ConDA: Continual Adaptation in Remote Sensing Via Visual Style Playback

This study focuses on continual adaptation in remote sensing semantic segmentation, addressing challenges posed by frequent data updates and model forgetting. Remote sensing images exhibit variations in visual styles due to factors like location, time, and weather conditions, creating distinct domains. To counter performance degradation in new domains, we introduce a new challenge task in remote sensing, termed Continual Domain Adaptation (ConDA). Our innovative Visual Style Replay method employs Variational Auto-Encoder (VAE) and knowledge distillation, enabling the model to continuously learn from historical domains without forgetting. The proposed approach, tested on the INRIA dataset under ConDA settings, outperforms existing methods in combating catastrophic forgetting in remote sensing segmentation tasks. This contributes to adapting deep learning models to real-world scenarios with continually evolving remote sensing data.

1. INTRODUCTION

Deep learning-based methods have shown remarkable performance in high-resolution remote sensing image interpretation tasks in recent years. However, remote sensing data undergoes frequent updates [1], requiring continuous model adaptations to accommodate new data [2]. Nevertheless, deep learning models often forget the originally trained content when faced with continually updated data [3]. This challenge arises from inherent variations in visual styles [4] among different batches of remote sensing images, attributed to changes in geographical locations, capture times, angles, and photographic weather conditions, essentially presenting distinct domains [5]. Due to these domain differences, models trained in a new domain inevitably experience performance degradation when applied to previous domains.