Reliable estimation of rice leaf nitrogen, chlorophyll, and SPAD supports precision nutrient management and crop phenotyping. However, measurements of narrow leaves are susceptible to background interference when the sensor footprint exceeds the target area. This study (1) compared black and white backgrounds on the same rice leaves under a wide measurement geometry in 2023 and (2) descriptively evaluated SPAD prediction using a redesigned narrow-geometry housing on an independent sample set collected in 2026. Spectra were collected from IR64 rice leaves at four growth stages and paired with SPAD, Kjeldahl nitrogen (%N), chlorophyll a, chlorophyll b, and total chlorophyll measurements. Under wide geometry, an 110 mm sensor-to-leaf distance and ±20.5° field of view produced an approximately 82 mm footprint, exposing an estimated 68–85% of the measurement area to the background. The white background increased raw signals by up to 10–11-fold in the blue region. Using five-fold outer and three-fold inner nested cross-validation, paired Full 18-channel analyses of 140 leaves yielded pooled OOF R² values for black versus white backgrounds of 0.620 versus 0.588 for SPAD, 0.262 versus 0.371 for %N, 0.619 versus 0.434 for chlorophyll a, 0.486 versus 0.267 for chlorophyll b, and 0.567 versus 0.399 for total chlorophyll. Paired bootstrap comparisons showed no clear background-dependent difference for SPAD or %N, whereas the black background consistently improved chlorophyll a, chlorophyll b, and total chlorophyll predictions, with 95% confidence intervals for paired differences in R², RMSE, MAE, and CCC excluding zero. Fold-local Random Forest selection retained 2–17 wavelengths across folds and targets but did not consistently outperform the Full 18-channel scenario. Under narrow geometry, a 10 mm sensor-to-leaf distance produced a nominal 7.5 mm optical footprint within a 30 mm sample opening. Using the Full 18-channel scenario, SPAD prediction reached R² = 0.665, RMSE = 2.720, MAE = 2.122, and CCC = 0.810. Overall, background exposure was associated with target-dependent differences in prediction performance under the wide geometry, while the numerical SPAD performance of the independent narrow-geometry dataset cannot establish a geometry effect.