Deep Self-Supervised Image Denoising for Joint Hyperspectral-Lidar Classification

The land cover classification is an essential task in remote sensing that assigns categories to pixels in a scene. Until now, it has been mainly solved using one single modality, which experiences limitations in complex scenes that contain different classes with similar spectral properties. Over the past few years, researchers have precisely solved the above challenge by analyzing hyperspectral and light detection and ranging data together. The analysis mainly involves the combination of image-level features acquired during their fusion and subsequent classification. Alternatively, the current method uses image denoising as self-supervised task for learning image-level features on both modalities separately. The denoising task ensures the extraction of the most relevant features from each data source. Afterwards, two instances of a ResNet50-based [1] classifier employ the individually acquired self-supervised features to learn with only a fraction of the available labels, encouraging their efficient usage. A decision fusion module takes each learned classifier’s weights and joins their individual decisions employing a weighted soft voting strategy. The validation process considers two benchmark datasets. Extensive experiments show that the representations learned through image denoising help the classifications to achieve results as accurate as the outcomes obtained using pre-trained ImageNet [2] weights. In addition, the method accomplishes proficient results without requiring manual labeling during representation learning, saving valuable human resources.