Parrots

ModelCraft and the CCP4iBuccaneer pipeline were run on each case in the three test sets. To prepare the data for the CCP4iBuccaneer pipeline, experimental phases were modified using five cycles of Parrot and models in the MR and AlphaFold test sets were refined with ten cycles of REFMAC. This was not performed for ModelCraft as density modification and model refinement are performed internally. Both pipelines were made to use the experimental phases in MLHL refinement with REFMAC, although ModelCraft stops using them once R_work decreases to 35%. An ablation study was also carried out on the molecular-replacement test set where individual steps in the ModelCraft pipeline were removed to see the effect on the overall performance. An additional test was carried out to reduce the number of cycles in ModelCraft from 25 to five, as was used in the original CCP4iBuccaneer pipeline.
Performance was measured using R_work, R_free, protein model completeness and the time it took the pipeline to finish. Completeness was calculated by moving the built model over the deposited model using CSYMMATCH and then taking the percentage of protein residues in the deposited model that had a matching residue in the built model. A residue was considered to match if N, C^α and C were all within 1 Å. CSYMMATCH uses the space-group symmetry operations to try and match each chain onto the reference model. It also accounts for possible differences in the cell origin, which is important as the molecular-replacement solutions are not guaranteed to have the same origin as the deposited model. Completeness is used as the primary metric for comparison between pipelines. Using R factors would be a less fair comparison because the CCP4iBuccaneer pipeline only builds protein and ModelCraft also builds nucleic acids and waters.

In 2020, a total of 11 atmospherically corrected Sentinel-2 L2A images (hereinafter, S2) were selected as suitable (containing no visible clouds or haziness at the experimental site). The acquisition times of the images spanned through the cropping season, from June 19 to September 5. In 2021, 15 images were selected, acquired from June 2 to September 10. All images were processed using the European Space Agency (ESA) Sentinel Application Platform (SNAP) v8.0.9 (ESA, http://step.esa.int). The bands selected for the NDVI calculation were $B8A$ , that is, the narrow-NIR vegetation band centred at 865 nm, and $B4$ , the red band centred at 665 nm. The selection of B8A instead of the wider-NIR B8 was performed following the suggestion by Kaplan and Rozenstein (2021 ) that band 8A is best suited for LAI estimations given its narrower NIR window. In both years, a Sequoia multispectral sensor (Parrot, Paris, France) was mounted on a Matrice 600 Pro drone (DJI, Shenzhen, China), hereinafter called the UAV. The camera has a resolution of 1280 × 960 pixels and measured reflected radiation in four spectral regions, namely, the green (530–570 nm), red (640–680 nm), red-edge (730–740 nm) and near-infrared (NIR, 770–810 nm) ranges. The camera was calibrated prior to each flight. The ground pixel resolution was 0.04 m on average across all flights, with a 40% side overlap between flight swaths. Due to the intrinsic overlapping nature of UAV images, the swaths were processed by maintaining the mutual independence of the recorded strips in all processing steps. This prevented swath mosaicking and pixel resampling on the overlapping areas until the final index product. After that, the nearest-neighbour resampling method was adopted for image mosaicking. The UAV surveys were conducted around midday under clear-sky conditions and took approximately 45 min to complete, thus minimizing possible changes in light conditions and solar zenith angle.
In both 2020 and 2021, LAI was estimated using a parametric fractional vegetation cover (FVC) based method as proposed by Zeng et al. (2000 (link)) and adapted to LAI calculation by Ali et al. (2015 (link)). The method relies on the definition of an extinction coefficient parameter $k(\vartheta )$ . For both UAV and Sentinel-2 parametric estimations, the 2020 ground measured LAI dataset was used for model parameter calibration. For 2021, UAV images and Sentinel-2 images that were closer in time to LAI ground surveys (June 9, July 29 and August 26) were then selected for validation. Following calibration on the 2020 dataset, a single sensor-specific $k(\vartheta )$ value was adopted, thus creating a suitable approach for model applications without further $k(\vartheta )$ estimations. The $k(\vartheta )$ was 0.59 for Sentinel-2 and 0.37 for UAV multispectral sensor. Further information about the method used for parametric LAI calculations from both Sentinel-2 and UAV is reported in supplementary materials.
Additionally, the Biophysical Processor Algorithm (LAI_{S2_MLA}) at a 10 m resolution included in SNAP for all Sentinel-2 data was used for all available dates in 2020 and 2021. The algorithm uses a trained neural network to derive LAI values from the top-of-canopy-level reflectance and is considered to perform reasonably well over a large array of vegetation types. In the text, LAI values estimated with the parametric method are addressed as LAI_S2, while the LAI values estimated with the machine learning algorithm included in the biophysical processor are addressed as LAI_{S2_MLA}.

An unmanned aerial vehicle (UAV) with mounted either multispectral Parrot Sequoia 1.2 MP, 1280 × 960 px (bands: green at 550 nm ± 40 nm; red at 660 nm ± 40 nm; red edge at 735 nm ± 10 nm; near-infrared at 790 nm ± 40 nm) or thermal sensor DJI Zenmuse XT uncooled Vox Microbolometer 640 × 512 px was operated at a flight altitude of about 100 m over the experimental field, resulting in 10 cm spatial resolution images. The orthomosaics of multispectral and thermal images were processed in Pix4Dmapper software (Pix4D SA, Prilly, Switzerland) with the use of “Ag Multispectral” and “Thermal camera” modes, respectively. The output was high-resolution GeoTIFF radiometrically calibrated images. The calibration of the multispectral images was performed with the MicaSense calibration panel images obtained prior to the flight and used during the processing in the Pix4Dmapper software. Thermal orthomosaics were calibrated according to the hottest (soil) and the coolest (wet canopy) pixels in the image. The flights were performed over the spring season of 2019 with additional diurnal flights in spring 2020. The flight times and main atmospheric parameters, such as air temperature, solar radiation, wind speed, relative humidity, and reference evapotranspiration, are presented in Table 2.