Development of a Novel One-Dimensional Nested U-Net Cloud-Classification Model (1D-CloudNet)

Abstract

Cloud classification is fundamental to advancing climate research and improving weather forecasting. However, existing cloud classification models are constrained by several limitations. For instance, simple statistical methods depend heavily on prior knowledge, leading to frequent misclassifications in regions with high latitudes or complex terrains. Machine learning approaches based on two-dimensional images face challenges such as data scarcity and high annotation costs, which hinder accurate pixel-level cloud identification. Additionally, single-pixel classification methods fail to effectively exploit the spatial correlations inherent in cloud structures. In this paper, we introduce the one-dimensional nested U-Net cloud-classification model (1D-CloudNet), which was developed using Himawari-8 and CloudSat data collected over two years (2016–2017), comprising a total of 27,688 samples. This model is explicitly tailored for the analysis of one-dimensional, multi-channel images. Experimental results indicate that 1D-CloudNet achieves an overall classification accuracy of 88.19% during the day and 87.40% at night. This represents a 3–4% improvement compared to traditional models. The model demonstrates robust performance for both daytime and nighttime applications, effectively addressing the absence of nighttime data in the Himawari-8 L2 product. In the future, 1D-CloudNet is expected to support regional climate research and extreme weather monitoring. Further optimization could enhance its adaptability to complex terrains.

1. Introduction

Clouds are condensations of water, consisting of tiny liquid droplets, ice crystals, or a combination of both, suspended in the atmosphere and typically not in contact with the ground [1]. Clouds play a crucial role in the Earth system. They are not only key components of the global water cycle but also participate in biogeochemical cycles and significantly impact the global radiative energy balance [2,3]. Furthermore, the presence and variation of clouds have a substantial influence on the evolution of weather systems and changes in the climate system [4]. Different clouds have distinct radiative characteristics, which reflect the dynamic and thermal processes in the atmosphere [5]. Accurate cloud detection and classification are crucial for many research efforts, specifically in the following aspects: firstly, they can improve the effectiveness of numerical model parameterization schemes, thereby enhancing model accuracy; secondly, they can enhance the analysis of precipitation intensity, temperature, and humidity fields; and finally, they contribute to increasing the accuracy of weather forecasts, ensuring flight safety [6,7,8,9].

Cloud classification research often employs retrieval methods based on the principle that different types of clouds exhibit distinct absorption and reflection characteristics across various wavelength ranges. By analyzing radiative data, meteorological parameters such as cloud-top temperature and cloud optical thickness can be retrieved, allowing cloud types to be categorized according to their physical properties [10]. The data sources for cloud classification rely on various observation technologies and platforms, with satellite remote sensing being the most common and widely used source, encompassing both active and passive remote sensing methods. Active remote sensing systems emit signals and receive information from the reflection or scattering by clouds, offering high-resolution details of cloud structure and characteristics. This method is well suited for detailed cloud analysis within a focused observational range, though it typically covers a more limited area. In contrast, passive remote sensing utilizes natural light to observe clouds, enabling rapid acquisition of cloud distribution and dynamics over broad areas [11]. While this approach has a wider observational range, it is constrained by weather conditions and time of day and has limited penetration through thick cloud layers.

Currently, mainstream cloud classification methods fall into two major categories: simple statistical approaches and machine learning-based techniques. Yu et al. [12] proposed a unit feature space classification method (UFSCM), which identifies six types of clouds based on a feature space constructed from visible light reflectance and infrared brightness temperature. Purbantoro et al. [13] utilized a split-window algorithm (SWA) and classified clouds into nine types by selecting different thresholds for brightness temperature (BT) and brightness temperature difference (BTD) from satellite data for winter and summer seasons. With the rapid advancement of computing technology, complex machine learning methods have proven to be highly effective for cloud classification. Compared to the aforementioned simple statistical approaches, these methods generally achieve significantly higher classification accuracy. Wohlfarth et al. [14] trained a support vector machine (SVM) classifier using four channels (color + thermal) from Landsat-8 satellite data, achieving an overall classification accuracy of 95.4% for 13 categories. Wang et al. [15] combined Himawari-8 multi-band radiation information with CloudSat cloud classification products and applied a random forest (RF) algorithm, training daytime and nighttime cloud classifiers with accuracies of 88.4% and 79.1%, respectively, across a dataset of seven cloud types. Gorooh et al. [16] proposed the Deep Neural Network Cloud-Type Classification (DeepCTC) model, using 16-channel features from the GOES-16 satellite to classify eight cloud types, achieving an average accuracy of 85%. Jiang et al. [17] introduced the CLP-CNN based on U-net, incorporating channel and spatial attention mechanisms to classify FY-4A L1 data into nine categories on a per-pixel basis, achieving a classification accuracy of 76.8% compared with Himawari-8 cloud classification products. Guo et al. [18] developed machine learning-based cloud classification models using data from the Advanced Geostationary Radiation Imager (AGRI) on the Fengyun-4A satellite, along with CPR-CALIOP merged-product classification results for training and validation. The models achieved accuracies of 83.4% for daytime and 79.4% for all-time single-layer cloud-type classification, outperforming Fengyun 4A Level-2 cloud products. Shang et al. [19] developed the CARE algorithm, which combines a threshold-based method and an extra randomized tree (ERT) model, using Himawari-8 full-disk measurements to achieve all-day cloud detection, with hit rates validated against CALIPSO measurements: daytime cloudy HR 76.95% and clear HR 87.97% and nighttime cloudy HR 73.19% and clear HR 81.77%.

Although the above methods have been successfully applied to cloud classification with commendable results, several unresolved issues remain. Firstly, simple statistical methods typically require prior knowledge from experts and are prone to misclassification in special regions, such as high-latitude desert areas. Additionally, these methods require the development of new algorithms tailored to different satellites. Machine learning methods based on two-dimensional images, on the other hand, generally rely on data from passive remote sensing. Both machine learning and statistical methods often involve manual labeling of cloud types, albeit to different extents. Statistical methods typically require labeled data to validate thresholds or refine empirical rules, whereas machine learning methods rely on extensive labeled datasets for data-driven feature extraction. This dependence on large labeled datasets limits dataset size and hampers model generalization. Consequently, these models tend to achieve high classification accuracy only on specific datasets of two-dimensional cloud images, without the capability for fine-grained, pixel-level cloud classification. Finally, single-pixel-based cloud classification combines data from both active and passive remote sensing and performs automatic cloud labeling during the spatiotemporal matching process, enabling pixel-level classification results. However, this approach disregards the spatial correlation among pixels within a cloud cluster, treating each pixel as an isolated entity. A cloud cluster typically spans multiple pixels within a certain area, and the cloud types of these neighboring pixels are often identical. This spatial correlation is crucial for improving the accuracy and effectiveness of cloud classification tasks.

In this study, we introduce the one-dimensional nested U-Net cloud-classification model (1D-CloudNet), which is designed to address the specific limitations of existing cloud classification methods. Our model aims to (1) mitigate the reliance on expert prior knowledge and the need for satellite-specific algorithm development by utilizing a data-driven approach, (2) overcome the challenges of data scarcity and high annotation costs associated with two-dimensional image-based machine learning methods by leveraging spatiotemporal matching of active and passive remote sensing data, and (3) improve upon single-pixel classification methods by considering the spatial correlations among pixels within cloud clusters. While we acknowledge that these advancements do not completely resolve all limitations of the mentioned methods, they represent significant steps towards enhancing the accuracy and efficiency of cloud classification.

For developing and validating 1D-CloudNet, we utilized data from the passive Himawari-8 and the active CloudSat satellites. Himawari-8 provides two-dimensional (longitude and latitude) images in 16 bands, along with solar and satellite angles, available both day and night (except visible and near-infrared bands at night). CloudSat, due to malfunction, operates only during the day, providing 125 vertical layers of cloud type distribution along its flight track. To fully leverage Himawari-8’s wide observational range and CloudSat’s high vertical resolution, we performed spatiotemporal matching, resulting in one-dimensional images with corresponding cloud types for each pixel.

The introduction of 1D-CloudNet offers a novel perspective on cloud classification. While existing methods have made considerable progress, this model is unique in its focus on one-dimensional and multi-channel image analysis. A key innovation is its ability to exploit spatial correlations inherent in cloud structures, a feature underexplored in previous models. By linking continuous pixels from spatiotemporal matching of active and passive remote sensing data, 1D-CloudNet enhances classification accuracy and extends its operational timeframe to include both day and night. This innovation in leveraging one-dimensional spatial correlations advances cloud classification methodology.

To guide this study, the following research questions are formulated: (1) How does the performance of 1D-CloudNet in cloud classification compare to traditional methods in terms of accuracy? And (2) what is the impact of adjusting input channel combinations, training epochs, and output layers on the cloud classification accuracy of 1D-CloudNet?

The research objectives for the 1D-CloudNet model are to (1) improve overall classification accuracy by at least 3% compared to traditional models, (2) reduce misclassification rates in nighttime data, and (3) effectively utilize both active and passive remote sensing data to enhance the model’s robustness across diverse terrains and weather conditions. Achieving these objectives will contribute to more accurate and reliable cloud classification, vital for climate research and weather forecasting.

Accurate cloud classification is essential for improving weather forecasting and climate modeling precision. In forecasting, better classification can enhance predictions of precipitation, temperature, and wind patterns, impacting sectors like agriculture, aviation, and disaster management. For climate modeling, better cloud distribution understanding can refine cloud feedback mechanisms in global models, improving long-term climate predictions. Furthermore, accurate real-time cloud classification aids decision-making in weather-sensitive industries such as renewable energy, where cloud cover affects solar and wind energy production. The 1D-CloudNet’s improved classification capability will bridge theoretical advancements with practical applications.

To provide a clear structure for this study, the paper is organized as follows: Section 2 details the methodology, including the design of the 1D-CloudNet, the datasets used, and the preprocessing steps. Section 3 presents the results of the model evaluation, including accuracy metrics and comparisons with traditional methods. Section 4 discusses the practical applications of the model, with case studies and long-term analysis. Finally, Section 5 concludes the paper with a summary of the findings and suggestions for future research.

2. Model Development

2.1. Model Design

U-Net, proposed in 2015, is a deep learning network structure designed for image segmentation, widely applied in the field of medical image analysis [20]. Its primary function is to segment input images into pixel-level predictions, classifying each pixel into its respective category. U-Net follows an encoder–decoder structure. The encoder progressively downsamples to reduce the size of the feature maps, extracting high-level semantic information from the image. The decoder then gradually upsamples the feature maps back to their original size, using skip connections to merge features from different levels, thus generating precise segmentation results.

U-Net++, proposed in 2018, is an improved version of U-Net with enhanced image segmentation capabilities [21]. Compared to the original U-Net, the main advancements of U-Net++ include extending the paths of skip connections, using deep supervision for loss computation, and model pruning. In U-Net++, the skip connections are augmented with a series of convolution operations, creating intermediate feature maps that blend image features from different levels in a more diverse manner. This approach allows U-Net++ to better capture contextual information and reduces the semantic gap between the encoder and decoder feature mappings. Deep supervision involves combining outputs from multiple layers to compute the loss function. This technique has the advantage of better updating parameters across multiple layers of the network and allows for easier selection of appropriate network depths for model pruning. Pruning involves comparing the output results of different layers on the validation set and selecting the network layer with the optimal evaluation metrics, rather than simply the highest layer. This way, non-optimal paths can be removed during the testing phase, reducing the model parameters and improving the efficiency of image segmentation during testing.

The original U-Net and its improved version, U-Net++, were both designed for two-dimensional images and cannot be directly applied to the one-dimensional cloud classification problem. To address this, we developed a cloud classification model based on U-Net++ that is tailored for one-dimensional images, referred to here as 1D-CloudNet for simplicity. Figure 1 illustrates the structure of 1D-CloudNet. For daytime input data, the model uses 24 channels, while for nighttime input data, it uses 14 channels, combining spectral and angular data features in both cases. The model outputs the probability of five cloud types: Clear, High Cloud, Middle Cloud, Low Cloud, and Ns/DC. In the figure, circles represent feature maps. Let X represent the input feature map, and let the two convolution layers with kernel size 3 be denoted as Conv1 and Conv2, followed by the batch normalization (BN) and ReLU activation functions. The processing steps are:

$$X_{mid}=ReLU\left(BN\left(Conv2\left(X\right)\right)\right)$$

(1)

$$X_{out}=ReLU\left(BN\left(Conv2\left(X_{mid}\right)\right)\right)$$

(2)

Here, $X_{out}$ represents the output of the “DoubleConv” module. This sequential flow ensures that features are extracted and refined over two stages with identical kernel sizes, enhancing representational capacity while maintaining simplicity. The red arrows indicate downsampling, which is implemented via one-dimensional max pooling and double convolution layers to achieve dimensionality reduction and feature extraction. The green arrows represent upsampling, which is accomplished through one-dimensional transposed convolution layers and double convolution layers to restore dimensions. Gray dashed lines depict skip connections for feature fusion. The blue arrows indicate deep supervision, where outputs from different levels contribute to the loss function calculation. The left ends of the blue arrows, labeled $X^{0,1}$ to $X^{0,4}$, correspond to output levels from model Level-1 (L1) to Level-4 (L4), respectively, allowing us to select the level with the best performance through pruning.

The model utilizes one-dimensional image data throughout the training, validation, and testing phases. In practical applications, L1 data from the geostationary satellite Himawari-8 is first loaded into the model. Since L1 data are in the form of two-dimensional images, we split each row into continuous one-dimensional image segments, with each segment treated as an independent input for the model. The model then performs downsampling through successive layers, gradually reducing spatial dimensions while increasing feature dimensions to extract richer discriminative information. Following this, the upsampling layers incrementally restore the spatial dimensions, mapping the feature information back to high-resolution outputs, thereby enabling fine-grained predictions of cloud type.

During training, we use a cross-entropy loss function to measure the discrepancy between the model’s output and the actual cloud type, aiming to improve classification performance [22]. The model parameters are adjusted using the Adam optimization algorithm, which minimizes the loss function to enhance both the model’s generalization capability and convergence speed [23]. To ensure robustness and reproducibility in the training process, we have meticulously set several key hyperparameters. The learning rate is dynamically adjusted based on the number of training epochs: it starts at 0.001 for the first 100 epochs, then decreases to 0.0001 from epoch 101 to 200, and further reduces to 0.00001 for epochs beyond 200. This adaptive learning rate strategy helps in effectively navigating the loss landscape and fine-tuning the model parameters. Additionally, we set the batch size to 16, which balances the computational efficiency and the stability of the training process. To further stabilize the learning process, accelerate convergence, and indirectly prevent overfitting, we incorporate BatchNorm1d, a batch normalization function provided by PyTorch’s torch.nn module. This method normalizes the activations of each layer to have a mean close to 0 and a standard deviation close to 1 (based on mini-batch statistics). Ultimately, the model outputs the probability distribution across five cloud types, with the type having the highest probability designated as the prediction result. Additionally, the sequentially predicted one-dimensional cloud classification results are reassembled to match the original two-dimensional image shape, ensuring that the cloud classification results correspond precisely to their locations on the satellite image.

2.2. Study Area

This study selects Southeast Asia as the research area for several reasons. First, as shown in the digital elevation map in Figure 2, Southeast Asia exhibits rich geographical diversity, encompassing a range of landforms such as the mainland, peninsulas, islands, and vast oceanic regions, as well as varied topography, including plains, mountains, and hills. This diversity offers an ideal environment to evaluate the model’s generalization ability across different terrains, climates, and surface conditions. The geographical coordinates of Southeast Asia span from 92°E to 140°E and 10°S to 28°N, covering multiple climate zones. Furthermore, this region is continuously influenced by active convective activities near the equator and a dynamic monsoon system, contributing to complex and variable meteorological conditions that increase the challenges and research value of cloud classification [24].

Additionally, Southeast Asia has accumulated years of geostationary satellite observations, providing a wealth of remote sensing data with high spatiotemporal resolution, which forms a solid foundation for regional cloud classification. However, current studies on the long-term trends and short-term evolution of cloud types in this region remain limited, with a lack of systematic analysis on cloud classification patterns. Therefore, selecting Southeast Asia as the research area not only enables the effective use of existing satellite observation data but also presents an opportunity to deepen understanding of the cloud distribution characteristics and their temporal dynamics in this region.

2.3. Data and Methods

2.3.1. Himawari-8/9 Data

Himawari-8 and Himawari-9, the new generation of Japanese geostationary meteorological satellites, are currently positioned in geostationary orbit at approximately 140.7°E, 35,800 km above the equator. Himawari-8 entered operational service on 7 July 2015. On 13 December 2022, the Japan Meteorological Agency transferred satellite operations from Himawari-8 to Himawari-9. They are equipped with the Advanced Himawari Imager (AHI), which includes three visible (VIS) bands, three near-infrared (NIR) bands, and ten infrared (IR) bands [25]. The observational bands, central wavelengths, and spectral response functions of the AHI onboard Himawari-8 and Himawari-9 are essentially identical. Therefore, models trained using Himawari-8 data can be directly applied to Himawari-9 data. Additional information regarding the observation bands of the Himawari-8/9 AHI is provided in

Table 1. Observation bands of Himawari-8/9 AHI.

Waveband	No.	Wavelength (µm)	Measurement/Purpose
VIS	1	0.47	Aerosol over land, coastal water, composite imaging
	2	0.51	Composite imaging
	3	0.64	Vegetation, burn scars, aerosol over water, winds, composite imaging
NIR	4	0.86	Daytime cirrus clouds
	5	1.6	Daytime cloud-top phase and particle size, snow
	6	2.3	Daytime land/cloud properties, particle size, vegetation, snow
IR	7	3.9	Surface and clouds, nighttime fog, fire, winds
	8	6.2	High-altitude atmospheric water vapor, winds, precipitation
	9	6.9	Mid-altitude atmospheric water vapor, winds, precipitation
	10	7.3	Low-altitude water vapor, winds, sulfur dioxide
	11	8.6	Total atmospheric water, cloud phase, dust, precipitation, sulfur dioxide
	12	9.6	Ozone, atmospheric turbulence, winds
	13	10.4	Surface and clouds, atmospheric window
	14	11.2	Clouds, precipitation, atmospheric window
	15	12.4	Total water, ash, atmospheric window
	16	13.3	Air temperature, cloud heights, carbon dioxide

.

This study utilizes Himawari-8/9 L1 gridded data and L2 cloud property data. The full-disk L1 data cover the observation area from 60°S to 60°N and 80°E to 160°W, with a temporal resolution of 10 min and a spatial resolution of 5 km. The L1 data variables used in this study include albedo (reflectance * cos (SOZ)) for bands 01 to 06, brightness temperature for bands 07 to 16, satellite zenith angle (SAZ), satellite azimuth angle (SAA), solar zenith angle (SOZ), and solar azimuth angle (SOA). The coverage area and spatiotemporal resolution of the L2 cloud property data are identical to those of the L1 data; however, the L2 data are only available during daytime. The L2 cloud product provides the following cloud variables: cloud optical thickness, cloud effective radius, cloud-top temperature, cloud-top height, and cloud type (ISCCP definition), with this study utilizing the cloud-type variable [26,27].

2.3.2. CloudSat Data

CloudSat, a satellite developed to measure the vertical structure of clouds, was launched into a nominal 705 km sun-synchronous orbit on 28 April 2006. CloudSat flies as part of the A-Train of satellites, positioned approximately 58 s behind NASA’s EOS Aqua and 17.5 s in front of CALIPSO (lidar), which itself flies 1 min ahead of the CNES PARASOL [28]. Since 2011, due to a battery malfunction, it has only provided observations during daytime. CloudSat is equipped with the Cloud Profiling Radar (CPR), a 94 GHz nadir-looking radar. CPR samples profiles at 625 kHz, corresponding to a range sampling distance of 240 m. One profile consists of 125 vertical bins, recorded every 0.16 s by averaging approximately 600 pulses. CPR provides an instantaneous footprint of approximately 1.3 km across track by 1.7 km along track at mean sea level [29].

The CloudSat L2 product used in this study is 2B-CLDCLASS, which classifies clouds into either stratus (St), stratocumulus (Sc), cumulus (Cu), nimbostratus (Ns), altocumulus (Ac), altostratus (As), deep convective (DC), or cirrus (Ci). A granule of 2B-CLDCLASS is defined as an orbit containing approximately 37,088 cloud profiles, each containing 125 layers of cloud types in the vertical direction [30]. In this study, clouds were categorized into four types based on the layer in which they are located and their correlation with precipitation: High Cloud, Middle Cloud, Low Cloud, and Ns/DC. For convenience, “All Cloud” represents the totality of these four cloud types. By merging clouds of the same level, we aim to reduce complexity, thereby simplifying the analysis and enhancing interpretability.

Table 2. Cloud types and properties in 2B-CLDCLASS.

Cloud Type	Cloud Base	Vertical Dimension	Horizontal Dimension	Precipitation	Cloud Type of This Study
Ci	>7.0 km	Moderate	1000 km	None	High Cloud
As	2.0–7.0 km	Moderate	1000 km	None	Middle Cloud
Ac	2.0–7.0 km	Shallow or Moderate	1000 km	Virga Possible
St	0–2.0 km	Shallow	100 km	None or Slight	Low Cloud
Sc	0–2.0 km	Shallow	1000 km	Drizzle or Snow Possible
Cu	0–3.0 km	Shallow or Moderate	1 km	Drizzle or Snow Possible
Ns	0–4.0 km	Thick	1000 km	Prolonged Rain or Snow	Ns/DC
DC	0–3.0 km	Thick	10 km	Intense Shower of Rain or Hail Possible

lists the cloud types and their properties in 2B-CLDCLASS, including the height of the cloud base, horizontal and vertical features, and the amount of precipitation for each type, as well as mapping them to the four cloud types defined in this study. In addition, the absence of clouds was classified as Clear, resulting in a total of five cloud types.

2.3.3. Spatiotemporal Matching

In the past, cloud classification tasks required researchers to manually select samples from satellite images and label them, which demanded high professional skills. Moreover, manual classification methods were highly subjective, often leading to classification errors. The 2B-CLDCLASS product provides a large number of samples for cloud classification tasks through a standardized judgment criteria and generation process, significantly enhancing the performance of machine learning methods.

We used the cloud scenario profiles from the CloudSat 2B-CLDCLASS product as the ground truth for cloud classification and the L1 images from Himawari-8 as the input features for the model. However, due to the different spatiotemporal resolutions of the two datasets, appropriate spatiotemporal matching is required to obtain a large number of training and testing samples. Specifically, we downloaded all the 2B-CLDCLASS product data from 2016 and 2017 and extracted the cloud scenario profiles that were located within the study area and had passed quality control, as indicated by the Data_quality flag in the 2B-CLDCLASS product. More precisely, we selected profiles where the Data_quality flag was 0, which indicates good quality data. The timestamps of these profiles were rounded down to the nearest ten minutes, and the profiles within each ten-minute interval were packed into a single file. Based on the timestamps of these files, we downloaded the corresponding Himawari-8 L1 data to ensure that the time difference between the CloudSat observations and the closest Himawari-8 scans were within ten minutes. Spatially, for each grid point in the Himawari-8 L1 data, we extracted the cloud profiles whose central coordinates differed by no more than 0.025 degrees in both longitude and latitude. Figure 3 illustrates the spatiotemporal matching process.

2.3.4. Data Preprocessing

A series of data preprocessing steps is necessary to enhance data usability and model accuracy after spatiotemporal matching. Firstly, the Himawari-8 L1 data contain some missing values, where correct albedo or brightness temperature results are absent, necessitating the removal of these outlier samples prior to processing. Additionally, due to significant differences in observational principles and viewing angles between Himawari-8’s AHI and CloudSat’s CPR, the cloud-type samples from these two instruments cannot be strictly equivalent, especially for samples far from Himawari-8’s nadir point, which introduces larger errors. To reduce these errors, only samples from the interior of cloud bodies are selected, while samples from broken clouds, thin cirrus, or cloud edges are discarded. Specifically, a sliding window approach is used where, for every consecutive set of five samples with consistent cloud types, the middle sample is chosen as representative of the cloud interior. Conversely, when the cloud types within the window are not uniform, the central pixel is part of broken clouds or cloud edges. This approach allows us to effectively filter out samples that do not represent the homogeneous regions of clouds, thereby improving the quality and representativeness of our training data. During processing, consecutive samples along CloudSat’s flight track are treated as pixels and concatenated into one-dimensional images for subsequent segment processing.

Figure 4 shows a flowchart of data preprocessing steps. In Figure 4a, each square represents a grid point from the Himawari-8 satellite, where gray squares indicate unmatched points, and blue squares represent points that have been successfully matched with CloudSat. These matched points are sequentially numbered from 1 to 18. Points numbered 4 to 15 are covered by a High Cloud cluster, while points 1–3 and 16–18 are Clear. Figure 4b displays the one-dimensional sequence extracted from these matched pixels, where each pixel is labeled with its corresponding cloud type. A sliding window of length five is applied to this sequence, moving along the direction of the sequence to extract pixels located within the cloud interior. Figure 4c presents the high-quality cloud classification samples obtained after preprocessing. By selecting the middle sample from each sliding window, the method reduces boundary effects and focuses on the most representative data points within the cloud cluster. This approach enhances the accuracy of cloud classification by ensuring that the selected samples are less influenced by noise or extreme values at the edges of the cloud.

In this study, the frequencies of these complete one-dimensional image lengths were analyzed, and a segment length of 64 was chosen to balance the retention of sample quantity with the preservation of contextual information needs. Next, these segmented one-dimensional images are processed in reverse order to enhance the diversity of training data. Data standardization is a critical step in the preprocessing process, aimed at enhancing the stability and effectiveness of model training by eliminating the dimensional impact of original variables. Specifically, for albedo in the VIS and NIR bands, we divide it by the cosine of SOZ, converting it to reflectance, resulting in six channels. For brightness temperature in the IR bands, we apply min-max normalization to constrain the range between 200 K and 350 K, a range determined through statistical analysis, yielding 10 channels. For the angular data—SAZ, SAA, SOZ, and SOA—we calculate both their cosine and sine values, resulting in eight channels. In total, the combined spectral and angular data comprise 24 channels. Finally, a total of 27,688 samples were obtained. For the matched data from 2016 and 2017, the data from the first 27 days of each month are split into training and validation sets in an 8:1 ratio, with the remaining data allocated to the test set. This ensures sufficient training data and independent test data to evaluate the model’s generalization ability.

Figure 5 illustrates the distribution of cloud types within the dataset. The Clear class constitutes the largest proportion, indicating that the study area is predominantly characterized by clear skies. The proportions of High Cloud, Middle Cloud, and Low Cloud decrease sequentially, reflecting the inherent characteristics of satellite observations. In cases of multi-layered clouds, satellites typically detect only the highest cloud layer, as the uppermost clouds tend to obscure those beneath them due to the perspective of satellite imaging from above. The distribution of cloud types in the dataset reveals a notable class imbalance, which may influence both model training and evaluation. Specifically, models may exhibit a bias towards the dominant class (Clear), potentially leading to suboptimal performance on less prevalent classes, such as Low Cloud. Therefore, it is essential that model evaluation goes beyond overall accuracy and incorporates performance metrics like precision, recall, and F1 scores for each cloud type. This approach ensures a more thorough understanding of the model’s effectiveness across all classes, with attention to those that are less frequent.

These preprocessing steps, including data cleaning, sample selection, segment processing, data standardization, and dataset partitioning, effectively eliminate outliers and noise from samples, ensure data consistency across different scales, and guarantee the stability and reliability of the model in practical applications, thereby achieving outstanding performance in cloud classification and recognition tasks.

3. Model Evaluation

3.1. Evaluation Metrics

In this study, cloud classification is treated as a multi-class classification problem, thus employing metrics suitable for multi-class evaluation [31]. For the model’s predictions on the test set, let TP_i denote the number of samples where the actual cloud type is i and the predicted cloud type is i. Let FN_i denote the number of samples where the actual cloud type is i but the predicted cloud type is not i. FP_i represents the number of samples where the actual cloud type is not i but the predicted cloud type is i. Finally, TN_i stands for the number of samples where neither the actual cloud type is i nor the predicted cloud type is i. Therefore, the formulas for precision, recall, and F1 score for cloud-type i are as follows:

$${\mathrm{Precision}}_i={\displaystyle \frac{{\mathrm{TP}}_i}{{\mathrm{TP}}_i+{\mathrm{FP}}_i}}$$

(3)

$${\mathrm{Recall}}_i={\displaystyle \frac{{\mathrm{TP}}_i}{{\mathrm{TP}}_i+{\mathrm{FN}}_i}}$$

(4)

$$\mathrm{F}1{\mathrm{score}}_i={\displaystyle \frac{2·{\mathrm{Precision}}_i·{\mathrm{Recall}}_i}{{\mathrm{Precision}}_i+{\mathrm{Recall}}_i}}$$

(5)

These metrics reflect the classification performance of the model for a specific cloud type. Recall indicates the model’s ability to correctly identify positive instances, while precision measures its ability to distinguish negative instances. F1 score combines both precision and recall into a single metric. Higher values for these three metrics indicate better classification performance of the model for this particular cloud type. The formulas for evaluating the overall classification performance of the model across all cloud types are as follows:

$$\mathrm{Accuracy}={\displaystyle \frac{\sum {\mathrm{TP}}_i}{\mathrm{Number}\mathrm{of}\mathrm{test}\mathrm{samples}}}$$

(6)

Accuracy is an important metric, representing the proportion of correctly classified samples out of the total number of samples. Higher values for this metric indicate stronger overall classification performance of the model across all cloud types.

3.2. Validation Results

1D-CloudNet produces four output levels, from L1 to L4, corresponding to the outputs of the network’s shallowest to deepest layers, respectively. During the training process of 1D-CloudNet, we used the training set for model training and evaluated various parameter combinations on the validation set to select the optimal configuration. The number of training epochs and output levels were identified as primary factors to optimize first. Most existing cloud classification studies focus on spectral data, often overlooking the potential contribution of angular data. However, previous research indicates that the normalization of visible albedo is significantly affected by SOZ, SAZ, SAA, and SOA [32]. Therefore, we hypothesize that incorporating angular data can provide a more comprehensive feature set for the cloud classification model, potentially enhancing classification performance.

In our experimental design, we fixed the model input to normalized spectral and angular data, totaling 24 channels, and adjusted the training epochs and output levels. The results are presented in Figure 6. As shown, the accuracy on the validation set gradually increases with the number of training epochs, with a particularly notable improvement in the first 50 epochs, after which the rate of improvement diminishes. By the 200th epoch, the accuracy for all levels exceeded 89%. Observing the trend of the loss function, we saw that after 200 epochs, the loss largely stabilized. Although increasing the training epochs further could have slightly improved accuracy, it would also have increased computational cost and might have reduced model generalizability. Thus, we determined 200 epochs to be the optimal training duration, balancing performance and computational efficiency.

At 200 epochs, the accuracy differences across L1 to L4 were less than 1%. To compare performance in practical applications, we selected Himawari-8 observation data from 4:00 on 14 September 2018 and used the model to retrieve cloud-type distributions at four output levels across Southeast Asia, as shown in Figure 7. From L1 to L4, the cloud structures exhibit an increasing trend of lateral stretching, likely related to decreased spatial resolution during the downsampling process. Specifically, at the L1, cloud structures are compact and well defined, maintaining the original shape of the clouds; however, at the L4, due to further resolution reduction, clouds appear significantly stretched horizontally, with blurred boundaries and a tendency for some cloud types to blend. As the level increases, the spatial integrity of cloud structures gradually distorts, with horizontal stretching impacting the precision and accuracy of cloud classification. This phenomenon of lateral stretching due to resolution reduction suggests that L1 is more advantageous for representing and analyzing cloud structures as it more accurately reflects the actual features of cloud clusters.

After determining the training epochs and output levels, we designed six input channel combinations based on spectral and angular information to identify the optimal configuration. Since VIS bands, NIR bands, SOZ, and SOA are only available during the day, while IR bands, SAZ, and SAA are available throughout the day, these combinations aim to fully utilize spectral and angular data, optimizing the model for both day and night training. Figure 8 presents the performance of each combination on the validation set. By comparing the effectiveness of different channel combinations, we observed the performance of combinations (1) VIS+NIR, (2) VIS+NIR+SAZ+SOZ+SAA+SOA, (3) VIS+NIR+IR, (4) VIS+NIR+IR+SAZ+SOZ+SAA+SOA, (5) IR, and (6) IR+SAZ+SAA. Overall, combinations that incorporated both spectral and angular data achieved approximately 1% higher classification accuracy than those using only spectral data, confirming the effectiveness of angular data in improving classification performance. Further analysis of the F1 scores across cloud types revealed that the classification performance of all cloud types improved, particularly in the classification of low clouds. For daytime input channel selection, combinations (1), (2), (3), and (4) were all viable candidates, with combination (4) achieving the highest classification accuracy of 89.62% on the validation set. For the nighttime model, combinations (5) and (6) were considered, with combination (6) outperforming combination (5), achieving a classification accuracy of 87.35% on the validation set.

Finally, the optimal input channel combination for the daytime model was determined to be combination (4) (VIS+NIR+IR+SAZ+SOZ+SAA+SOA), comprising a total of 24 channels, while the optimal input channel combination for the nighttime model was determined to be combination (6) (IR+SAZ+SAA), comprising 14 channels. It is important to note that since the CloudSat satellite provides only daytime observations, our dataset does not include actual nighttime data. Therefore, following the method of Wang et al. [15], we simulated nighttime observation conditions by excluding channels available only during the day and using channels accessible throughout the day. While this approach allows us to train and test the model for nighttime conditions, it may not fully capture the unique characteristics of nighttime clouds. The lack of real nighttime data could potentially lead to a bias in the model’s performance, as it might not be exposed to the full range of cloud types and conditions that occur at night. This could result in lower classification accuracy for certain cloud types, especially those that have distinct nighttime features.

3.3. Model Comparison

The random forest (RF) algorithm is a classic machine learning method that makes predictions based on an ensemble vote from multiple decision trees. RF can efficiently process high-dimensional data and reduce the risk of overfitting, making it suitable for both classification and regression tasks. Wang et al. [15] developed an RF model specifically for cloud classification. In contrast, a deep neural network (DNN) is composed of multiple dense layers in a fully connected structure, enabling it to effectively learn and represent complex nonlinear features. Gorooh et al. [16] designed the DeepCTC model using this DNN architecture. Both models have shown high classification accuracy in cloud classification tasks.

To assess the classification performance of the 1D-CloudNet model designed in this study, we performed a comparison with the RF and DeepCTC models. For fairness, the 1D-CloudNet, RF, and DeepCTC models used the same input channels and were trained on the same training set, followed by performance evaluation on the test set. Figure 9 shows the classification results for each model on the test set. The RF and DeepCTC models showed similar classification performance, achieving an overall accuracy of about 85% during the day and 83% at night. In comparison, 1D-CloudNet significantly improved classification performance, increasing overall accuracy by approximately 3% during the day and 4% at night. 1D-CloudNet also demonstrated varying degrees of improvement across cloud types, with the most notable enhancements in the classification of Middle Cloud and Low Cloud.

Analyzing the reasons for this improvement, we observe that traditional models use features from individual samples independently, without considering spatial correlations between samples. The unique design of 1D-CloudNet allows it to capture not only the feature correlation across different spectral bands within each sample but also the spatial continuity among neighboring samples. This approach preserves information across multiple bands while fully utilizing spatial context, resulting in higher cloud classification accuracy.

The confusion matrix in Figure 10 displays the performance of 1D-CloudNet in classifying different cloud types during day and night. In conjunction with the analysis in Figure 7, we gain further insight into the model’s classification performance. Under daytime conditions, 1D-CloudNet achieved an overall classification accuracy of 88.19%, with a precision of 96.35% for Clear and over 77% for High Cloud and Ns/DC. The high recall and F1 scores for these cloud types indicate strong reliability in classification. During nighttime conditions, overall classification accuracy declined slightly to 87.40%. While precision remained stable for Clear and High Cloud, it dropped significantly for Ns/DC.

Both day and night, the model exhibited lower classification accuracy for Middle Cloud and Low Cloud, averaging between 50% and 60%. The confusion matrix reveals that misclassifications primarily occur in the following cases: Middle Cloud misclassified as High Cloud and Ns/DC, and Low Cloud misclassified as Clear and Middle Cloud. This misclassification may be due to the interference of multilayer clouds, which complicates model judgment. Geostationary satellites, employing passive observation techniques, often detect only upper-level clouds in multilayered conditions, as lower clouds are frequently obscured. Additionally, some thin cumulus clouds have limited thickness and coverage, with high radiative transmissivity. Their albedo is similar to that of the surface, which can lead the model to confuse these clouds with the ground.

3.4. Practical Application Evaluation

In previous experiments, our 1D-CloudNet model demonstrated higher accuracy in daytime and nighttime classification tasks compared to existing models, a result validated solely on the test set. Due to the characteristics of the input channels, we noted that the daytime model operates effectively only during the day, while the nighttime model performs reliably throughout the day and night. To further assess the model’s practical application effectiveness, we selected a different daytime period—4:00 on 7 July 2015—for testing, distinct from the prior dataset.

As shown in Figure 11a, two distinct typhoon cloud clusters, along with several scattered cloud groups, were clearly visible within the study area. The RGB image provides rich visual information, greatly aiding in the intuitive identification of these cloud clusters. Comparing the retrieval results in Figure 11b–d, we observed that 1D-CloudNet achieved high accuracy in retrieving High Cloud and Ns/DC, consistent with the evaluation results from the test set. Additionally, we noted certain limitations in the Himawari-8 L2 product. In some areas, the product exhibits data gaps, and due to the process of its retrieval algorithms, it is unavailable at night. In contrast, 1D-CloudNet offers significant advantages in all-day applicability, providing accurate cloud type retrieval results regardless of daytime or nighttime conditions. Moreover, the model performs reliably under complex meteorological conditions, without failure in cloud-type retrieval. However, the model also has some limitations. For instance, both the daytime and nighttime models underestimate scattered Middle Cloud and Low Cloud, with the nighttime model showing a more pronounced underestimation of Low Cloud.

4. Model Application

4.1. Long-Term Analysis

Since 2015, the Himawari-8 satellite has provided data services, switching to Himawari-9 in 2022. A full-year dataset from 2016 to 2023 has now been collected, containing hourly L1 data. When SOZ in an image exceeds 90°, the image is classified as nighttime; otherwise, it is considered daytime. Cloud classification was conducted using daytime and nighttime models to obtain high spatiotemporal resolution cloud type over Southeast Asia, followed by statistical analysis of cloud type and cloud fraction.

Figure 12 shows the temporal variation in cloud fraction across Southeast Asia. Monthly averages reveal a distinct seasonal pattern, with higher cloud fraction observed during the rainy season from May to October, when precipitation is most concentrated, and lower cloud fraction during the dry season from November to April. Annual averages show long-term trends, with peak cloud fraction occurring in 2017 and 2022 and relatively low values in 2019 and 2023. A linear trend analysis indicates an overall downward trend in total cloud fraction over time, particularly in the long term. Although High Cloud shows an increasing trend in monthly averages, it decreases in annual averages, while other cloud types exhibit a consistent decline across both time scales.

Figure 13 presents the spatial distribution of cloud fraction. All Cloud is concentrated between 10°S and 10°N and near 28°N, particularly around the Maoke Mountains in New Guinea. Peak Middle Cloud fraction corresponds with areas of the highest All Cloud fraction, likely due to the high elevation and persistent snow cover in the Maoke Mountains, which the classifier may mistakenly label as Middle Cloud. High Cloud is primarily distributed between 10°S and 10°N, with Sumatra showing peak High Cloud fraction. Middle Cloud is found around the Maoke Mountains, the surrounding seas near Sumatra, and the northwest of the study area. Low Cloud is concentrated over the mainland regions of northern Southeast Asia, while Ns/DC is mainly distributed over low-latitude oceans.

Figure 14 illustrates the seasonal distribution of All Cloud fraction, showing that the Maoke Mountains consistently experience substantial cloud fraction across all four seasons. In summer, an area with elevated cloud fraction appears in the northern Andaman Sea, while in autumn, regions with greater cloud fraction include northern Sumatra and its surrounding seas. In winter, the area with increased cloud fraction shifts to the Java Sea. Overall, the regions with substantial cloud fraction display a north-south migration pattern across spring, summer, autumn, and winter. The broader distribution of areas with substantial cloud fraction in summer further supports the earlier statistical conclusion that cloud fraction is highest during this season.

Figure 15 shows the seasonal distribution of four cloud types. The enhanced cloud fraction area for High Cloud similarly undergoes a north–south shift. Middle Cloud displays a broader area of elevated cloud fraction in summer and autumn, primarily covering the western part of the study area. In winter, Low Cloud exhibits a significant region of increased cloud fraction in central and western China. Throughout the year, Ns/DC shows regions of increased cloud fraction mainly over the ocean, with peak areas located in the Andaman Sea during summer and autumn.

In summary, analysis of cloud fraction based on Himawari-8/9 satellite data from 2016 to 2023 indicates that cloud fraction is higher during the rainy season from May to October and significantly lower during the dry season from November to April, with an overall declining trend in All Cloud fraction. Additionally, elevated cloud fraction regions shift seasonally along the north–south axis, with the broadest distribution in summer and more concentrated coverage in winter. However, due to the complex terrain and climatic conditions in Southeast Asia, there are classification errors, particularly over high-altitude regions like the Maoke Mountains. These errors may result from interference in cloud classification caused by terrain and snow cover, leading to confusion between mid-level clouds and surface features. For instance, the persistent snow cover in the Maoke Mountains can have similar radiative properties to certain cloud types, making it challenging for the model to distinguish between them. Additionally, the complex topography can cause variations in the observed cloud characteristics, further complicating the classification process. In some cases, the model may misclassify high-altitude snow-covered areas as Middle Cloud, especially when the cloud cover is sparse, or the snow cover is extensive. This highlights the need for incorporating more contextual information, such as terrain data, to improve the model’s accuracy in such regions. Future work will focus on enhancing the model’s ability to recognize and differentiate between high-altitude features and actual cloud formations, thereby reducing misclassification in complex environments.

4.2. Short-Term Analysis

To investigate cloud evolution characteristics in short-term weather events over Southeast Asia, we selected Typhoon Mangkhut in September 2018 as a case study. Typhoon Mangkhut formed on 7 September and gradually dissipated on 17 September. Figure 16 illustrates the evolution of cloud types associated with Typhoon Mangkhut over Southeast Asia from 11 September to 16 September, recorded at six-hour intervals (04:00, 10:00, 16:00, and 22:00). These time-series images enable us to observe the dynamic changes in the typhoon’s cloud structure, capturing the entire process of formation, movement, landfall, and dissipation, thus allowing an analysis of the spatial distribution and evolution characteristics of the typhoon’s cloud system.

On 11 September, Typhoon Mangkhut’s cloud cluster first entered Southeast Asia, displaying a well-defined structure. The central region was dominated by Ns/DC, transitioning to Middle Cloud and High Cloud toward the outer regions. This cloud distribution reflects the typical structure of a typhoon system, with denser cloud, lower cloud-top temperature, and more intense convective activity concentrated at the center. On 12 September, the typhoon’s cloud cluster expanded further, with increased coverage of Ns/DC, indicating intensified typhoon strength and an expanded precipitation zone. Between 13 and 14 September, the cloud system continued moving westward and made its first landfall in the Philippines. On 15 September, Typhoon Mangkhut crossed the Philippines into the South China Sea. After making landfall along the Chinese coast on 16 September, the typhoon gradually dissipated. The coverage of Ns/DC decreases significantly, while High and Middle Clouds progressively become dominant. This transition marks the weakening phase of the typhoon system, as the typhoon cloud clusters gradually dissipate after landfall and eventually vanish.

In summary, these time-series images reveal the characteristic cloud evolution patterns of Typhoon Mangkhut. During the intensification phase, Ns/DC dominates the core region of the typhoon, with Middle Cloud and High Cloud progressively forming the outer layers. Following landfall, Ns/DC gradually diminishes, transitioning into Middle Cloud and High Cloud as the cloud cluster become increasingly sparse, indicating the dissipation of the typhoon system. These observations provide a clear basis for studying the spatial distribution and evolution of cloud types within a typhoon system, which is crucial for understanding the changes in typhoon-associated heavy precipitation zones and cloud structure. Additionally, the images reveal notable differences in cloud type results between day and night. There are inconsistencies in cloud type distribution over the same area at different times, particularly in the distribution range and intensity of Ns/DC. This day-night inversion inconsistency may be attributed to various factors, such as observational conditions, variations in solar radiation, and the optical properties of the atmosphere. For instance, daytime solar radiation may alter the scattering and absorption characteristics of cloud structures within the atmosphere compared to nighttime, thereby affecting cloud-type identification accuracy. Moreover, diurnal variations in cloud-top temperature and cloud optical thickness could introduce biases in cloud classification results. The existence of these day–night inversion inconsistencies indicates that there is room for improvement in enhancing the robustness and consistency of current models in cloud-type recognition.

5. Conclusions

This study introduces a novel cloud classification model, 1D-CloudNet, developed using 2016 and 2017 data from Himawari-8 and CloudSat satellites. Based on modifications to U-Net++, 1D-CloudNet is specifically designed for one-dimensional multi-channel images, enabling effective use of contextual information from neighboring pixels to enhance cloud classification accuracy. Experimental validation determined that the optimal input data for the daytime model includes all spectral and angular data (VIS, NIR, IR, SAZ, SOZ, SAA, SOA), comprising a total of 24 channels, while the optimal input data for the nighttime model consist of infrared spectral data and satellite angular data (IR, SAZ, SAA), comprising 14 channels. For both models, the Level-1 (L1) output of 1D-CloudNet most closely matched the true cloud type.

A comparison with existing models (RF and DeepCTC) showed that 1D-CloudNet achieved an overall accuracy of 88.19% for daytime classification, representing an improvement of approximately 3%, and 87.40% for nighttime classification, representing an improvement of around 4%. Clear classification performed best, with a precision exceeding 96%, while High Cloud and Ns/DC reached precision levels between 70% and 80%. Middle Cloud and Low Cloud exhibited lower classification performance, with precision between 50% and 60%. Compared with the Himawari-8 L2 product, 1D-CloudNet demonstrated the advantage of all-day applicability and reliable retrieval without classification failures, although it had limitations in underestimating scattered Middle Cloud and Low Cloud.

In practical application, 1D-CloudNet was used to analyze long-term cloud fraction and cloud type variations over Southeast Asia, as well as cloud characteristics in short-term weather events. The long-term analysis of cloud fraction from 2016 to 2023 revealed higher cloud fraction during the rainy season from May to October and significantly lower cloud fraction during the dry season from November to April, with an overall downward trend. Spatially, elevated cloud fraction regions migrated seasonally along the north–south axis, with the broadest distribution in summer and more concentrated coverage in winter. For short-term analysis, Typhoon Mangkhut in 2018 was examined, revealing structural features of the typhoon cloud cluster: the innermost ring consisted of Ns/DC, surrounded by Middle Cloud in the middle ring and High Cloud in the outer ring, with cloud optical thickness decreasing outward. After landfall, Ns/DC gradually transitioned to High Cloud and Middle Cloud and eventually dissipated.

In summary, the key findings are as follows: (1) The development of 1D-CloudNet significantly improves cloud classification accuracy using optimized inputs for daytime and nighttime. (2) The model demonstrates competitive performance compared to existing methods, particularly for Middle and Low Cloud. (3) Practical applications reveal insights into long-term cloud fraction trends and the structural evolution of cloud clusters during typhoons.

However, 1D-CloudNet demonstrated certain limitations in practical applications. First, high-altitude snow cover was always misclassified as Middle Cloud. Second, there were notable differences between the daytime and nighttime model results in certain areas, indicating the need for further improvements in the model’s robustness and consistency in cloud type recognition.

Future work will focus on optimizing 1D-CloudNet in the following ways: (1) incorporating terrain data or enhancing the model’s ability to recognize high-altitude features to improve adaptability in complex environments and reduce misclassification; (2) exploring more consistent feature fusion methods or adjusting the model architecture to increase consistency between daytime and nighttime classification results; (3) expanding the training dataset, especially by adding Middle Cloud and Low Cloud samples, or introducing dedicated modules for Middle Cloud and Low Cloud recognition to improve classification performance for these cloud types; (4) extending the dataset to encompass diverse regions and climate conditions, incorporating higher-resolution observational data and integrating data from other satellites that provide nighttime cloud classification labels to enhance the model’s generalization capability. With these improvements, we anticipate that 1D-CloudNet will provide more reliable cloud classification results for regional climate studies and extreme weather monitoring and can be further extended for global cloud classification and climate research.