AI-driven radiometric correction of crop covered-net distortions for proximal multispectral vineyard imaging

Abstract

Net covers are widely used in vineyards to mitigate biotic and abiotic stressors. However, these structures introduce wavelength-dependent attenuation that alters multispectral reflectance, thus compromising the comparability of vegetation indices across illumination conditions required for remote sensing applications. Existing radiometric correction strategies typically rely on calibration panels, irradiance sensors, or paired open-sky reference acquisitions, which limit their applicability in netted vineyards. This study proposes a proximal multispectral correction framework to remove radiometric distortions caused by covering nets. The approach uses a combination of reference-based adjustment and a lightweight artificial neural network (ANN) model capable of reference-free correction. Multispectral images were acquired at the canopy level under shade-net and open-sky conditions using a fixed proximal imaging configuration. A snapshot experimental design was adopted to isolate illumination effects while minimizing temporal variability. Region-of-interest preprocessing and percentile-based distribution matching were used to generate training targets for ANN modeling. The ANN successfully inferred open-sky radiometric intensity distributions from shaded inputs across visible and near-infrared bands, achieving high correlations with reference imagery (R = 0.97) and slopes close to unity. Vegetation indices derived from corrected images closely matched unnetted grapevine values, with RGB-based indices reaching correlations up to R = 0.99 and multispectral indices up to R = 0.97. After supervised training with paired shaded and open-sky images, the proposed ANN-based model enables deployment-stage correction of shaded multispectral images without auxiliary instrumentation or concurrent reference acquisition. This ability to operate without references during inference restores radiometric consistency and supports robust computation of multispectral indices in netted vineyards.