Traditional Neural Networks (NN) have been popular in the satellite remote sensing community for the last 25 years. For coarse resolution infrared or microwave instruments, NN algorithms have been used at the pixel level. New neural architectures such as Convolutional Neural Networks (CNN) use the Deep Learning (DL) approach to solve complex problems at the image scale. For instance, CNNs have been applied to high resolution instruments (SAR or in the visible domain) to detect surface waters or vegetation. High resolution data is better suited for image processing techniques because spatial features are stronger and pixel noise can be an important issue. CNNs applications are generally related to image classification or segmentation, less for regression problems dealing with the estimation of a variable in each pixel of the image. The objective of this paper is to better understand how and on which conditions CNNs work, and how beneficial they can be for coarse resolution instruments such as IASI (Infrared Atmospheric Sounding Interferometer). The CNN and DL approaches are tested in a regression mode, to estimate the Sea Surface Temperature (SST) at the image scale. The CNN technique is compared to a traditional pixel-based NN: both have a SST retrieval error of 0.3 K. An instrument noise and a missing data sensitivity studies are conducted. It is shown that the use of the CNN approach in this simple-experiment context is beneficial only under some conditions: when the variable to retrieve has enough spatial coherency (simple smoothness or presence of spatial features in the images), and when the instrument noise at the pixel scale is larger than a threshold. This study is a preliminary illustration of what can be expected from CNNs for coarse resolution instruments such as IASI.

在过去的 25 年中，传统的神经网络 (NN) 在卫星遥感界一直很流行。对于粗分辨率红外或微波仪器，已在像素级使用神经网络算法。卷积神经网络 (CNN) 等新神经架构使用深度学习 (DL) 方法来解决图像尺度的复杂问题。例如，CNN 已应用于高分辨率仪器（SAR 或在可见域中）以检测地表水或植被。高分辨率数据更适合图像处理技术，因为空间特征更强，像素噪声可能是一个重要问题。CNN 的应用通常与图像分类或分割有关，较少涉及处理图像每个像素中变量估计的回归问题。本文的目的是更好地了解 CNN 的工作方式和条件，以及它们对 IASI（红外大气探测干涉仪）等粗分辨率仪器有何好处。CNN 和 DL 方法在回归模式下进行测试，以估计图像尺度的海面温度 (SST)。将 CNN 技术与传统的基于像素的 NN 进行比较：两者的 SST 检索误差均为 0.3 K。进行了仪器噪声和缺失数据敏感性研究。结果表明，在这种简单的实验环境中使用 CNN 方法仅在某些条件下是有益的：当要检索的变量具有足够的空间连贯性（图像中的简单平滑或存在空间特征）时，以及当仪器噪声在像素尺度上大于阈值。