Hyperspectral remote sensing is a new remote sensing technology that emerged in the early 1980s. Hyperspectral data has the advantages of fine spectral resolution, numerous bands, and wide spectral range. It provides almost continuous spectral curves for each pixel. Its rich spectral information of ground objects can broaden the scope and enhance the depth of remote sensing applications and improve the accuracy and reliability of quantitative analysis. In recent years, hyperspectral technology has developed rapidly in China. The launch of hyperspectral satellites such as GF-5, ZY-1 02D, GF-5B, and ZY-1 02E has enriched abundant hyperspectral data sources and has greatly promoted the development of hyperspectral remote sensing in China. However, hyperspectral satellites, like multispectral satellites, will inevitably be affected by clouds and cloud shadows in the imaging process. Thick clouds in the atmosphere totally cover the reflected surface information, while thin clouds attenuate the reflection of the surface, such as cirrus clouds and haze. Cloud shadow will also degrade the image quality. Therefore, how to accurately identify clouds and cloud shadows has become the key to ensuring the level of further applications. An improved method is proposed to detect clouds and cloud shadows based on domestic hyperspectral satellites.

As a mature cloud detection algorithm, the Fmask algorithm has been widely used and has become the operational algorithm of Landsat and Sentinel product systems. In this algorithm, clouds and cloud shadows are recognized by multiple threshold criteria and flood filling, respectively. Finally, it uses similarity matching to reconfirm cloud shadows, and the detection accuracy of clouds and cloud shadows for Landsat can reach 96.41% and 70%, respectively. However, previous studies have revealed that the detection accuracy of Fmask is relatively low and limited for data without thermal infrared bands. For example, the cloud and cloud shadow detection accuracy of Sentinel-2 data is about 89% and 50%, respectively. It is much lower than the accuracy of Fmask applied to multispectral data. Therefore, an improved Fmask algorithm is proposed specifically for domestic hyperspectral satellites. We optimize the structure of cloud and cloud shadow detection procedures on the basis of the original Fmask algorithm. For urban areas prone to cloud-detected confusion, we add auxiliary judgments to detect bright ground objects. At the same time, the improved algorithm can distinguish the cloud shadow from terrain shadows and improve the accuracy accordingly. 20 hyperspectral images of GF-5 and ZY-1 02D are used to verify the improved algorithm, covering three typical classes, such as urban, mountainous, and flat areas.

The experimental results indicate that the improved Fmask algorithm performs well in cloud and cloud shadow recognition, highly consistent with the visual recognition results under various underlying surfaces (Fig. 7). The improved Fmask algorithm is compared with the original Famsk algorithm and the other two algorithms, and the cloud and cloud shadow recognition accuracy of all algorithms are calculated, in terms of the overall accuracy, user accuracy, and producer accuracy. The user accuracy and producer accuracy of the improved Fmask algorithm for cloud detection can reach 91.26% and 99.97%, respectively, while the accuracy of cloud shadow detection can reach 78.66% and 79.41%, respectively. The accuracy of all algorithms is illustrated by thermal diagrams (Fig. 8). Evidently, the accuracy of the improved Fmask algorithm is significantly better than the original Fmask algorithm for the scenes containing cities and mountains. Compared with the other two threshold-based algorithms, the improved Fmask algorithm shows significant superiority in aspects of clouds and cloud shadows.

This work improves the Fmask algorithm in cloud and cloud shadow recognition to make it suitable for domestic hyperspectral data. The improved algorithm has been tested in 20 hyperspectral images containing typically different underlying surfaces, and the results are highly consistent with the visual recognition. It is also significantly higher than the validation algorithm in terms of user accuracy and producer accuracy in over 60% of images. The improved Fmask algorithm has advantages in terms of robustness, high accuracy, and versatility. The cloud and cloud shadow recognition procedures include an adjustable threshold, which makes the algorithm more flexible to meet different requirements for cloud and cloud shadow recognition. In addition, the improved algorithm does not need extra auxiliary data, running fast and implementing easily. It can be used for high-precision identification of clouds and cloud shadows in hyperspectral data and enable the operational processing of domestic hyperspectral satellite data.