A crop’s spectral signature is worth a compressive text

Abstract

The accuracy of crop mapping based on remotely sensed hyperspectral imagery 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 deploy and optimize. Inspired by studies in natural language processing, we consider the spectral signature corresponding to each pixel as text. Specifically, we first feed the hyperspectral image (HSI) data into the Channel2Vec module to generate channel embeddings. Based on the channel embeddings, we use a lossless compressor and Normalized Compression Distance (NCD) to create a spectral tokenizer. It can segment the spectral signature corresponding to each pixel into multiple windows along the channel dimension, and then extract local sequence information from each window. By combining the local sequence information with the original HSI data, we construct spectral embeddings. Finally, we again use the lossless compressor to compute the NCD between the spectral embeddings, and then classify using only the $k$-nearest-neighbor classifier ($k$NN). The proposed framework is ready-to-use and lightweight. Without any training, it achieves results competitive with deep learning models on three benchmark datasets. It outperforms the average of 11 advanced deep learning methods trained at scale. Moreover, it outperforms more than half of these models in the few-shot scenario, where there are not enough labels to effectively train a neural network.