Progressive Multimodal Integration for RGB-Based NDVI Inference in Golf Course Turfgrass

The normalized difference vegetation index (NDVI) is a key indicator of turfgrass condition.However, direct NDVI measurement requires a multispectral sensor that captures the red and near-infrared (NIR) bands, which has increased interest in deep-learning methods for inferring NDVI from conventional red-green-blue (RGB) imagery.Existing multimodal studies have mainly compared the performance of final models that simultaneously incorporate multiple environmental variables, while few have systematically quantified the incremental predictive contribution of individual environmental-information blocks.This study therefore proposes a Progressive Multimodal Integration (P-series) framework that adds environmental information stepwise under a common U-Net-style encoder-decoder fusion topology (nearly identical capacity across stages) and an identical training protocol.Four models (P1b, P2, P3, and P4) were trained from scratch using the same RGB imagery, course-zone information, and multilayer-perceptron branch for processing environmental information.P1b served as a zero-weather baseline in which the environmentalvariable inputs were set to zero.The analysis used 121,683 co-registered RGB-NDVI image chips from 87 site-dates acquired at four golf courses in a climatic transition zone where cool-and warm-season turfgrasses coexist.NDVI targets were generated using an operational NIR/Red image-processing workflow without reflectance calibration.Models were evaluated using a hole-based leak-free split and five random seeds.On the validation data, P1b achieved a mean absolute error (MAE) of 0.0230 and a coefficient of determination (R²) of 0.975.Adding basic meteorological and seasonal information in P2 significantly improved performance (Holm-adjusted p < 0.001).In contrast, adding air-temperaturebased growing degree days in P3 yielded no statistically significant additional benefit (p = 0.273), while adding extended environmental information in P4 significantly degraded generalization performance (p < 0.001).Redundancy analysis showed that most newly added variables were largely explained by the preceding information block (mean R² ≈ 0.90), demonstrating that adding variables does not necessarily increase predictive information.In an auxiliary analysis of six future captures, three showed errors comparable to historical performance, two remained within the historical error range but with visibly reduced accuracy under atypically cool meteorological inputs, and one showed a marked degradation most plausibly associated with a capture-level Red-band radiometric anomaly, supporting the need for capture-level quality control.Overall, improving generalization through environmental information depends more on selecting predictive information that is not redundant with existing inputs than on increasing the number of variables.