**1. Introduction**

Severe convective weather refers to convective weather accompanied by thunderstorm, gale, hail, tornado, local heavy precipitation, and other severe weather phenomena. It is a typical small to medium scale disastrous weather event that seriously threatens the safety of aviation, ship navigation, and occurs frequently in the sea [1–3]. The accurate monitoring and forecasting of severe convective weather are difficult and significant [4]. At present, one of the main means to monitor severe convective weather is achieved by monitoring radar echoes. Radar composite reflectivity >35 dBZ is generally considered as an indicator of the occurrence of severe convective weather [5]. However, in some regions, such as oceans, radar cannot be deployed.

**Citation:** Yu, X.; Lou, X.; Yan, Y.; Yan, Z.; Cheng, W.; Wang, Z.; Zhao, D.; Xia, J. Radar Echo Reconstruction in Oceanic Area via Deep Learning of Satellite Data. *Remote Sens.* **2023**, *15*, 3065. https://doi.org/10.3390/ rs15123065

Academic Editor: Silas Michaelides

Received: 7 April 2023 Revised: 15 May 2023 Accepted: 26 May 2023 Published: 12 June 2023

**Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

It has been shown that radar echoes (e.g., composite reflectivity, vertically integrated liquid), precipitation, and other data can be inverted to monitor severe convective weather based on satellite data with wide coverage [6]. For example, some scholars proposed the Geostationary Operational Environmental Satellites (GOES) Precipitation Index (GPI) by using the physical properties of cold and warm clouds to establish the relationship between cloud top infrared temperature and rainfall probability and intensity [7–9]. Then, in order to improve the inversion accuracy, the results obtained by the GPI method are accumulated over a longer time scale [10,11]. Further, researchers introduced more characteristic variables, such as relative humidity and precipitable water, and developed the Hydro-Estimator algorithm [12]. On this basis, some scholars used exponential functions and quadratic curves to estimate the rainfall intensity, and improved the satellite inversion precipitation algorithm using humidity correction factors and cloud growth rate correction factors [9,13]. Traditional satellite inversion methods are usually based on the understanding of physical processes, and rely on parametric relationships between cloud properties and rainfall and convective processes [14].

With the development of artificial intelligence science and technology, machine learning algorithms have been gradually introduced into the field of atmospheric science in the context of meteorological big data. Machine learning has nonlinear mapping capability and is good at finding patterns in input and output signals, which can better solve nonlinear problems compared with traditional statistical regression methods [15]. Several studies have shown that models based on deep learning network structures outperform traditional methods in experiments to invert severe convective weather [16]. For example, some scholars have conducted preliminary research on precipitation reconstruction based on artificial neural networks (ANN), and the results show that the performance of ANN satellite inversion algorithms is superior to traditional linear methods [17,18]. Later, with the emergence of convolutional neural networks (CNN) [19], more and more scholars have used CNN to invert precipitation and vertically integrated liquid [20,21], demonstrating the effectiveness of CNN in fusing spatial data under different underlying surfaces, and combining data with the physical multichannel inputs in infrared spectroscopy precipitation estimations. On this basis, for the improvement and development of CNN, U-Net is widely used in the field of image segmentation [22]. The U-Net-based reconstruction algorithm is also used to reconstruct radar reflectivity fields to improve short-term convective-scale forecasts of high impact weather hazards and to identify the location, shape, and intensity of convective systems [23–27].

However, most studies that use satellite information to reconstruct data such as radar echoes and precipitation to monitor severe convective weather are based on data from the land area in order to construct the reconstruction model. These studies have defaulted that the models can be applied directly to the oceanic area, whereas there is no assessment of the applicability of the reconstruction model to the ocean. Due to the existence of underlying surface differences, differences in climate situations, lightning, and storm characteristics can occur. This indicates that it is not rigorous to apply the satellite reconstruction model constructed by datasets in non-maritime regions to the oceans directly. However, there are also many problems if radar data from ocean are directly used for data reconstruction. Given that the radar base stations are located on land, with the increase in offshore distance, the elevation of radar detection radiation is too high, and the composite reflectivity of the area far from the radar base station only contains a small amount of the basic reflectivity factors of elevation, which is biased from the real data [28]. The accuracy of radar data in oceanic surface is affected. Therefore, it is urgent to find a data reconstruction method suitable for the ocean.

In addition, with the rapid development of deep learning, it is difficult for us to understand the deep learning model and fully trust it. Therefore, the interpretability of the model has also been highly valued by scholars. In 2004, the academic community proposed the concept of interpretable artificial intelligence [29]. After that, methods of interpretable research such as Local Interpretable Model-Agnostic Explanations (LIME), Layer-wise Relevance Propagation (LRP), Shapley Additive Explanation (SHAP), saliency map, attention mechanism, and DeepLIFT were proposed [30–36]. In previous reconstruction studies, few studies have focused on the differences in the feature importance of models when underlying surface conditions change. Due to the differences between land and ocean, it is extremely necessary to conduct interpretable research on deep learning models generated on different underlying surfaces.

In this study, we build deep learning models for the reconstruction of composite reflectivity from satellite bright temperature data using U-Net with jump connections under different underlying surfaces (land, coast, offshore, and sea). The accuracy is compared to derive a deep learning reconstruction method that is relatively more applicable to the ocean. Then, the importance of the features on different underlying surfaces is analyzed to obtain an interpretable reconstruction model. The model achieves more accurate and credible monitoring of severe convective weather in remote areas without radar deployment.
