Accurate in situ quantification by surface-enhanced Raman spectroscopy (SERS) on biological tissues with uneven surfaces is a persistent challenge due to signal variability from surface irregularities and the coffee-ring effect, which severely limits reproducibility and reliability. Overcoming these limitations is critical for advancing SERS toward practical, high-accuracy applications in biological, agricultural, and clinical settings. Herein, we present a streamlined strategy integrating minimal sample preparation, a low-cost SERS substrate with a tailored one-dimensional convolutional neural network (1D-CNN) for reproducible SERS quantification on uneven biological surfaces. Detection of thiabendazole (TBZ) on apple skin served as a representative model. We highlight the critical role of sample preparation and SERS substrate selection in minimizing spectral variation. Gold nanoparticles (AuNPs) reduced the enhancement non-uniformity compared with silver nanoparticles (AgNPs). Despite preprocessing, substantial intensity variation persisted (RSD: 47.99% for AuNPs; 80.50% for AgNPs). To address this, a customized 1D-CNN with decreasing kernel sizes (57–37–11–3) was compared with single-peak intensity calibration (SPIC), partial least squares regression, random forest, and fixed-kernel 1D-CNNs (3 and 5). The tailored 1D-CNN consistently outperformed all other models while maintaining a lightweight structure and short training time (242 s), enabling rapid retraining for in situ analysis under field-relevant conditions. It achieved accurate quantification under high signal variability, improving the R² for AuNP-enhanced TBZ quantification from 0.332 (SPIC) to 0.935. This work establishes a generalizable framework integrating machine learning with in situ SERS to mitigate surface irregularities, enabling accurate quantification on real-world biological tissues and advancing SERS toward field-deployable sensing applications.