<p>With the speedy advancement of Convolutional Neural Networks (CNNs), deep hashing is extremely significant in remote sensing image retrieval scenarios due to the low storage cost and quick searching speed. However, under multi-label remote sensing retrieval scenarios, conventional proxy-based methods only consider the dependency relationships between proxies and samples while neglecting semantic relations among data points. This limitation fails to convey complex label correlations and causes conflicts in the embedding space. Besides, current remote sensing image retrieval frameworks tend to utilize global descriptors to characterize the image and hardly reflect the local spatial information associated with multi-label semantics, which further impedes the feature representation capability. To resolve the mentioned problems, a new Deep Semantic-dependence Proxy Hashing Framework (DSdPH) is proposed to improve the efficiency of model retrieval in multi-label remote sensing retrieval. Specifically, an attention-guided feature extraction framework is employed to obtain the fine-grained representations by jointly learning the raw samples’ global semantic information and the inherent spatial structures. Based on the irrelevant pairs constraint, the dependence-guided proxy loss is developed to construct a robust embedding space, in which the semantic relationships of multi-label data are well expressed by jointly exploring the proxy-data dependency and the correlation of irrelevant samples. In addition, to reduce information loss during the conversion from continuous feature vectors to discrete binary hash codes, we introduce a semantic location quantization strategy to enhance the discriminative ability of discrete coding. Extensive experiments on three benchmarks demonstrate DSdPH’s superior performance. Code is available at <a href="https://anonymous.4open.science/r/DSdPH">https://anonymous.4open.science/r/DSdPH</a>.</p>