Remote　sensing　(RS)　audio-visual　cross-modal　retrieval　is　a　challenging　task　in　the　search　of　meaningful　RS　information.　Nevertheless,　the　impact　of　multiscale　features　and　associated　redundant　information　in　the　RS　images　cannot　be　overlooked　in　the　retrieval　task.　In　addition,　how　to　deal　with　the　completely　different　physical　expressions　of　different　modal　information　is　crucial　for　cross-modal　retrieval　tasks.　To　tackle　these　issues,　we　propose　a　Scale-aware　Adaptive　Refinement　and　Cross-Interaction　(SARCI)　network.　The　Quaternion-attention　Dominated　Multiscale　Visual　Refinement　(QDMVR)　module　in　SARCI　is　suggested　to　learn　multiscale　visual　features　and　further　optimize　features　containing　redundant　information　for　different　scale　features.　To　better　integrate　channel　attention　and　spatial　attention　for　adaptively　learning　of　meaningful　visual　semantics,　we　propose　the　symmetric　quaternion　attention　(SQA)　within　the　QDMVR　module　to　enhance　RS　visual　features.　The　SQA　mechanism　acts　on　both　high-level　and　low-level　features　to　explore　salient　RS　vision　information　across　different　scales.　In　order　to　allow　information　from　different　modalities　to　interact　more　valuably,　we　propose　the　Instruction-based　Cross-Learning　Module　(ICLM)　to　perform　cross-modal　feature　interaction　based　on　the　characteristic　of　the　two　modalities.　SARCI　network　demonstrates　state-of-the-art　performance　on　three　public　RS　cross-modal　datasets:　Sydney,　UCM,　and　RSICD　audio-visual　datasets.　The　code　is　available　at:　https://github.com/WUTCM-Lab/SARCI.