Leaf area index (LAI) is one of key variables for depicting vegetation structures in land ecosystems. Land surface models necessitate uniform LAI inputs at varying spatial scales to ensure accurate outputs at multiscale levels, however, operational satellite LAI products are acquired only at low spatial resolutions, inhibiting their application at finer spatial scales. Spatial downscaling methods are beneficial for the spatial enhancement of LAI products, and the emergence of deep learning methods has provided promising options for land surface parameter downscaling. However, the potential of deep learning has not been well explored in LAI downscaling. To address this research gap, this study designed an original hierarchical downscaling approach facilitated by generative adversarial network (GAN), transfer learning (TL), and data augmentation techniques to retrieve LAI at fine spatial resolutions, leveraging multiscale satellite images, and cascading from 500-m to 250-m and then to 30-m scales. First, an improved super-resolution GAN (ISRGAN) model was pre-trained using the GLASS LAI and MOD09Q1 products to bridge the general non-linear relationships of LAI between the 500-m and 250-m resolutions. Subsequently, limited reference LAI images were applied to fine-tune this pre-trained ISRGAN model to address the domain shift in the 250-m resolution LAI estimations. Then, the fine-tuned LAI values and the 30-m resolution LAI reference images were utilized as the ISRGAN inputs to produce fine-resolution LAI maps. Finally, the downscaled LAI values derived from the proposed approach were separately validated against reference LAI maps and field measurements across the 250-m and 30-m resolutions. Results show that the fine-tuned transfer learning technique outperforms the pre-trained ISRGAN model and GLASS LAI, with a lower RMSE (0.78) and higher R2 (0.83) at the 250-m resolution. Moreover, the proposed hierarchical downscaling framework achieves better performances for 30-m resolution LAI estimations, regardless of the validation accuracy (R2 = 0.76; RMSE=0.95) and spatiotemporal distributions, than the ISRGAN model which was directly trained by the 500-m and 30-m resolution images. This study highlights that a hierarchical downscaling is valuable for fine-resolution LAI estimations, which leverages multiscale and multisource satellite observations via deep learning.

中文翻译：

---

通过深度学习利用多源和多尺度观测生成高分辨率叶面积指数的分层降尺度方案


叶面积指数（LAI）是描述陆地生态系统植被结构的关键变量之一。地表模型需要在不同空间尺度上进行统一的 LAI 输入，以确保多尺度水平的准确输出，然而，运行卫星 LAI 产品只能在低空间分辨率下获取，这限制了其在更精细的空间尺度上的应用。空间降尺度方法有利于LAI产品的空间增强，深度学习方法的出现为地表参数降尺度提供了有前景的选择。然而，深度学习在 LAI 降尺度方面的潜力尚未得到充分开发。为了解决这一研究空白，本研究设计了一种原始的分层降尺度方法，该方法由生成对抗网络 (GAN)、迁移学习 (TL) 和数据增强技术促进，以精细空间分辨率检索 LAI，利用多尺度卫星图像，并从 500 个卫星图像进行级联-m 到 250 米，然后到 30 米尺度。首先，使用 GLASS LAI 和 MOD09Q1 产品对改进的超分辨率 GAN (ISRGAN) 模型进行预训练，以弥合 LAI 在 500 米和 250 米分辨率之间的一般非线性关系。随后，应用有限的参考 LAI 图像来微调此预训练的 ISRGAN 模型，以解决 250 米分辨率 LAI 估计中的域偏移问题。然后，微调的 LAI 值和 30 米分辨率的 LAI 参考图像被用作 ISRGAN 输入，以生成精细分辨率的 LAI 地图。最后，根据参考 LAI 地图和 250 米和 30 米分辨率的现场测量分别验证了从所提出的方法得出的缩小的 LAI 值。 结果表明，微调的迁移学习技术优于预训练的 ISRGAN 模型和 GLASS LAI，在 250 米分辨率下具有较低的 RMSE (0.78) 和较高的 R2 (0.83)。此外，与直接由 500 米和30 米分辨率的图像。这项研究强调，分层降尺度对于精细分辨率 LAI 估计很有价值，它通过深度学习利用多尺度和多源卫星观测。