Noise disruption inspired NAS with efficient focal attention for hyperspectral image classification

Neural Architecture Search automatically optimizes neural network architectures, reducing the need for manual design and expert experience. While effective for high-dimensional hyperspectral data, Neural Architecture Search faces challenges like long processing times and reduced robustness with noisy labels. To address these, this paper propose Efficient Focal Attention Neural Architecture Search, which integrates the Efficient Channel Attention module and Focal Loss function within a multi-scale spectral attention Neural Architecture Search framework to enhance model convergence and robustness. The Efficient Channel Attention module focuses on key channels to improve feature capture in hyperspectral data, while the Focal Loss function reduces the impact of noisy labels by prioritizing hard-to-classify samples. The Enhanced Position and Scale Aware mechanism further improves adaptability and accuracy by considering position and scale variations. Efficient Focal Attention Neural Architecture Search shows superior performance in handling noisy labels, as demonstrated by experimental results on the Xuzhou dataset, outperforming other deep learning models. This validates the potential of Neural Architecture Search in hyperspectral image classification and offers insights for other complex datasets.