Assessing reclamation potential of abandoned dryland is increasingly critical amid escalating water scarcity and rising food security demands. We develop a knowledge-guided machine learning framework integrated with remote sensing to inform water-carbon spatial optimization. The framework predicts 1-km resolution estimates of irrigation water use, net primary productivity, and soil organic carbon, and is effectively validated across the Middle Reaches of Yellow River Basin, the largest irrigated dryland agricultural region in China. Spatial optimization identifies nearly 40 % of abandoned dryland (similar to 800,000 hectares) as suitable for reclamation, with an annual economic return of USD 40 million. We find that reclamation performance does not follow a monotonic relationship with vegetation growth, and it is optimized under water-conserving regimes or high-potential zones. Reclamation potential remains stable under low- and medium-emission scenarios through 2031-2050 but declines under high-emission trajectories. The proposed framework demonstrates stability across multiple model layers and data preprocessing pipelines but requires further refinement in input variables and algorithm configuration. Despite ecological and economic co-benefits, reclamation could be constrained by the mismatch between high upfront investment (USD 240 million per year) and a long payback period (17 years). This indicates the need for sustained fiscal support during the early implementation phase to enable a transition toward long-term financial self-sufficiency. Overall, our work presents a transferable and cost-effective framework for evaluating reclamation potential and guiding climate-adaptive resource management, with broad applicability for dryland regions worldwide.