An efficient method for cloud detection based on the feature-level fusion of Landsat-8 OLI spectral bands in deep convolutional neural network

Cloud segmentation is a critical pre-processing step for any multi-spectral satellite image application. In particular, disaster-related applications e.g., flood monitoring or rapid damage mapping, which are highly time and data-critical, require methods that produce accurate cloud masks in a short time while being able to adapt to large variations in the target domain (induced by atmospheric conditions, different sensors, and scene properties). This research presented a deep convolutional neural network for cloud detection in the Landsat-8 dataset at the pixel level. Two key components of the proposed network are convolutional layers in the decoder branch with two convolution kernels in various scales. The near-infrared band in this study was added to the network inputs, including red, green, and blue bands, to improve the network performance. In the proposed network architecture, the encoder-decoder branches symmetrically with the density of feature maps resulting from the multiplicity of filters and the design of multi-dimension filters, providing a local and general context for accurate identification of the cloud and its margins to extracted spatial features in high-level scales are used. However, Multi-scales feature maps will be sampled and integrated and used to generate output with high accuracy. Finally, the proposed method uses 3500 patches of Landsat-8 satellite images with various cloud challenges by using several kernels in sizes 3 x 3 and 5 x 5 with an F1-score of 96.6 and a Jaccard index (JI) of 93.5 provides higher accuracy than other methods. In general, the suggested method outperformed the alternatives in the same, uncorrected data set in terms of accuracy, particularly in regions with bright surfaces. Due to the effectiveness of the proposed framework, it has a lot of potential for practical application with different types of satellite images.