STPRNet: A Point Cloud Deep Learning Framework for Blending Level 2 Sea surface temperature Satellite Data(https://doi.org/10.63386/620246)

Authors: Jiaye Luan^1*

 Titles & Affiliations:

^1*Undergraduate, Reading Academy, Nanjing University of Information Science and Technology, Nanjing, China, 210044.

Emails of Authors:

First author: 202283100034@nuist.edu.cn

Corresponding author: 202283100034@nuist.edu.cn

ORCID Links of Authors:

First author: https://orcid.org/0000-0000-8334-6872

Authors’ bio:

Jiaye Luan is currently pursuing the B.S. degree in Mathematics and Applied Mathematics from Nanjing University of Information Science and Technology, Nanjing, China. His research interests include deep learning architectures, neural network modeling, and machine learning applications in environmental remote sensing. He has contributed to research in cloud optical parameter retrieval from geostationary satellite imagery using deep learning frameworks and is currently participating in the National Natural Science Foundation project ”Digital Empowerment and Smart Regulation Project” focused on water ecosystem observation technology services. Mr.Luan has co-authored publications in peer-reviewed journals in the field of remote sensing.

STPRNet: A Point Cloud Deep Learning Framework for Blending Level 2 Sea surface temperature Satellite Data

Jiaye Luan*

Reading Academy, Nanjing University of Information Science and Technology, Nanjing 210044, China.

*e-mail: 202283100034@nuist.edu.cn

Abstract: Sea surface temperature (SST) derived from satellite measurements is fundamental for understanding ocean-atmosphere interactions and monitoring climate change. While infrared (IR) sensors provide high spatial resolution SST measurements, they are limited by cloud cover. Conversely, microwave (MW) sensors offer all-weather capability but at coarser resolutions. These two sensor types provide complementary capabilities that enhance the overall accuracy and coverage of SST observations when used in conjunction. Existing methods for merging IR and MW data typically rely on daily Level 3 (L3) or Level 4 (L4) products that sacrifice temporal granularity and spatial detail due to the inherent limitations of regular gridding. To address these limitations, we propose Spatio-Temporal Point Regression Network (STPRNet), a novel point cloud-based deep learning framework that directly operates on Level 2 (L2) satellite data, treating instantaneous observations as three-dimensional points in space and time. By leveraging a modified PointNet++ architecture, our model can process L2P data from both MODIS and AMSR2 sensors to produce enhanced SST estimates. The STPRNet framework demonstrates superior performance, achieving nighttime root-mean-square errors (RMSE) of 0.52°C for pure microwave-based SST estimation and 0.20°C when utilizing hybrid point clouds combining both IR and MW observations.

Keywords: data blending, Level 2 satellite data, point cloud deep learning, remote sensing, sea surface temperature (SST), spatio-temporal data analysis

1. Introduction

Sea surface temperature (SST) is a crucial parameter for understanding ocean-atmosphere interactions and plays a significant role in climate dynamics. Variations in SST can impact weather patterns, marine ecosystems, and the development of tropical storms and hurricanes, as well as contribute to extreme weather events like droughts and floods [1], [2]. SST is also a key indicator for climate change and is essential for accurate weather forecasting and long-term climate monitoring [3], [4]. Satellite remote sensing has enabled the collection of high-resolution SST data, which is vital for understanding global and regional climate processes [5], [6].

Satellite infrared (IR) and microwave (MW) observations together form the primary foundation for current global SST products [1], [7]. While IR sensors-such as the Advanced Very High Resolution Radiometer (AVHRR) and the Moderate Resolution Imaging Spectroradiometer (MODIS)-can deliver high spatial resolutions (on the order of 1-4 km), they are easily obstructed by cloud cover and high aerosol concentrations, causing data gaps and reduced coverage [8], [9]. In contrast, MW sensors-including instruments like the Advanced Microwave Scanning Radiometer (AMSR) series-can penetrate clouds more effectively, enabling near all-weather observations and more complete spatial coverage [10], [11]. However, MW datasets are characterized by coarser spatial resolution (often around 25 km or more), and their accuracy can degrade near coastline [11], [12]. Because these two measurement approaches are complementary-IR excels in spatial detail but lacks coverage under cloudy conditions, while MW offers broader coverage but lower resolution- merging or synergizing their strengths has emerged as a promising pathway for generating SST fields that combine high resolution, accuracy, and spatiotemporal completeness [7].

Numerous methods have been proposed for merging IR and MW SST data over the past several decades. Traditional geostatistical techniques such as Objective Analysis (OA) and Optimum Interpolation (OI) have been widely applied [13]-[18]. However, these approaches suffer from smoothing effects that reduce fine spatial characteristics, particularly in coastal areas where detail is crucial [9], [12]. They also require strong assumptions about error statistics that are often difficult to ascertain in practice [13], [19]. Similarly, data assimilation methods-including Variational (VAR) approaches and Kalman Filter (KF)-introduce subjective error specifications that affect accuracy [9], [12], while KF requires scale trans- formations that can introduce additional uncertainties [6]. The change of support (COS) problem remains a fundamental challenge when integrating data of different resolutions [20]- [22], as traditional resampling or local averaging methods inevitably lose spatial variability within coarser pixels and are adversely affected by missing data [5], [11]. More recently, deep learning-based super-resolution techniques have made significant strides in reconstructing high-resolution SST fields [23]-[27]. However, these approaches typically rely on gridded datasets at L3 or L4 resolutions and do not fully exploit the instantaneous spatio-temporal relationships found in Level 2 (L2) data [26]. Moreover, some deep learning models incorporate only the single nearest microwave point for each IR pixel [28], overlooking surrounding MW observations that may contain crucial local context. What is needed is a methodology that can directly process L2 observations in their native irregular point format, preserving the rich spatio-temporal relationships that are typically lost in gridded products.

The key idea of our approach stems from point-based deep learning methods, notably PointNet and PointNet++ [29], [30], which were originally developed to handle 3D geometric data in tasks such as shape classification and segmentation. Unlike traditional convolutional neural networks (CNNs) that operate on regular grids (e.g., images or voxels), point-based models can directly process irregular, unordered sets of points by learning spatial encodings and aggregating them into deeper feature representations. Inspired by this flexibility, we ob- served that Level 2 satellite data-being collected at irregular spatio-temporal coordinates-can similarly be viewed as point clouds in a three-dimensional space (latitude, longitude, and time). Moreover, hierarchical operations in PointNet++ capture both local and global structures, making it well-suited to learn complex relationships among nearby microwave observations distributed in time and space.

Our methodology employs the Spatio-Temporal Point Regression Network (STPRNet), a comprehensive point cloud- based approach for SST estimation. STPRNet begins with a data matching procedure that aligns observations from MODIS (infrared) and AMSR2 (microwave) sensors based on spatial proximity and temporal coincidence. The framework consists of three integrated modules: First, a Spatio-Temporal Data Preprocessing module transforms the matched satellite data into sensor-aware point clouds, where each point is represented by an eight-dimensional feature vector containing normalized spatial coordinates (longitude, latitude), temporal information, the target point’s coordinates, SST values, and a sensor type indicator (0 for microwave, 1 for infrared). Second, a Multi- Sensor Aware Hierarchical Feature Learning module hierarchically captures multi-scale features and spatiotemporal relation- ships from the irregular point data while employing attention- based pooling to weight observations from different sensors based on their inherent characteristics. Third, a Regression Network transforms the fused representations into accurate SST estimates. During training, the model is optimized using mean squared error as the loss function, enabling it to generate comprehensive SST maps from combined AMSR2 and MODIS inputs once deployed.

Our main contributions are summarized as follows:

1) Pioneering Point Cloud-Based Deep Learning for Spatiotemporal Data Analysis: We are the first to employ point cloud- based deep learning networks for oceanographic data analysis, establishing a novel paradigm that processes irregular spatiotemporal observations directly without gridding, enabling more effective learning of complex relationships in environmental data.

2) Exploiting Off-Grid L2 Data: Unlike existing works that rely on L3 or L4 products, our method leverages real-time L2 data with enhanced accuracy and higher temporal/spatial resolution, enabling a three-dimensional (latitude, longitude, and time) analysis of SST variabil- ity.

3) Developing Sensor-Aware Attention Fusion Method- ology: We introduced a novel attention-based fusion approach that explicitly distinguishes between microwave and infrared observations, dynamically weighting each sensor’s contribution based on learned quality characteristics. This method can effectively leverage the complementary strengths of each data source and enhance the robustness of SST estimation.

4) Superior Performance on Pure and Hybrid Inputs: Experiments show that our model achieves an RMSE of 0.52°C when using nighttime microwave-only point clouds and 0.20°C with the nighttime hybrid (microwave + infrared) input, highlighting the flexibility and accuracy of our framework.

The remainder of this paper is organized as follows. In Section 2, we describe our data sources, study region, and multi-temporal data matching approach for building point cloud datasets. Section 3 presents the STPRNet framework, including our sensor-aware preprocessing, hierarchical feature learning, and temperature regression components. Section 4 evaluates model performance through experiments with various input configurations and architectural optimizations. Finally, Section 5 summarizes our contributions and discusses future directions.

2.DATA

Our study employs a comprehensive data acquisition and processing workflow to enable accurate sea surface temperature (SST) estimation using a combination of infrared and microwave satellite observations. The workflow consists of four primary stages: data source selection, study region definition, multi-temporal data matching, and dataset construction for model training and evaluation. Fig. 1 illustrates this methodical approach, highlighting the progression from raw satellite data to structured point cloud datasets suitable for our deep learning framework.

Fig. 1. Data processing workflow showing the four stages: (1) data source selection from AMSR2 and MODIS satellites, (2) northwest Pacific region selection, (3) spatio-temporal data matching to create microwave and hybrid point clouds, and (4) final dataset construction for model training and testing.

2.1 Data Sources

Our analysis employs two primary datasets from the Group for High Resolution Sea Surface Temperature (GHRSST) Level 2P (L2P) product suite. The first dataset consists of IR measurements from the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard NASA’s Aqua satellite (MODIS_A-JPL-L2P-v2019.0) [31]. The second comprises MW observations from the Advanced Microwave Scanning Radiometer 2 (AMSR2), processed by Remote Sensing Systems (REMSS) under the designation AMSR2-REMSS-L2P- v8.2 [32]. Both datasets are distributed through NASA’s Physical Oceanography Distributed Active Archive Center (PO.DAAC) [33].

The MODIS instrument captures data in 36 spectral bands at varying spatial resolutions, with SST products derived from both day and night observations [34]. For SST retrieval, MODIS utilizes the long-wave IR channels (11 and 12 m wavelengths) and mid-infrared channels (3.95 and 4.05 m), achieving a spatial resolution of approximately 1 km at nadir [34]. The AMSR2 sensor, launched on the GCOM-W satellite by the Japan Aerospace Exploration Agency (JAXA), provides microwave-based SST measurements with a broader swath width of 1450 km [35]. The antenna rotates once per 1.5 seconds, enabling near-complete global coverage every two days through its conical scan mechanism [35].

We specifically utilize GHRSST L2P products, which represent “preprocessed” Level 2 data containing both geophysical variables and essential auxiliary fields [36]. These products maintain the native sensor resolution while including comprehensive quality flags and uncertainty estimates for each observation [36]. This choice of L2P data preserves the inherent spatial and temporal granularity of the measurements, which is crucial for our point cloud-based approach. While GHRSST also offers higher-level products (L3 and L4) that provide gridded or gap-free analyses, these products typically sacrifice some of the fine-scale detail present in L2P data [36].

2.2 Study Region

In this study, we focus on the northwest Pacific region spanning 0-60°N and 100-180°E (Fig. 2). This region experiences more tropical cyclones than any other area worldwide, with an annual average of approximately 35, of which about 80% develop into typhoons [37]. Around 26 tropical cyclones per year reach at least the intensity of tropical storms, accounting for about 31% of the global total [37]. The strong air–sea interaction in this region and the frequent variations in sea surface temperature (SST) make it an ideal testbed for SST estimation and data fusion studies.

Fig. 2. Study region in the northwest Pacific (0-60°N, 100-180°E) highlighted in red.

2.3 Data Matching

Our temporal framework encompasses observations from January 1, 2020, through June 30, 2020, providing sufficient seasonal variation for robust model development. To maintain consistency in the SST estimation process and account for diurnal variations, we separated the observations into distinct daytime and nighttime datasets.

For each target location, we implemented a multi-stage temporal and spatial point cloud construction strategy. Spatially, we employed the Haversine formula for great-circle distance calculations to identify neighboring points around each target location:

(1)

computes the great-circle distance  between two points on the Earth’s surface by taking into account the Earth’s spherical geometry. In this formula,  represents the mean radius of the Earth (approximately 6,371 km),  and  denote the latitudes, and  and  denote the longitudes of the two points, with all angular measures expressed in radians.

Temporally, we employed a primary matching protocol using a ±1-hour window centered on each MODIS acquisition time for same-day observations. Additionally, to evaluate our model’s temporal association capabilities, we extended this matching to include observations from the same locations on previous days (up to three days back). This approach produced several distinct point cloud configurations:

1) Same-day hybrid point clouds: For each target location, we first searched for available IR observations within defined spatial bounds (standardized as -0.04315°to 0.04224°for longitudinal deviation and -0.04236°to 0.04250°for latitudinal deviation based on statistical analysis). When IR observations were available but insufficient to reach 64 points, we supplemented with the nearest MW observations to create hybrid point clouds.

2) Previous-day IR point clouds: We constructed additional test datasets by collecting IR observations from the target location on previous days (one day back, two days back, and three days back). These temporal variants allow us to assess how well our model handles time-lagged observations when current IR data is unavailable.

3) Pure microwave point clouds: When no IR observations were available within the spatial window (typically due to cloud cover), we created point clouds consisting of the 64 nearest microwave observations. This configuration represents scenarios with complete IR data unavailability and tests our model’s ability to derive IR-equivalent SST solely from microwave measurements.

Each point in both IR and MW clouds is characterized by four fundamental variables: longitude, latitude, observation time, and sea surface temperature (SST). The spatial characteristics determined through standard deviation analysis serve as criteria for determining whether SST estimation is necessary for a given target point. When 64 IR points can be gathered within the defined spatial bounds, the target point is deemed to have sufficient observational density and does not require SST estimation.

2.4 Dataset Construction

Our experimental design incorporates two primary dataset configurations derived from the point cloud construction process. The first configuration, termed the microwave-only dataset, comprises 50% of the filtered files containing exclusively microwave point clouds. This dataset simulates conditions where infrared observations are completely unavailable due to cloud cover or other atmospheric interference.

The second configuration, designated as the mixed dataset, utilizes the remaining 50% of the filtered files to create hybrid point clouds. For each target location in this dataset, we calculated the number of additional points needed to achieve a total of 64 nodes, supplementing any existing IR observations with the nearest available microwave measurements. This approach represents realistic scenarios where partial IR coverage exists but contains gaps that require complementary microwave data to achieve complete spatial coverage.

The two primary configurations were first merged into a combined dataset, which was then randomly split into training and testing subsets using an 80-20 ratio. Subsequently, the dataset was supplemented with temporal variant data (previous-day IR observations) to facilitate a comprehensive model evaluation. This stratified splitting approach ensures that pure microwave, hybrid, and various temporal scenarios are adequately represented in both training and testing sets, enabling a thorough and robust assessment of model performance across diverse observational conditions.

3.Method

Our proposed framework, the Spatio-Temporal Point Regression Network (STPRNet), is an enhanced architecture for SST estimation built upon a modified PointNet++ architecture specifically tailored for regression tasks using irregular Level 2 satellite data. STPRNet introduces a sensor-aware attention mechanism that fuses microwave and infrared observations by leveraging their inherent quality differences. The method consists of three main modules: a Spatio-Temporal Data Preprocessing stage that normalizes, centers, and merges the point clouds while preserving sensor identity; a Multi-Sensor Aware Hierarchical Feature Learning module that employs attention-based pooling to weight points based on their sensor origin; and an Adaptive Temperature Regression Network that predicts the final SST value. Fig.3 illustrates the comprehensive architecture of STPRNet.

Fig. 3. Architecture of STPRNet model for SST estimation

In our formulation, the overall mapping function of the network is represented as:

(2)

where  is the estimated SST,  is the processed point cloud input, and  denotes the network parameters. The loss function is defined as the mean squared error (MSE):

(3)

where  is the ground truth SST for the -th sample. In the following sections, we describe each of these modules in detail.

3.1 Spatio-temporal Data Preprocessing

Raw Level 2 satellite data must be transformed into a unified point cloud representation suitable for our network. For each observation, let the central (target) point be defined as   and a surrounding point  for  Normalization is applied using fixed global bounds. For any scalar variable  (longitude, latitude, time, or SST), the normalized value is computed as:

(4)

Thus, the normalized feature vector for the -th point becomes

(5)

Subsequently, we center the point cloud relative to the target point by subtracting the normalized central coordinates  from each surrounding point:

(6)

Each point is then augmented with the target’s normalized coordinates and a sensor type indicator  to yield an eight-dimensional feature vector:

(7)

where  denotes the sensor type, with 0 representing microwave and 1 representing infrared observations. This sensor identity information is critical for the attention mechanism to automatically learn the optimal weighting between the two sensor types.

Finally, after processing, the microwave and infrared point clouds are merged into a single input matrix , which preserves both local spatial relationships, global contextual information, and sensor identity essential for accurate SST estimation.

3.2 Multi-Sensor Aware Hierarchical Feature Learning

The second module of STPRNet implements a multi-scale hierarchical feature learning process that intelligently integrates observations from different sensor types. Unlike traditional point cloud processing approaches, our method explicitly accounts for the heterogeneous nature of satellite observations by incorporating sensor-awareness throughout the feature extraction pipeline. This module consists of several key components, each contributing to the network’s ability to effectively fuse microwave and infrared data for accurate SST estimation.

3.2.1 Proximity-Prioritized Sampling

In conventional PointNet++ implementations, farthest point sampling (FPS) is employed to ensure comprehensive coverage of the entire point cloud. However, for the specific task of SST estimation at a target location, proximity to the central point is more relevant than uniform spatial distribution. Therefore, we introduce proximity-prioritized sampling through nearest point sampling (NPS).

For a given target point , the set of indices for the nearest  points is determined by:

(8)

3.2.2 Sensor-Aware Feature Extraction

An important aspect of STPRNet is the explicit integration of sensor type information during feature extraction. Each point in our preprocessed point cloud carries an additional dimension that identifies its origin as either microwave (0) or infrared (1) data. This sensor identity is preserved throughout the initial stages of feature extraction before being used to compute attention weights.

For each point  with feature vector , we extract the sensor type indicator  and process it through a small neural network  to obtain sensor-specific feature representations:

(9)

This allows the network to learn sensor-specific characteristics and account for the inherent differences between microwave and infrared observations.

3.2.3 Attention-Based Feature Aggregation

Original max pooling is effective in classification tasks because it captures the most dominant features, thereby enhancing discriminative power. However, in the context of SST regression, max pooling may disregard subtle variations that are critical for precise temperature estimation. In contrast, our approach introduces an attention mechanism that dynamically weights the contribution of each point based on its sensor type and feature characteristics.

The attention mechanism computes a weight  for each point in a local neighborhood  centered around a sampled point :

(10)

where  is a small neural network that maps sensor type to attention scores, and  is a sigmoid activation function. These attention scores are then normalized to obtain the final attention weights:

(11)

where  is a small constant for numerical stability. The normalized attention weights are then used to compute a weighted aggregate of the features:

(12)

This attention-based aggregation allows the network to automatically learn the optimal weighting between microwave and infrared observations based on their characteristics and relevance to the SST estimation task.

3.2.4 Multi-Scale Feature Hierarchy

Our feature learning process operates across multiple scales through a hierarchical architecture consisting of three set abstraction layers. Each layer operates at a different spatial scale and abstraction level, enabling the network to capture both localized patterns and broader contextual information.

In the first set abstraction layer (SA1), we sample  points within a radius  using  neighbors and process them through a multi-layer perceptron (MLP) with layer dimensions [4,4,8]. And the sensor-aware attention mechanism is applied:

(13)

where  denotes the nonlinear mapping implemented by the MLP.

The second set abstraction layer (SA2) operates at a broader scale, sampling  points within a radius  using  neighbors. These points are processed through an MLP with dimensions [8,8,16]:

(14)

The final global aggregation layer (SA3) consolidates all features into a unified representation using an MLP with dimensions [16,32,64], yielding a global feature vector

(15)

Additionally, we remove batch normalization layers from these MLP layers to prevent any potential bias or normalization artifacts that might degrade the precision of the regression output.

3.3 Regression Module

The final module of STPRNet transforms the rich multi-sensor feature representation into an accurate sea surface temperature prediction. This Adaptive Temperature Regression Network consists of a streamlined fully connected (FC) architecture that maps the aggregated global feature vector to a scalar SST value.

The regression network architecture is defined as:

(16)

where , and  are the corresponding bias terms. Each layer employs the ReLU activation function to introduce non-linearity.

3.4 Model Details

Our STPRNet architecture implements a point-based processing pipeline tailored for SST estimation. The network consists of two main components: a hierarchical point cloud processing module and a regression module. The point cloud processing module comprises three set abstraction layers (SA1, SA2, and SA3) operating at progressively broader spatial scales, with decreasing numbers of sampling points and increasing feature dimensions to capture both local patterns and global context. The regression module then transforms these extracted features into accurate SST predictions through a series of fully connected layers. Table 1 provides the detailed specifications of each layer in our network architecture, including input/output tensor dimensions, sampling parameters, and feature transformations implemented at each stage.

Table 1. Detailed specifications of each layer in the proposed network architecture

Module	Layer	Input Tensor	Output Shape	Sampling Parameters	Feature Transformation
Point Cloud Abstraction	SA1			,
	SA2
	SA3
Regression	FC1			–
	FC2			–
	FC3			–

4.Experiments

4.1 Setup

The model training and evaluation were conducted on a workstation equipped with a 12th Gen Intel® Core i7-12700H processor and an NVIDIA GeForce RTX 3060 Laptop GPU. The system features a display memory of 22276 MB, with 5996 MB dedicated and 16280 MB shared memory, and runs on the Windows 11 operating system. All experiments were executed using Python 3.8.19 and PyTorch 2.1.0. For optimization, we employed the Adam optimizer with the following hyperparameters: a learning rate of ,and a weight decay of  .Due to memory constraints and to ensure efficient training, the batch size was set to 24.

4.2 Metrics

To comprehensively assess the performance of our model, we adopted several evaluation metrics that capture different aspects of regression accuracy. The primary metrics used include:

4.2.1 Root Mean Square Error (RMSE)

RMSE quantifies the average magnitude of the estimation error. It is defined as:

(17)

where  is the number of samples,  is the ground truth SST for the th sample, and  is the corresponding estimated SST. A lower RMSE indicates a model that closely approximates the true values.

4.2.2 Mean Absolute Error (MAE)

MAE measures the average absolute differences between estimated values and actual observations:

(18)

This metric is particularly useful in understanding the overall error magnitude without penalizing large errors disproportionately.

4.2.3 Mean Bias (MB)

Mean Bias measures the average signed difference between estimated values and actual observations:

(19)

This metric indicates whether the model systematically overestimates (positive bias) or underestimates (negative bias) the true values. A value closer to zero indicates less systematic error in the estimation.

4.2.4 Relative Standard Deviation (RSD)

RSD quantifies the variability of estimation errors relative to the mean of the actual values:

(20)

where  is the standard deviation of the differences between estimated and actual values, and  is the mean of the actual values. RSD provides insight into how consistent the model estimations are relative to the magnitude of the observed values, with lower percentages indicating more consistent estimations.

4.2.5 Coefficient of Determination

evaluates how well the estimated values explain the variance in the ground truth. It is computed as:

(21)

where  is the mean of the ground truth values. An  value closer to 1 indicates a higher proportion of explained variance by the model.

4.2.6 Peak Signal-to-Noise Ratio (PSNR)

PSNR is typically used to assess the quality of reconstruction in imaging tasks and, in our case, it quantifies the similarity between the estimated SST map and the ground truth SST map. PSNR is given by:

(22)

where  is the maximum possible value of the SST (based on the dynamic range of the normalized data) and MSE is the mean squared error between the estimated and true SST maps. Higher PSNR values indicate a closer match between the two images, signifying superior reconstruction quality.

4.3 Estimation Errors in Pure Microwave Point Clouds

Our analysis of pure microwave point clouds reveals that the MAE during the day (0.40) is slightly higher than the nighttime MAE (0.37). Additionally, the RMSE values follow a similar pattern, with daytime RMSE measuring 0.54, compared to a marginally lower RMSE of 0.52 at night. This trend suggests that estimation accuracy is influenced by the time of day at which the remote sensing observations were made.

One potential explanation for the observed increase in error during daytime conditions lies in the complexity introduced by solar radiation. During the day, diurnal warming occurs, leading to the formation of a shallow warm layer that can affect SST estimates when strong winds are absent. Without sufficient vertical mixing to homogenize the water column, the microwave signals can become susceptible to inaccuracies due to the reflective and scattering effects induced by solar radiation [38]. In contrast, nighttime measurements are less impacted by solar interference, as the absence of sunlight reduces such contamination, leading to more reliable SST estimates [38]. Consequently, the use of nighttime data allows for more robust SST estimations, avoiding complications that may arise from high wave number measurements affected by solar radiation reflections. Based on the superior performance observed with nighttime data, we utilize nighttime observations for all subsequent model validation and experimentation.

Figs. 4-6 illustrate the error characteristics of our model when using pure microwave point clouds. The spatial distribution of RMSE (Fig. 4) shows higher errors in coastal regions, aligning with known limitations of passive microwave SST, which suffers from poor spatial resolution and susceptibility to coastal effects [39]. The error histograms (Fig. 5) confirm lower standard deviation during nighttime (0.5170) versus daytime (0.5394), while the visual comparison (Fig. 6) demonstrates that our model effectively preserves the temperature patterns present in the original AMSR2 data.

(a)                                   (b)

Fig. 4. Spatial distribution of RMSE across the study region for (a) daytime and (b) nighttime SST estimations using pure microwave point clouds.

(a)                                   (b)

Fig. 5. Error distribution histograms for (a) daytime and (b) nighttime SST estimations using pure microwave point clouds.

(a)                                   (b)

Fig. 6. Comparison of (a) original AMSR2 SST and (b) model-estimated SST results using pure microwave point cloud input

4.4 Blending Errors in Hybrid Point Clouds

The hybrid point cloud approach demonstrates a marked improvement in the accuracy of SST blending, showcasing substantially lower error values compared to the pure microwave datasets. For the hybrid configuration, the daytime MAE is recorded at 0.146, while the night exhibits a slightly higher MAE of 0.155. The RMSE values show a similar pattern, with daytime RMSE at 0.19 and nighttime RMSE at 0.20.

As illustrated in Fig. 7, the spatial distribution of RMSE for hybrid point clouds shows significantly lower values across the entire study region compared to pure microwave estimations, with most areas exhibiting RMSE values below 0.2K. Fig. 8 reveals that error distributions for hybrid blending results are tightly concentrated around zero, with notably smaller standard deviations (daytime: 0.1923, nighttime: 0.1788) than those observed in the pure microwave approach. The comparison in Fig. 9 demonstrates how our model effectively integrates information from both AMSR2 (Fig. 9a) and MODIS (Fig. 9b) to produce comprehensive SST fields (Fig. 9c) that maintain high spatial detail while providing complete coverage.

(a)                                   (b)

Fig. 7. Spatial distribution of RMSE for (a) daytime and (b) nighttime blended SST using hybrid point clouds.

(a)                                   (b)

Fig. 8. Error distribution histograms for (a) daytime and (b) nighttime hybrid point cloud blended SST.

(a)                         (b)                       (c)

Fig. 9. Comparison of (a) AMSR2 SST measurements, (b) MODIS SST measurements, and (c) model-blended SST results using hybrid point cloud input.

These results underscore the complexity and variability involved in fusing SST data derived from different satellites and in situ sensors. The differences in mean errors between daytime and nighttime observations emphasize the necessity of accounting for atmospheric and environmental conditions that may influence sensor readings. By recognizing these factors, the performance of SST estimation systems can be effectively enhanced through optimal data integration strategies that can accommodate time-of-day variability. Overall, the findings illuminate the advantages of hybrid data approaches for SST estimation and suggest avenues for future research focusing on data fusion methodologies that can further optimize accuracy.

4.5 Temporal Efficacy Analysis

To evaluate the model’s performance with infrared data from different temporal periods, we conducted a systematic analysis comparing various input configurations. This experiment assesses how well STPRNET leverages historical IR observations when same-day data is unavailable.

We tested five distinct input configurations: (1) pure microwave point cloud, (2) hybrid point cloud with same-day IR data, (3) hybrid point cloud with 1-day old IR data, (4) hybrid point cloud with 2-day old IR data, and (5) hybrid point cloud with 3-day old IR data. We maintained consistent model architecture and hyperparameters across all tests to isolate the impact of temporal displacement. As a baseline, we compared our results with a fully connected neural network that could only process pure microwave inputs.

Table 2. Performance comparison with different temporal inputs versus baseline model

Model Type	Input Type	RMSE(K)	MAE(K)	Mean Bias(K)	RSD(%)
Fully Connected	Pure MW	0.543559	0.407373	-0.115275	0.179843
STPRNET	Pure MW	0.539421	0.403179	-0.003917	0.182642
STPRNET	Same-day IR hybrid	0.192343	0.145788	-0.005222	0.065065
STPRNET	1-day old IR hybrid	0.487739	0.377095	0.026182	0.163894
STPRNET	2-day old IR hybrid	0.483791	0.367376	0.044159	0.161914
STPRNET	3-day old IR hybrid	0.494936	0.376232	0.022945	0.166765

The results in Table 2 reveal several important patterns. STPRNet with same-day IR data demonstrates strong performance (RMSE: 0.192343K), exhibiting a 64.3% error reduction compared to the pure microwave configuration. When using temporally displaced IR observations, performance metrics indicate increased error values, yet all hybrid configurations maintain efficacy above the pure microwave baselines. The conventional IDW approach provides a reference point for traditional spatial interpolation techniques, while the fully connected network, which processes feature vectors rather than point clouds, establishes context for evaluating STPRNet’s architectural advantages in handling spatial data structures.

An interesting pattern emerges where 2-day old IR data (RMSE: 0.483791K) yields slightly better results than 1-day old data (RMSE: 0.487739K), possibly indicating favorable observation conditions during that specific period. The bias analysis shows near-zero bias for same-day IR inputs (-0.005222K), while older IR data introduces a positive bias that peaks with 2-day old observations (0.044159K).

While same-day IR observations provide optimal accuracy, historical IR data from up to three days prior still enhances estimation quality. This temporal flexibility offers valuable redundancy for operational SST mapping in regions where consistent same-day IR observations may be unattainable due to cloud cover or other atmospheric interference.

4.6 Pointnet++ Model Architecture Optimization Experiments

4.6.1 Eliminate Batch Normalization and Dropout Layers

In this experiment, we utilized the original PointNet++ architecture and adapted it for a regression task by implementing modifications designed to enhance estimation accuracy for Sea Surface Temperature (SST) estimations. We aimed to evaluate the performance of the original PointNet++ model directly in regression mode and subsequently analyze the improvements achieved after modifying the architecture by removing batch normalization (BN) layers in both the point cloud abstraction layer and the fully connected layers, alongside the exclusion of dropout layers. This study posits that the original components of the PointNet++ architecture, tuned to enhance performance for classification tasks, could inadvertently introduce noise that detracts from accuracy in regression applications.

The first phase of this experiment employed the standard PointNet++ configuration, which included both BN and dropout layers optimized for classification tasks. As anticipated, the estimated values produced by this configuration exhibited substantial deviations from the actual SST observations. The scatter plot presented in Fig.10 illustrates a comparison of estimated SST values against actual SST values for the ‘pure microwave’ and ‘mixed point cloud’ scenarios prior to the adjustments. Notably, the estimated SST values, particularly at lower temperatures, significantly deviated from the ideal 1:1 line. The R² value was recorded at 0.97, and the Root Mean Square Error (RMSE) was calculated to be 1.25, indicating that the initial configuration faced challenges in accurately estimating lower SST values.

Further analysis of the ‘mixed point cloud’ scenario revealed that prior to modification, the RMSE for the hybrid point cloud was recorded at 1.24, which mirrored the RMSE of the pure microwave point cloud. This result indicated that the regression model was underutilizing the valuable infrared data inherent in the hybrid point cloud. However, after adjustments, the RMSE for the hybrid point cloud decreased dramatically to 0.37, surpassing the RMSE of the microwave point cloud, which was reduced to 0.55. This observation serves as compelling evidence that the modified model successfully leveraged the infrared contributions within the hybrid point cloud, optimizing both accuracy and estimation capability.

Fig. 10. Scatter plots comparing estimated versus true SST for (a) pure microwave before modification, (b) pure microwave after modification, (c) mixed point cloud before modification, and (d) mixed point cloud after modification, with corresponding statistics.

Transitioning to Fig.11, which presents a bar chart comparing estimation bias across varying SST ranges, highlighted further insights into our modifications. Initially, the blue bars across diverse SST values exhibited significant positive deviations, reflecting an overall trend of larger estimation errors. In stark contrast, the red bars consistently portrayed smaller deviations, affirming the relative efficiency of the modified model. Notably, in the higher SST ranges, deviations from actual values approached zero for two types of inputs, showcasing improved estimation performance.

(a)

(b)

Fig. 11. Average deviation (estimated minus true SST) across temperature ranges for (a) pure microwave estimations and (b) mixed point cloud estimations

4.6.2 Modified to Sensor-Aware Attention Pooling and Nearest Point Sampling

This experiment builds upon the initial modifications by focusing on refining the mechanisms for feature extraction and point cloud sampling within the PointNet++ architecture. Specifically, we address the method by which the model aggregates features and selects salient points for further processing. Recognizing that the original PointNet++ model employs max pooling to capture distinctive point cloud features, we hypothesized that a shift to sensor-aware attention-based pooling might better suit the nuanced requirements of our regression task. Our attention mechanism explicitly leverages the sensor type information to dynamically weight the contribution of microwave and infrared observations during feature aggregation. Moreover, the standard farthest point sampling (FPS) method was replaced with nearest point sampling (NPS) after centering the point cloud, under the premise that prioritizing data points near the centroid could provide a more efficient means of leveraging point cloud information.

To quantitatively assess the impact of these adjustments, we recorded changes in error metrics before and after implementing the described modifications, focusing on both pure microwave and hybrid point cloud inputs. For pure microwave point cloud inputs, the Mean Absolute Error (MAE) decreased from 0.39 to 0.37, and the Root Mean Square Error (RMSE) decreased from 0.55 to 0.52. These results indicate a modest yet consistent improvement in estimation accuracy. The sensor-aware attention pooling facilitates a more intelligent integration of local point features by learning optimal weighting strategies, while NPS ensures preferential selection of points closest to the centroid, optimizing the model’s focus on the most representative data.

The improvements were more pronounced when considering hybrid point cloud inputs. Before optimization, the MAE was 0.26, and the RMSE was 0.37. After applying the attention-based feature extraction and refined sampling strategies, the MAE substantially decreased to 0.16, and the RMSE fell to 0.20. This significant reduction in error underscores the effectiveness of our modifications in harnessing the additional information provided by the hybrid point cloud. The attention mechanism particularly benefits the hybrid case by automatically learning to balance the contributions of microwave and infrared measurements based on their inherent quality characteristics, giving appropriate weight to the higher-resolution infrared data while maintaining the robustness of microwave observations.

Fig.12 presents histograms comparing the distributions of estimated and true SST values before and after model adjustments. The “After” plots demonstrate a closer alignment between the estimated (blue) and true (red) distributions, with fewer large deviations observed. This visual representation corroborates the quantitative results, highlighting the enhanced estimation accuracy achieved through our modifications.

Fig. 12. Frequency distributions of SST values comparing true observations (red solid lines) and model results (blue dashed lines): (a) pure microwave before modification, (b) pure microwave after modification, (c) mixed point cloud before modification, and (d) mixed point cloud after modification.

Further insights are gained from the violin plots in Fig.13, which illustrate the distributions of absolute errors before and after model modifications for both pure microwave and hybrid point cloud inputs. For pure microwave inputs, the changes are subtle, reflecting the moderate improvements observed in the error metrics. However, for mixed point cloud inputs, the violin plots reveal a marked reduction in the spread of errors after modification. This visual evidence further supports the claim that our sensor-aware attention pooling is particularly effective in fusing the complementary information from different sensor types, leading to substantial improvements in estimation accuracy.

Fig. 13. Violin plots showing absolute error distributions before and after model modification: (a) pure microwave before modification, (b) pure microwave after modification, (c) mixed point cloud before modification, and (d) mixed point cloud after modification.

4.7 Impact of Infrared Data Percentage on SST Estimation Accuracy

This experiment investigates the impact of varying percentages of infrared (IR) data on the accuracy of Sea Surface Temperature (SST) estimation using a hybrid point cloud approach. Cloud masking is a critical step in estimating SST from satellite observations, as clouds interfere with remote sensing data, limiting usable data and creating geographical bias [40]. To simulate different cloud coverage scenarios, a complete IR SST truth map was progressively masked, ranging from 100% (representing a pure microwave point cloud input) to 90%, with hybrid point clouds used as input for model estimation and filling of the IR SST data.

The methodology involved creating eleven distinct masking scenarios, each with a different percentage of IR data masked, ranging from 100% to 90%. At 100% masking, the input consisted solely of microwave data, effectively simulating complete cloud cover over the IR data. As the masking percentage decreased, the model received an increasing amount of IR data within the hybrid point cloud. The model was then tasked with inverting and filling the masked regions, leveraging the available microwave and IR data to reconstruct a complete SST map. The results of these reconstructions across the eleven scenarios are visually presented in Fig.14, demonstrating the model’s performance under different data availability conditions.

Fig. 14. Visual comparison of SST reconstruction with varying IR data availability. Each panel shows results with different masking percentages (100%-90%), with the true value shown in the bottom right.

To quantitatively assess the accuracy of the reconstructed SST maps, the Peak Signal-to-Noise Ratio (PSNR) was calculated for each scenario, comparing the model-filled plots to the original, unmasked truth plots. Higher PSNR values indicate better image quality, signifying minimal differences between the original and processed image [41]. The detailed performance metrics for each masking percentage are documented in Table 3, including PSNR, Mean Bias, Relative Standard Deviation (RSD), and RMSE values.

Table 3. Quantitative assessment of SST estimation performance across different masking levels

Mask Level	PSNR (dB)	Mean Bias (K)	RSD (%)	RMSE (K)
100%	21.4	0.3099	0.15	0.5433
99%	22.51	0.2507	0.14	0.4785
98%	23.63	0.1951	0.13	0.4205
97%	24.64	0.1489	0.12	0.3744
96%	25.39	0.1086	0.11	0.3433
95%	26.43	0.0691	0.1	0.3046
94%	27.23	0.0353	0.1	0.2777
93%	27.89	0.008	0.09	0.2573
92%	28.4	-0.0146	0.08	0.2428
91%	28.88	-0.0347	0.08	0.2297
90%	29.51	-0.0541	0.07	0.2137

 

Table 3 reveals a clear trend: as the masking percentage decreases (more IR data becomes available), the PSNR increases from 21.4 dB at 100% masking to 29.51 dB at 90% masking. Simultaneously, the RMSE decreases from 0.5433 K to 0.2137 K, demonstrating the significant improvement in estimation accuracy with the inclusion of even small amounts of IR data.

To further investigate this relationship, we conducted an additional experiment examining the impact of including specific numbers of IR data points in the model input. Unlike the previous masking experiment which focused on specific regions, this additional experiment was conducted across the entire test dataset to provide a more comprehensive assessment. Table 4 presents the performance metrics as a function of IR points used, from 0 (pure microwave) to 10 points, evaluated across all test samples. The results show a dramatic improvement in accuracy with just a few IR data points, with RMSE decreasing from 0.5213 K with no IR data to 0.2085 K with just 4 IR points.

Table 4. Impact of increasing infrared (IR) data points on estimation accuracy metrics

IR Points	Samples	RMSE(K)	MAE(K)	Mean Bias(K)	RSD(%)
0	55862	0.5213	0.3706	0.0929	0.1732
1	55750	0.4255	0.3029	0.0507	0.1417
2	55633	0.3198	0.2277	0.0015	0.1062
3	55489	0.2490	0.1839	-0.0344	0.0833
4	55326	0.2085	0.1582	-0.0639	0.067
5	55141	0.1951	0.1512	-0.0826	0.0597
6	54950	0.1902	0.1499	-0.0965	0.0553
7	54754	0.1904	0.1508	-0.1021	0.0542
8	54552	0.1911	0.1521	-0.1058	0.0538
9	54308	0.1908	0.1521	-0.1065	0.0534
10	54079	0.1903	0.1519	-0.1067	0.0532

The initial steep decline in RMSE underscores the significant impact of even a small number of IR data points on the accuracy of SST reconstruction. When the number of IR points is minimal, the RMSE is high, indicating substantial discrepancies between the model’s output and the true SST values. However, as the model gains access to just a few IR data points, the RMSE decreases dramatically, suggesting that these initial IR observations provide critical information for constraining the solution space and guiding the model towards a more accurate reconstruction. This rapid improvement highlights the value of even sparse IR data in refining SST estimates derived primarily from microwave observation.

5.Conclusion

In conclusion, this paper presents STPRNet, a point cloud-based deep learning framework that processes Level 2 satellite data as three-dimensional points in space-time. Our methodology preserves the irregular format of instantaneous observations, enabling effective capture of complex spatio-temporal relationships without the limitations of gridded products. By adapting the PointNet++ architecture with a sensor-aware attention fusion mechanism, our framework successfully integrates the complementary strengths of MODIS infrared sensors (high spatial resolution) and AMSR2 microwave instruments (all-weather capability). Experimental results demonstrate the effectiveness of this approach, achieving nighttime root-mean-square errors of 0.52°C with microwave-only point clouds and 0.20°C with hybrid inputs. This advancement enhances real-time, high-resolution SST mapping capabilities crucial for climate monitoring and establishes a flexible paradigm that could extend to other environmental remote sensing applications requiring irregular spatio-temporal data and multi-sensor data integration.

References

1. Wentz, F.J.; Gentemann, C.; Smith, D.; Chelton, D. Satellite Measurements of Sea Surface Temperature through Clouds. Science 2000, 288, 847–850, doi:10.1126/science.288.5467.847.
2. Us Epa, O. Climate Change Indicators: Sea Surface Temperature Available online: https://www.epa.gov/climate-indicators/climate-change-indicators-sea-surface-temperature (accessed on 21 February 2025).
3. Reynolds, R.W.; Rayner, N.A.; Smith, T.M.; Stokes, D.C.; Wang, W. An Improved In Situ and Satellite SST Analysis for Climate. J. Clim. 2002, 15, 1609–1625, doi:10.1175/1520-0442(2002)015<1609:AIISAS>2.0.CO;2.
4. Reynolds, R.W.; Smith, T.M. A High-Resolution Global Sea Surface Temperature Climatology. J. Clim. 1995, 8, 1571–1583, doi:10.1175/1520-0442(1995)008<1571:AHRGSS>2.0.CO;2.
5. Guan, L.; Kawamura, H. Merging Satellite Infrared and Microwave SSTs: Methodology and Evaluation of the New SST. J Oceanogr 2004, 60, 905–912, doi:10.1007/s10872-005-5782-5.
6. Zhu, Y.; Bo, Y.; Zhang, J.; Wang, Y. Fusion of Multisensor SSTs Based on the Spatiotemporal Hierarchical Bayesian Model. Journal of Atmospheric and Oceanic Technology 2018, 35, 91–109, doi:10.1175/JTECH-D-17-0116.1.
7. Zhang, X.; Chen, N. Reconstruction of GF-1 Soil Moisture Observation Based on Satellite and In Situ Sensor Collaboration under Full Cloud Contamination. IEEE Trans. Geosci. Remote Sensing 2016, 54, 5185–5202, doi:10.1109/TGRS.2016.2558109.
8. Guan, L.; Kawamura, H. [No Title Found]. J. Oceanogr. 2003, 59, 201–209, doi:10.1023/A:1025543305658.
9. Tang, S.; Yang, X.; Dong, D.; Li, Z. Merging Daily Sea Surface Temperature Data from Multiple Satellites Using a Bayesian Maximum Entropy Method. Front. Earth Sci. 2015, 9, 722–731, doi:10.1007/s11707-015-0538-z.
10. Reynolds, R.W. Impact of Mount Pinatubo Aerosols on Satellite-Derived Sea Surface Temperatures. J. Clim. 1993, 6, 768–774, doi:10.1175/1520-0442(1993)006<0768:IOMPAO>2.0.CO;2.
11. Chao, Y.; Li, Z.; Farrara, J.D.; Hung, P. Blending Sea Surface Temperatures from Multiple Satellites and In Situ Observations for Coastal Oceans. Journal of Atmospheric and Oceanic Technology 2009, 26, 1415–1426, doi:10.1175/2009JTECHO592.1.
12. Li, A.; Bo, Y.; Zhu, Y.; Guo, P.; Bi, J.; He, Y. Blending Multi-Resolution Satellite Sea Surface Temperature (SST) Products Using Bayesian Maximum Entropy Method. Remote Sens. Environ. 2013, 135, 52–63, doi:10.1016/j.rse.2013.03.021.
13. Bretherton, F.P.; Davis, R.E.; Fandry, C.B. A Technique for Objective Analysis and Design of Oceanographic Experiments Applied to MODE-73. Deep Sea Res. Oceanogr. Abstr. 1976, 23, 559–582, doi:10.1016/0011-7471(76)90001-2.
14. Reynolds, R.W. A Real-Time Global Sea Surface Temperature Analysis. J. Clim. 1988, 1, 75–87, doi:10.1175/1520-0442(1988)001<0075:ARTGSS>2.0.CO;2.
15. Reynolds, R.W.; Smith, T.M. Improved Global Sea Surface Temperature Analyses Using Optimum Interpolation. J. Clim. 1994, 7, 929–948, doi:10.1175/1520-0442(1994)007<0929:IGSSTA>2.0.CO;2.
16. Reynolds, R.W.; Zhang, H.; Smith, T.M.; Gentemann, C.L.; Wentz, F. Impacts of in Situ and Additional Satellite Data on the Accuracy of a Sea‐surface Temperature Analysis for Climate. Int. J. Climatol. 2005, 25, 857–864, doi:10.1002/joc.1168.
17. Thiébaux, J.; Rogers, E.; Wang, W.; Katz, B. A New High-Resolution Blended Real-Time Global Sea Surface Temperature Analysis. Bull. Am. Meteorol. Soc. 2003, 84, 645–656, doi:10.1175/BAMS-84-5-645.
18. Chelton, D.B.; Wentz, F.J. Global Microwave Satellite Observations of Sea Surface Temperature for Numerical Weather Prediction and Climate Research. Bull. Am. Meteorol. Soc. 2005, 86, 1097–1116, doi:10.1175/BAMS-86-8-1097.
19. McIntosh, P.C. Oceanographic Data Interpolation: Objective Analysis and Splines. J. Geophys. Res.: Oceans 1990, 95, 13529–13541, doi:10.1029/JC095iC08p13529.
20. Atkinson, P.M.; Curran, P.J. Defining an Optimal Size of Support for Remote Sensing Investigations. IEEE Trans. Geosci. Remote Sensing 1995, 33, 768–776, doi:10.1109/36.387592.
21. Atkinson, P.M.; Curran, P.J. Choosing an Appropriate Spatial Resolution for Remote Sensing Investigations. 1997.
22. Woodcock, C.E.; Strahler, A.H. The Factor of Scale in Remote Sensing. Remote Sens. Environ. 1987, 21, 311–332, doi:10.1016/0034-4257(87)90015-0.
23. Ducournau, A.; Fablet, R. Deep Learning for Ocean Remote Sensing: An Application of Convolutional Neural Networks for Super-Resolution on Satellite-Derived SST Data. In Proceedings of the 2016 9th IAPR Workshop on Pattern Recogniton in Remote Sensing (PRRS); December 2016; pp. 1–6.
24. Khoo, J.J.D.; Lim, K.H.; Pang, P.K. Deep Learning Super Resolution of Sea Surface Temperature on South China Sea. In Proceedings of the 2022 International Conference on Green Energy, Computing and Sustainable Technology (GECOST); October 2022; pp. 176–180.
25. Lloyd, D.T.; Abela, A.; Farrugia, R.A.; Galea, A.; Valentino, G. Optically Enhanced Super-Resolution of Sea Surface Temperature Using Deep Learning. IEEE Trans. Geosci. Remote Sensing 2022, 60, 1–14, doi:10.1109/TGRS.2021.3094117.
26. Ping, B.; Su, F.; Han, X.; Meng, Y. Applications of Deep Learning-Based Super-Resolution for Sea Surface Temperature Reconstruction. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 14, 887–896, doi:10.1109/JSTARS.2020.3042242.
27. Izumi, T.; Amagasaki, M.; Ishida, K.; Kiyama, M. Super-Resolution of Sea Surface Temperature with Convolutional Neural Network- and Generative Adversarial Network-Based Methods. Journal of Water and Climate Change 2022, 13, 1673–1683, doi:10.2166/wcc.2022.291.
28. Xiao, C.; Hu, C.; Chen, N.; Zhang, X.; Chen, Z.; Tong, X. A Genetic Algorithm–Assisted Deep Neural Network Model for Merging Microwave and Infrared Daily Sea Surface Temperature Products. Front. Environ. Sci. 2021, 9, doi:10.3389/fenvs.2021.748913.
29. Qi, C.R.; Su, H.; Mo, K.; Guibas, L.J. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation 2017.
30. Qi, C.R.; Yi, L.; Su, H.; Guibas, L.J. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space 2017.
31. Nasa/Jpl GHRSST Level 2P Global Sea Surface Skin Temperature from the Moderate Resolution Imaging Spectroradiometer (MODIS) on the NASA Aqua Satellite (GDS2) 2020.
32. Nasa/Jpl GHRSST Level 2P Global Subskin Sea Surface Temperature Version 8.2 (v8.2) from the Advanced Microwave Scanning Radiometer 2 (AMSR2) on the GCOM-W Satellite by REMSS 2023.
33. PO.DAAC | NASA Earthdata Available online: https://www.earthdata.nasa.gov/centers/po-daac (accessed on 21 February 2025).
34. Information (NCEI), N.C. for E. GHRSST Level 2P Global Sea Surface Skin Temperature from the Moderate Resolution Imaging Spectroradiometer (MODIS) on the NASA Terra Satellite (GDS Versions 1 and 2) Available online: https://www.ncei.noaa.gov/access/metadata/landing-page/bin/iso?id=gov.noaa.nodc:GHRSST-MODIS_T-JPL-L2P (accessed on 21 February 2025).
35. Information (NCEI), N.C. for E. GHRSST Level 2P Global Subskin Sea Surface Temperature Version 8a from the Advanced Microwave Scanning Radiometer 2 on the GCOM-W Satellite (GDS Version 2) Available online: https://www.ncei.noaa.gov/access/metadata/landing-page/bin/iso?id=gov.noaa.nodc:GHRSST-AMSR2-REMSS-L2P (accessed on 21 February 2025).
36. Ghrsst_Adm GHRSST – Products. Ghrsst.
37. Wu, Z.; Jiang, C.; Conde, M.; Chen, J.; Deng, B. The Long-Term Spatiotemporal Variability of Sea Surface Temperature in the Northwest Pacific and China Offshore. Ocean Science 2020, 16, 83–97, doi:10.5194/os-16-83-2020.
38. Okura, S.; Hosokawa, S. Representativeness and Certainty of Sea Surface Temperature from MODIS in Semi-Enclosed Bays. Coastal Eng. J. 2022, 64, 428–441, doi:10.1080/21664250.2022.2116775.
39. Zhang, M.; Xu, N.; Chen, L. Fusion SST from Infrared and Microwave Measurement of FY-3D Meteorological Satellite. J. Trop. Meteorol. 2024, 30, 89–96, doi:10.3724/j.1006-8775.2024.009.
40. Agabin, A.; Prochaska, J.X. Reconstructing Sea Surface Temperature Images: A Masked Autoencoder Approach for Cloud Masking and Reconstruction 2023.
41. Hore, A.; Ziou, D. Image Quality Metrics: PSNR vs. SSIM. 2010 20th International Conference on Pattern Recognition 2010, 2366–2369, doi:10.1109/ICPR.2010.579.