In recent years, block-based compressive sensing (block-based CS) integrated with deep neural networks has emerged as a promising approach to efficiently addressing image compressive sensing (ICS) problems. It demonstrates notable advantages in both data acquisition and reconstruction and significantly improves image reconstruction quality. However, existing methods often treat image blocks uniformly when handling various image features. In addition, their adaptive sampling resource allocation strategies struggle to accommodate the majority of image types. Moreover, achieving high-accuracy image reconstruction with limited sampling data remains a significant challenge. This paper proposes a WVD (Weighted Variance and DCT-based) sampling allocation strategy, which dynamically evaluates local variance and frequency domain saliency to ensure highly discriminative and differentiated processing among image blocks. Additionally, a multi-channel sampling network is constructed, initialized with Gaussian random matrices, to support the flexible operation of multiple sampling rates. An iterative optimization process is implemented through a novel two-stage training strategy, using a shared dataset stratified by sampling rates. Each sampling matrix is fine-tuned to align with its corresponding target sampling rate. Experimental evaluations on multiple standard datasets reveal that the proposed method consistently surpasses current state-of-the-art deep learning models in objective performance measures, regardless of the sampling rate.