High-quality and continuous satellite data are essential for land surface studies such as monitoring of land surface phenology, but factors such as cloud contamination and sensor malfunction degrade the quality of remote sensing images and limit their utilization. Filling gaps and recovering missing information in time series of remote sensing images are vital for a wide range of downstream applications, such as land surface phenology extraction. Most existing gap-filling and cloud removal methods focus on individual or multi-temporal image reconstruction, but struggle with continuous and overlapping missing areas in time series data. In this study, we propose a Temporal Attention-Based Multi-Scale Generative Adversarial Network (TAMGAN) to reconstruct time series of Moderate Resolution Imaging Spectroradiometer (MODIS) data. TAMGAN leverages a Generative Adversarial Network (GAN) with a 3-dimensional Convolution Neural Networks (3DCNN) in its generator to reconstruct the missing areas in the annual time series of remote sensing images simultaneously. The temporal attention blocks are designed to capture the changing trends of surface reflectance over time. And multi-scale feature extraction and progressive concatenation are introduced to enhance spectral consistency and provide detailed texture information. Experiments are carried out on MOD09A1 products to evaluate the performance of the proposed network. The results show that TAMGAN outperformed the comparison methods across various evaluation metrics, particularly in handling large and continuous missing areas in the time series. Furthermore, we showcase an example of downstream application by extracting phenological information from the gap-filled products. By effectively filling gaps and removing clouds, our method offers spatial-temporal continuous MODIS surface reflectance data, benefiting downstream applications such as phenology extraction and highlighting the potential of artificial intelligence technique in remote sense data processing.

中文翻译：

---

一种基于时间注意力的多尺度生成对抗网络，用于填补 MODIS 数据时间序列中的空白，用于地表物候提取


高质量和连续的卫星数据对于地表物候监测等地表研究至关重要，但云污染和传感器故障等因素会降低遥感图像质量并限制其利用。填补遥感图像时间序列中的空白和恢复缺失信息对于广泛的下游应用至关重要，例如地表物候提取。大多数现有的间隙填充和云移除方法都侧重于单个或多时态图像重建，但难以处理时间序列数据中连续和重叠的缺失区域。在这项研究中，我们提出了一个基于时间注意力的多尺度生成对抗网络 （TAMGAN） 来重建中分辨率成像光谱仪 （MODIS） 数据的时间序列。TAMGAN 在其生成器中利用生成对抗网络 （GAN） 和 3 维卷积神经网络 （3DCNN） 同时重建遥感图像年度时间序列中的缺失区域。时间注意力块旨在捕获表面反射率随时间的变化趋势。引入多尺度特征提取和渐进级联，增强光谱一致性，提供详细的纹理信息。对 MOD09A1 产品进行实验，以评估所提出的网络的性能。结果表明，TAMGAN 在各种评估指标上都优于比较方法，尤其是在处理时间序列中较大且连续的缺失区域方面。此外，我们通过从间隙填充的产品中提取物候信息，展示了一个下游应用的例子。 通过有效填补空白和去除云层，我们的方法提供了时空连续的 MODIS 表面反射数据，有利于物候提取等下游应用，并突出了人工智能技术在遥感数据处理中的潜力。