Improving Fractional Vegetation Cover Estimation With Shadow Effects Using High Dynamic Range Images

Measured fractional vegetation cover (FVC) on the ground is very important for validation of the remote sensing products and algorithms. However, because of the influence of some factors such as the angle of illumination and vegetation density, the existence of vegetation shadows limits the accuracy of FVC estimation. This article proposes a deep learning method to reduce the FVC estimation error based on high dynamic range (HDR) images with vegetation shadows (HDR REC-DL method). The HDR REC-DL method can accurately extract FVC from HDR images with complex texture information on vegetation shadows. This method is based on the U-Net convolutional network structure for semantic segmentation of images containing vegetation shadows, and the segmentation results are less affected by vegetation types. Results from the HDR REC-DL method were highly similar to the vegetation segmentation results from visual interpretation. Values of the kappa coefficient, F1 score (F1), recall, and mean intersection over union of the HDR REC-DL method were 0.926, 0.942, 0.924, 0.916 for sunny weather and 0.903, 0.974, 0.983, and 0.895 for cloudy weather, respectively. Compared with the vegetation segmentation accuracy of the shadow-resistant algorithm, the HDR REC-DL method increases the kappa coefficient, F1, and mIOU by 21%, 16%, and 29% for sunny weather, and by 11.1%, 3.6%, and 10.3% for cloudy weather, respectively. The HDR REC-DL method provides a novel method for accurately estimating FVC from images containing vegetation shadows.