 47:165. https://doi.org/10.1007/ s12046-022-01928-w Jagannathan P, Rajkumar S, Frnda J, Divakarachari PB, Subramani P (2021) Moving vehicle detection and classification using gaussian mixture model and ensemble deep learning technique. Wirel Commun Mob Comput 2021:1 - 5. https://doi.org/10.1155/ 2021/5590894 Vishnu R, Krishna Prakash N (2022) Mobile application-based virtual assistant using deep learning. Soft computing and signal processing. Advances in intelligent systems and computing, vol 1340. Springer, Singapore. https://doi.org/10.1007/978-981-16-1 249-7_57 Rajan R, Sivan S, Multi-channel (2023) CNN-based rāga recognition in carnatic music using sequential aggregation strategy. Circuits Syst Signal Process. https://doi.org/10.1007/s00034-02 3-02301-w Zhao B, Hua L, Li X, Lu X, Wang Z (2019) Weather recognition via classification labels and weather-cue maps. Pattern Recogn 95:272 - 284. https://doi.org/10.1016/j.patcog.2019.06.017 Kalkan M, Bostancı GE, Güzel MS, Kalkan B, Özsarı Ş, Soysal Ö, Köse G (2022) Cloudy/clear weather classification using deep learning techniques with cloud images. Comput Electr Eng 102:108271. https://doi.org/10.1016/j.compeleceng.2022.108271 Chu W-T, Zheng X-Y, Ding D-S (2017) Camera as weather sensor: estimating weather information from single images. J Vis Commun Image Represent 46:233 - 249 ISSN 1047 - 3203. https ://doi.org/10.1016/j.jvcir.2017.04.002 Kuril S, Saini I, Saini BS (2013) Cloud classification for weather information by artificial neural network. Int J Appl Phys Math 3(1):28. https://doi.org/10.7763/IJAPM.2013.V3.167 Laban N, Abdellatif B, Ebied HM, Shedeed HA, Tolba MF (2020) Multiscale satellite image classification using deep learning approach. In: Hassanien A, Darwish A, El-Askary H (eds) machine learning and data mining in aerospace technology. Studies in computational intelligence, vol 836. Springer, Cham. https: //doi.org/10.1007/978-3-030-20212-5_9 Yan X, Luo Y, Zheng X (2009) Weather recognition based on images captured by vision system in vehicle. In: Advances in 13 Y. H. Bhosale neural networksISNN 2009: 6th International symposium on neural networks, ISNN 2009 Wuhan, China, Proceedings, Part III 6 (pp. 390398). Springer Berlin Heidelberg, 26 - 29 May 2009 18. Li X, Wang Z, Lu X (2017) A multi-task framework for weather recognition. In: Proceedings of the 25th ACM international conference on Multimedia. Oct 19 (pp. 13181326) 19. Xia J, Xuan D, Tan L, Xing L (2020) ResNet15: weather recognition on trafÏc road with deep convolutional neural network. Adv Meteorol 2020:1 - 1. https://doi.org/10.1155/2020/6972826 20. Xu X, Chen Y, Zhang J, Chen Y, Adhiyaman Manickam (2021) A novel approach for scene classification from remote sensing images using deep learning methods. Eur J Remote Sens 54(sup2):383 - 395. https://doi.org/10.1080/22797254.2020.179 0995 13 21. Al-Haija QA, Gharaibeh M, Odeh A (2022) Detection in adverse weather conditions for autonomous vehicles via deep learning. AI 3(2):303 - 317. https://doi.org/10.3390/ai3020019 22. Dahmane K, Duthon P, Bernardin F, Colomb M, Chausse F, Blanc C (2021) Weathereye-proposal of an algorithm able to classify weather conditions from trafÏc camera images. Atmosphere 12(6):717. https://doi.org/10.3390/atmos12060717 23. Dataset (2023) Weather image recognition, Kaggle, Online, Accessed 15 Feb. https://www.kaggle.com/datasets/jehanbhathe na/weather-dataset 24. ImageNet. Accessed Online. http://www.image-net.org Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional afÏliations.