Normalized difference vegetation index (NDVI), derived from multi-spectral (MS) images, is a metric widely used to evaluate the growth status of vegetation in remote sensing. Existing methods for generating high-resolution (HR) NDVI are typically based on pan-sharpening, which often result in huge errors even in case of tiny spectral distortions. To overcome this challenge, from a novel perspective, this paper introduces an HR vegetation index (HRVI) to realize direct fusion with a low-resolution NDVI rather than pan-sharpening an HRMS image. In particular, we propose a two-branch network based on the multi-scale and attention mechanism, termed as NDVI-Net, to obtain the HRNDVI with small distortion. In our network, the multi-scale channel enhancement blocks are used in both NDVI and HRVI branches, in which multi-scale convolution is used to capture structural information with different reception fields and channel attention mechanism is adopted to perform feature selection. Meanwhile, the spatial features are injected unidirectionally from the HRVI into NDVI branches, so as to further improve the quality of features in the NDVI branch. Subsequently, the spatial intensify block is adopted only in the NDVI branch to implement selective enhancement for the previously obtained features along the spatial position, strengthening the retention of local detail features. Finally, HRNDVI is reconstructed based on the high-representation NDVI features, which contains clear texture details and precise intensity. Experimental results demonstrate the significant advantage of our method over the current state-of-the-art in terms of both subjective visual effect and quantitative metrics. Moreover, we apply the HRNDVI generated by our method to vegetation detection and enhancement, and land cover mapping in remote sensing, which can achieve the best performance.

来自多光谱（MS）图像的归一化植被指数（NDVI）是一种广泛用于评估遥感中植被生长状况的指标。现有的用于生成高分辨率（HR）NDVI的方法通常基于泛锐化，即使在频谱失真很小的情况下，也经常会导致巨大的误差。为了克服这一挑战，本文从新颖的角度介绍了一种HR植被指数（HRVI），以实现与低分辨率NDVI的直接融合，而不是对HRMS图像进行全景锐化。特别是，我们提出了一种基于多尺度和注意力机制的两分支网络，称为NDVI-Net，以获得失真较小的HRNDVI。在我们的网络中，NDVI和HRVI分支都使用了多尺度通道增强模块，其中多尺度卷积用于捕获具有不同接收场的结构信息，并采用信道关注机制进行特征选择。同时，将空间特征从HRVI单向注入NDVI分支，以进一步提高NDVI分支的特征质量。随后，仅在NDVI分支中采用空间增强块，以沿空间位置对先前获得的特征进行选择性增强，从而增强对局部细节特征的保留。最后，HRNDVI是基于具有高代表性的NDVI特征而重建的，该特征包含清晰的纹理细节和精确的强度。实验结果表明，就主观视觉效果和定量指标而言，我们的方法相对于当前的最新技术具有显着优势。此外，我们将通过我们的方法生成的HRNDVI应用于植被检测和增强，以及遥感中的土地覆被制图，可以达到最佳性能。