CA-STIM: an interpolation model with spatio-temporal evolution characteristics and cross-attention mechanism for 2D island morphology sequences

Abstract

Coral islands are highly susceptible to climate change, making it crucial to understand their spatial morphological evolution for sustainable development and management. Recent advancements in high spatio-temporal resolution earth observation technologies have facilitated the analysis and prediction of island morphology over medium to short timescales. However, issues such as cloud cover and atmospheric interference often lead to poor image quality, resulting in significant missing in the 2D morphology sequences extracted from remote sensing images. To address this issue, we propose a spatio-temporal interpolation model (CA-STIM) that integrates both external environmental dynamics and the intrinsic spatio-temporal evolution characteristics of island morphology using a convolutional neural network-long short-term memory network (CNN-LSTM) framework with a cross-attention mechanism and a weighted binary cross-entropy loss function. The cross-attention mechanism incorporates external environmental factors, enhancing the interpolation accuracy, while the weighted binary cross-entropy loss function effectively addresses the challenge of directional heterogeneity. Using three coral islands in the South China Sea (Beizi, Mahuan, and Xiyue) as case studies, we perform 2D spatial morphology series interpolation. Experimental results demonstrate that our model outperforms baseline methods, achieving Dice scores of 0.9681, 0.9675, and 0.975 and Intersection-over-Union (IOU) scores of 0.9383, 0.9373, and 0.9513 on Beizi, Mahuan, and Xiyue Island, respectively.

Keywords

• Spatial morphology evolution of islands
• interpolation
• cross attention
• feature fusion
• deep learning