Implementation of a Novel Adaptive Deep Learning-based Change Detection Framework under Flood Environment

Abstract

An important and efficient Remote Sensing (RS) application is change detection, as it aids in locating the crucial changed areas and offers respective time-series data with the assistance of RS imagery. Here, one of the common and unexpected tragedies is flood, which affects the lives of people and the basic needs of the public. Thus, it is essential to tackle several issues that take place in the classical change detection models. In this research, a new learning-based model is introduced for flood change detection. At the beginning, essential validation data are gathered from benchmark deep learning sources. After data collection, images are fed into a developed deep learning framework. The detection of flood-related changes is then carried out through the SCA-MCA-E-ADDUNet[Formula: see text] model, which integrates Spatial Cross Attention and Multi-Convolution Attention Fusion Encoder within an Adaptive-Dilated DenseUNet[Formula: see text] structure. Moreover, various parameters in ADDUNet[Formula: see text] are optimised using the Updated Alpha Value-based Enhanced Wild Gibbon Optimisation Algorithm (UAV-EWGOA), which aids in improving the change detection efficiency. Finally, the change detection outcomes are obtained from SCA-MCA-E-ADDUNet[Formula: see text]. Later, various experiments are executed to verify the overall change detection efficiency of existing models. Here, Dice coefficient of SCA-MCA-E-ADDUNet[Formula: see text]-UAV-EWGOA is 94.81%, Intersection over Union (IoU) is 90.14%, accuracy is 94.9%, specificity is 95.08% and F1-score is 94.81%, respectively. Thus, the result was highly effective in detecting changes, potentially surpassing the capabilities of traditional or classical models in this specific change detection task.