Solid and progressive early learning for geospatial scene classification

The task of geospatial scene classification involves dividing images into meaningful land use and land cover categories based on the feature information contained. Deep neural networks have achieved significant success, with their excellent performance relying on a large number of accurately labeled samples. However, collecting such a large-scale dataset with precise labels is impractical. Real-world datasets often contain noise, which is commonly referred to as label noise. To address this issue, the early stopping method based on the network's memorization effect is employed, aiming to terminate training early to reduce the model's learning of noise. While applying early stopping effectively reduce noise interference, the models also inhibit the further learning of clean samples. To enhance the model's ability to learn from clean samples during training, this paper proposed a new learning approach, Solid and progressive early learning (SPEL), whose core lies in adopting a progressive updating strategy to gradually strengthen the ability to learn from clean samples while effectively inhibiting the learning of noise. We have established strict baseline standards and model evaluation strategies. In addition, SPEL was compared with other advanced methods and achieved state-of-the-art results.