During pulsed laser deposition (PLD) of YBa₂Cu₃O_7-δ (YBCO) superconducting thin films, the lattice state is governed not only by preset process parameters but also by dynamic conditions such as plasma-plume evolution, temperature drift, and chamber-pressure fluctuations. Conventional manually recorded parameters therefore provide an incomplete description of the actual growth process, while the small number of labeled samples available in materials experiments further limits data-driven modeling. Here, we develop an in situ multimodal monitoring and robust small-sample learning framework for PLD-grown YBCO films. Plume images are acquired synchronously with laser pulses, and optical character recognition is used to extract temperature and vacuum-pressure time series, yielding process data that combine nominal parameters, dynamic thermal/vacuum information, and plume morphology. U-Net and SegFormer are employed to segment the high-intensity core and diffuse envelope of the plume, respectively, from which physics-guided descriptors of energy-density distribution, spatial expansion, morphological stability, and core-envelope coupling are constructed. A nested modeling procedure combining repeated random subsampling for stable feature ranking, minimum redundancy-maximum relevance constraints, high-correlation filtering, and partial least-squares regression is then used to predict the X-ray diffraction (005) peak position. Leave-one-out cross-validation shows that static growth parameters have little predictive power (R²=0.0484), and adding temperature/vacuum dynamics yields only a modest improvement. In contrast, plume morphology markedly improves performance: plume features alone give the best result (R²=0.7193), slightly outperforming the combination of all feature groups (R²=0.6872). Ablation analysis indicates that the core and envelope contain complementary information and that their spatial coordination is critical for describing the YBCO c-axis lattice state. A label-randomization test (p=0.0196) further suggests that the observed predictive performance is unlikely to arise from chance fitting. This work provides a low-cost and scalable route for in situ monitoring, interpretable modeling, and data-driven analysis of PLD thin-film growth under small-sample conditions.