摘要
• Deep learning was applied to generate high resolution Sentinel-2 satellite image. • Modeling was done by combining of a 2D Transformer and customized 1DTransformer to take image as input. • Close model performances between super-revolution Sentinel-2 and hyperspectral images. Within the recent decades, a multitude of earth observation (spaceborne) and airborne remote sensing approaches have been developed to monitor crop growth and nitrogen status to improve crop management by providing in-season status information and decision support in wheat ( Triticum aestivum L.) production. Both space- and airborne remote sensing approaches come with inherent advantages and disadvantages. The main differences lie in the imagery’s spatial and spectral resolution, and, in consequence, its potential for quantitative modeling of crop growth and nitrogen status. To combine the advantages of two of the cheapest and most widely available air- and spaceborne imagery sources in agriculture, high spatial but low spectral resolution UAV-borne RGB imagery and low spatial but high spectral resolution Sentinel-2 satellite imagery, geometrical super-resolution (SR) was performed to generate a super-resolved high spatial and spectral resolution Sentinel-2 satellite (HRS2S) image using a deep learning (DL) based super-resolution model. The resulting HRS2S image preserved the original spectral and radiometric resolution of the Sentinel-2 satellite image while altering its spatial resolution to match the high resolution of the UAV-borne RGB image. The HRS2S image was then used to monitor crop status by predicting the physiological parameters dry matter (DM), nitrogen content (NC) and nitrogen uptake (NU) of spring wheat. Although most studies in literature still apply classical empirical methods, e.g. vegetation index-based regressions or machine learning, this study applied pure transformer-based DL models, ResTrans21, which combine a 2D Transformer (SimpleViT) with a custom 1DTransformer, in which skip connections between Transformer components are implemented via feature concatenation to enhance sequential modeling. The performance of the ResTrans21 models on HRS2S was validated against ground truth samples, yielding high coefficients of determination R 2 = 0.84, 0.85 and 0.86 for DM, NC and NU, respectively. The performance was then compared to the performance of corresponding ResTrans21 models based on UAV-borne imagery of two hyperspectral imagers (Rikola HSI, 29 wavebands, 450–800 nm; Mjolnir V-1240, 200 wavebands, 400–1000 nm). Despite utilizing more wavebands as predictor variables, the performance of these models was only slightly better, yielding R 2 = 0.86, 0.86 and 0.90 (Rikola HSI) and R 2 = 0.89, 0.85 and 0.92 (Mjolnir V-1240) for DM, NC and NU, respectively. Given these results, it appears that fusing satellite imagery with corresponding high spatial resolution UAV-borne RGB imagery is a suitable alternative to the acquisition and use of costly UAV-borne hyperspectral imagery for crop status monitoring.