CADE: A lightweight change-adaptive dual-path encoder framework for predicting and generating future changes in satellite imagery using semi-supervised learning
TL;DR
The findings indicate that CADE is precise and computationally efficient, showing that it is suitable for low-resource remote sensing uses and minimizes dependence on extensive annotated datasets as well.
Abstract
Satellite image generation is critical in disaster management, environmental monitoring, and urban planning applications. Traditional methods of change detection often face challenges related to noise and atmospheric conditions, while recent deep learning methodologies, including Generative Adversarial Networks and transfer learning, typically require substantial computational resources. This study presents a novel, lightweight, Change-Adaptive Dual-Path Encoder (CADE) designed to mitigate these challenges, enabling the efficient generation of predictive satellite images with reduced resource demands. Our model enhances image generation capabilities by utilizing a Residual block in a separated dual encoder and attention module, even with limited labeled data. Additionally, semi-supervised training significantly strengthens the model’s ability to learn from small datasets. CADE achieved a PSNR of 30.45 and an SSIM of 0.80 when tested on the LEVIR-CD dataset, surpassing top diffusion and GAN models. For DSFIN-CD, a PSNR of 28.93 and an SSIM of 0.76 were achieved. Furthermore, CADE obtained a 92.81% F1-score and 87.72% IoU on LEVIR-CD, and a 91.44% F1-score and 85.02% IoU on DSFIN-CD, all while using fewer parameters and reducing computational costs compared to other methods. This study contributes to Artificial Intelligence by proposing a semi-supervised encoder-decoder architecture that utilizes both labeled and unlabeled satellite imagery for temporal understanding and predictive modeling. Our findings indicate that CADE is precise and computationally efficient, showing that it is suitable for low-resource remote sensing uses. This semi-supervised method minimizes dependence on extensive annotated datasets as well. The use of this technology in engineering is pivotal for improving autonomous satellite systems that are responsible for effective Earth observation, evaluating natural disasters, and planning for sustainable environmental management.