A Comprehensive Review of Weight Initialization Strategies for CNNs in Remote Sensing

Abstract

Weight initialization is a critical factor in training convolutional neural networks (CNNs), particularly in remote sensing (RS) applications where data is often high-dimensional, imbalanced, and noisy. This review explores traditional methods such as Random, Xavier, and He initialization, as well as modern strategies like MetaInit, GradInit, RGB-based, Laor, and convergence-aware approaches. These techniques are evaluated in the context of RS-specific challenges and benchmarked across architectures like VGG, ResNet, and DenseNet. Case studies in urban segmentation, SAR target recognition, and infrastructure quality assessment show that domain-sensitive initialization improves convergence speed, training stability, and classification accuracy. The review emphasizes the importance of selecting adaptive and data-aware initialization schemes that align with RS data characteristics. Overall, weight initialization should be treated not as a fixed preprocessing step but as a key design element for building robust, accurate, and scalable deep learning models in RS.