Downscaling of MODIS NDVI by Using a Convolutional Neural Network-Based Model with Higher Resolution SAR Data
Downscaling of MODIS NDVI by Using a Convolutional Neural Network-Based Model with Higher Resolution SAR Data
复制标题
DOI：
10.3390/rs13040732
复制
发表时间：
2021-02
期刊：
影响因子：
--
通讯作者：
Ryota Nomura;K. Oki
中科院分区：
文献类型：
--
作者：
Ryota Nomura;K. Oki
归一化植被指数（NDVI）是一种简单而有效的观测绿色活体植被的指标。自1970年代引入以来，归一化差异植被指数已被广泛用于土地管理、粮食安全和物理模型。对于这些应用，以高空间分辨率和高时间分辨率获取NDVI是优选的。然而，在使用卫星图像时，时间分辨率和空间分辨率之间一般都有一个折衷。为了解决这个问题，在这项研究中提出了一个基于卷积神经网络（CNN）的降尺度模型。该模型能够利用Sentinel-1 10米分辨率合成孔径雷达（SAR）数据，从MODIS（中分辨率成像光谱仪）250米分辨率的NDVI中估算出10米高分辨率的NDVI。首先，通过使用2019年在目标区域获取的数据，对该降尺度模型进行训练，以根据升尺度的250米分辨率Sentinel-2 NDVI和10米分辨率Sentinel-1 SAR数据的组合来估计Sentinel-2 10米分辨率的NDVI。通过对2020年实测数据的分析，验证了该模型的通用性，结果表明，该模型对NDVI的预测精度较高（平均MAE = 0.090，ρ = 0.734）。接下来，250米的NDVI从MODIS数据被用来作为输入，以确认该模型的条件下复制一个实际的应用案例。尽管原始MODIS数据和Sentinel-2数据存在不匹配，但模型对NDVI的预测精度仍然可以接受（平均MAE = 0.108，ρ = 0.650）。最后，利用2020年1月1日~ 2020年12月31日的MODIS和Sentinel-1数据，应用该模型对高空间分辨率的NDVI进行了预测。利用高空间分辨率（约×2.5）的NDVI影像增强时间分辨率，观测到了在原始MODIS分辨率下无法观测到的甘蓝复种现象。所提出的方法，使生产10米分辨率的NDVI数据具有可接受的精度时，无云的MODIS NDVI和哨兵-1 SAR数据，并可以提高时间分辨率的高分辨率10米的NDVI数据。
The normalized difference vegetation index (NDVI) is a simple but powerful indicator, that can be used to observe green live vegetation efficiently. Since its introduction in the 1970s, NDVI has been used widely for land management, food security, and physical models. For these applications, acquiring NDVI in both high spatial resolution and high temporal resolution is preferable. However, there is generally a trade-off between temporal and spatial resolution when using satellite images. To relieve this problem, a convolutional neural network (CNN) based downscaling model was proposed in this research. This model is capable of estimating 10-m high resolution NDVI from MODIS (Moderate Resolution Imaging Spectroradiometer) 250-m resolution NDVI by using Sentinel-1 10-m resolution synthetic aperture radar (SAR) data. First, this downscaling model was trained to estimate Sentinel-2 10-m resolution NDVI from a combination of upscaled 250-m resolution Sentinel-2 NDVI and 10-m resolution Sentinel-1 SAR data, by using data acquired in 2019 in the target area. Then, the generality of this model was validated by applying it to test data acquired in 2020, with the result that the model predicted the NDVI with reasonable accuracy (MAE = 0.090, ρ = 0.734 on average). Next, 250-m NDVI from MODIS data was used as input to confirm this model under conditions replicating an actual application case. Although there were mismatch in the original MODIS and Sentinel-2 NDVI data, the model predicted NDVI with acceptable accuracy (MAE = 0.108, ρ = 0.650 on average). Finally, this model was applied to predict high spatial resolution NDVI using MODIS and Sentinel-1 data acquired in target area from 1 January 2020~31 December 2020. In this experiment, double cropping of cabbage, which was not observable at the original MODIS resolution, was observed by enhanced temporal resolution of high spatial resolution NDVI images (approximately ×2.5). The proposed method enables the production of 10-m resolution NDVI data with acceptable accuracy when cloudless MODIS NDVI and Sentinel-1 SAR data is available, and can enhance the temporal resolution of high resolution 10-m NDVI data.