Adaptive Early-Exit Inference in Graph Neural Networks Based Hyperspectral Image Classification

Hyperspectral image (HSI) classification is a prominent and active research topic in the field of remote sensing. The unique capabilities of hyperspectral imaging, which captures detailed spectral information across a wide range of wavelengths, make it a valuable tool for various applications. Recently, graph neural networks (GNNs) demonstrated outstanding performance in classifying HSIs. Due to the issue of performance deterioration, as the number of layers increases, the GNN models developed for HSI classification are shallow-layer models. However, multi-layer or deep GNNs can extract more HSI features and associations from distant neighbors and classify them more accurately. At the same time, the characteristics and complexity of the HSIs in a dataset vary, and a significant portion of the HSIs can be classified using fewer graph neural layers. Only a few complex HSIs necessitate the execution of all layers for an accurate prediction. This paper introduces an input-adaptive early-exit deep GNN model for HSI classification in which exit branches are added to the intermediate layers of the standard GNN model, allowing simple HSIs to exit early with correct prediction while complex images propagate further up to the final layer. The proposed model, EEGNN, is evaluated on popular HSI datasets such as the Indian Pines dataset and the Pavia University dataset, and it significantly reduces computational costs and execution time while improving accuracy.