Hongrui Ye

The accuracy of crop mapping, based on hyperspectral images collected by remote sensing, has been significantly improved through the use of deep learning. However, traditional deep learning can be computationally intensive, requiring millions of parameters, which can make it `expensive' to utilize and optimize. Inspired by research on natural language processing, we consider the spectral signature corresponding to each pixel as a `spectral text.' Specifically, we initially feed the hyperspectral (HS) images into the proposed Channel2Vec module to convert them into channel embeddings. Then we use a lossless compressor and Normalized Compression Distance (NCD) to create a dynamic Spectral Tokenizer. It tokenizes `spectral text' into `spectral tokens', extracting the average reflectance values within the token as local sequence information. Finally, we concatenate the local sequence information with the original HSI data to generate spectral embeddings. We then use the lossless compressor again to identify the regularity between spectral embeddings, which we classify based on NCD with the \(k\)-nearest-neighbor classifier (\(k\)NN). Without any trainable parameters, our method achieved results that are competitive with deep learning methods across three sub-scenes of the UAV-HSI-Crop dataset. It even outperforms the average of 8 advanced deep learning methods. Based on the established framework, explorations are conducted from the perspectives of tokenizer settings, compressor selection, and so on. These experiments further validate the effectiveness, feasibility, and lightweight nature of our framework.