Aquaculture is a critical component of sustainable food systems, particularly in the Amazon Basin, where its expansion must balance production goals with ecological preservation. Monitoring aquaculture systems is essential for ensuring sustainable development, yet traditional satellite-based approaches face significant challenges in identifying aquaculture ponds amidst the region’s complex tropical landscapes. These challenges are exacerbated by the variability in pond sizes, shapes, and land-use contexts, requiring innovative solutions. Here, we present AISciVision-Aqua, a novel framework that integrates Visual Retrieval-Augmented Generation (VisRAG) and domain-specific interactive tools to enhance aquaculture pond detection using high-resolution satellite imagery. Leveraging Large Multimodal Models (LMMs), AISciVision-Aqua incorporates domain-specific interactive tools, such as panning and zooming, that emulate human workflows for image classification and enable dynamic exploration of spatial contexts. The framework was trained and validated on a robust dataset combining PlanetScope imagery and a validated database of aquaculture pond locations in Rondônia state, Brazil. Our results show that AISciVision-Aqua achieves significantly higher precision and recall compared to baseline methods, including random forest classifiers and supervised deep-learning models. Additionally, the framework provides interpretable transcripts of its decision-making process, fostering transparency and facilitating expert validation. By integrating spatial analysis tools and iterative reasoning, AISciVision-Aqua enables real-time refinement of predictions, offers a scalable, accessible solution for aquaculture monitoring in complex landscapes like the Amazon, and supports effective decision-making to balance food security with environmental conservation in tropical ecosystems. Our approach highlights the transformative potential of artificial intelligence in addressing sustainability challenges in freshwater science.