Chlorophyll inversion in walnut leaves based on improved spectral indices and texture features

This study focuses on the walnut cultivar 'Wen 185' in the arid region of southern Xinjiang. Addressing the limitation that traditional vegetation indices often experience saturation during peak growth periods, which restricts inversion accuracy in complex orchard environments, we obtained full-growth period images using a multispectral unmanned aerial vehicle platform. We incorporated the blue (450 nm) and red-edge (750 nm) bands to develop enhanced spectral index schemes. By integrating machine learning algorithms, we achieved improved estimation of SPAD values. The results indicate that the multiplicative improved spectral index scheme using the blue band (Scheme c), which multiplies and recombines the traditional index with the blue light band, significantly improves sensitivity to chlorophyll variations and effectively mitigates the spectral saturation phenomenon. In algorithm evaluations, the Convolutional Neural Network (CNN) model achieved the highest individual validation predictive accuracy under Scheme c (validation R² = 0.814, RPD = 2.359). However, its performance varied across different index combinations. In contrast, the Random Forest (RF) model demonstrated consistent cross-scheme stability and generalization capability across all spectral enhancement formulations (validation R² ≥ 0.70). Therefore, while CNN provides peak performance under specific optimal feature configurations, RF offers a more stable and robust alternative for practical applications involving diverse spectral inputs. While the spatial distribution map, generated by integrating geographical coordinates, visually illustrates the dynamic evolution of walnut SPAD-characterized by an initial increase followed by a decrease during the growth period-subsequent multi-stage variance analysis revealed that the superiority of the W1F1 water and fertilizer coupling treatment is stage-dependent. Specifically, W1F1 achieved significantly higher SPAD values than other combinations during the oil transformation stage (p < 0.05), whereas differences across other growth periods were not statistically significant. The findings provide a scientific foundation for precise nitrogen diagnosis and informed decision-making in the smart fruit and forestry sector in arid regions.