Incremented Scence Classification on Satellite Imagery Using Feature Map Distillation and Classifier Discrepancy

This paper presents an incremental learning method that aims to address the challenges of retraining deep neural networks with old data for new categories in remote sensing image classification. While deep learning models can quickly update the existing model with new data, they face challenges such as "catastrophic forgetting" and the "stability-plasticity dilemma." To address these challenges, this study proposes an incremental learning method that uses feature map-based knowledge distillation to enable the new model to memorize the old model's knowledge. The discrepancy loss between multiple classifiers is calculated to enhance performance on new tasks. The proposed method is compared to the classical incremental learning method Learning without Forgetting (LwF) on the RESISC45 dataset. The results show that our method improves the overall accuracy by 3.2% and reduces the forgetting rate by 2.0% compared to LwF after continuously adding one to five tasks. The proposed method can effectively reduce the degree of catastrophic forgetting and provide a feasible solution for remote sensing image classification problems.