Multimodal convolutional neural network for NDVI recovery under cloud cover using temperature context

Purpose of tesearch. Cloud cover disrupts the continuity of the time series of the normalized vegetation difference index NDVI in satellite monitoring, making it difficult to assess critical phenological stages and stress factors. Traditional methods demonstrate limited effectiveness: the Savitsky-Goley filter suppresses short-term phenological phases, spatial substitution ignores soil heterogeneity, and generative-adversarial networks create physiologically implausible textures. The aim of the study is to develop a multimodal convolutional network for NDVI recovery under clouds using spectral-temporal context and meteorological data. Methods. An architecture with an early fusion of spectral and meteorological modalities is proposed. The model processes an 8-day sequence of six spectral indices and a green RGB image channel, as well as a vector of weather signs for the target day. A cloud detector based on RGB analysis with limitations on brightness, whiteness, and textural uniformity has been developed, reducing false alarms on light soils. The experiments were conducted on two independent sunflower fields with three sets of weather signs. The quality assessment was performed exclusively on cloudy areas of the image. Results. The minimum set of features (relative humidity, average air temperature, average atmospheric pressure) showed the highest accuracy: R2 = 0.9319, MSE = 0.00130, PSNR = 28.86 dB. The temperature set provided comparable quality R2 = 0.9245 with improved structural similarity SSIM = 0.7883. The complete set of weather signs did not improve the R2 = 0.9216 metric, which indicates the redundancy of some of the signs. Conclusion. A multimodal network with a minimal set of weather signs ensures highly accurate NDVI recovery under clouds while maintaining physiological plausibility. The advantages of the approach include resistance to false alarms of the cloud detector and compact architecture. The results open up opportunities for continuous monitoring of stressful events during critical phases of sunflower development.