Images of agricultural products—such as tobacco leaves, tea leaves, Chinese medicinal herbs, edible fungi, and dehydrated vegetables—captured in enclosed environments often suffer from poor illumination, color distortion, and significant background noise. These challenges impede accurate monitoring of color changes and reliable quality assessment. Existing low-light enhancement methods in smart agriculture are constrained by their reliance on large, high-quality annotated datasets. To address this, we propose DynZero-DCE, a rule-based dynamic zero-reference deep curve estimation framework tailored for agricultural low-light imaging. The method introduces: (i) a rule-based dynamic adjustment that adapts enhancement to per-image statistics, (ii) a cubic curve with piecewise luminance control for exposure stabilization, (iii) cross-scale local feature fusion and an edge-preserving denoising residual that suppresses noise amplification while maintaining leaf venation details. Trained without paired references, our losses enforce exposure consistency, color fidelity, and structural preservation. On a curated dataset, DynZero-DCE improves clarity by up to 32.9× and brightness by up to 15.7×. These results demonstrate superior luminance balancing, color faithfulness, and detail rendering under extreme lowlight conditions.