Mapping water bodies under cloud cover using remotely sensed optical images and a spatiotemporal dependence model

Optical remote sensing imagery is commonly used to monitor the spatial and temporal distribution patterns of inland waters. Its usage, however, is limited by cloud contamination, which results in low-quality images or missing values. Selecting cloud-free scenes or combining multi-temporal images to produce a cloud-free composite image can partially overcome this problem at the cost of the monitoring frequency. Predicting the spectral values of cloudy areas based on the spectral characteristics is a possible solution; however, this is not appropriate for water because it changes rapidly. Reconstructing cloud-covered water areas using historical water-distribution data has good performance, but such methods are typically only suitable for lakes and reservoirs, not over vast and complex terrain. This paper proposes a category-based approach to reconstruct the water distribution in cloud-contaminated images using a spatiotemporal dependence model. The proposed method predicts the class label (water or land) of a cloudy pixel based on the neighboring pixel labels and those at the same position in images acquired on other dates according to historical spatiotemporal water-distribution data. The method was evaluated through eight experiments in different study regions using Landsat and Sentinel-2 images. The results demonstrated that the proposed method could yield high-quality cloud-free classification maps and provide good water-extraction accuracy and consistency in most hydrological conditions, with an overall accuracy of up to 98%. The accuracy and practicality of the method render it promising for applications across a wide range of future research and monitoring efforts.

Automated summary

This paper proposes a new method to reconstruct the water distribution in cloud-contaminated images using a spatiotemporal dependence model. The results demonstrate that the method can produce high-quality cloud-free classification maps and provide good water-extraction accuracy and consistency in various hydrological conditions, with an overall accuracy of up to 98%. The accuracy and practicality of the method make it promising for future research and monitoring efforts across a wide range of applications.