Imagery, Guided

Annual (September to August) ice area flux through Nares Strait for 2016–2019 was determined using an established technique^{2 ,18 (link),38} . First, sea ice motion from each sequential pair of Sentinel-1 imagery (~0.5 to 1-day time separation) was determined using the Komarov and Barber tracking algorithm^{39 (link)}. Sea ice motion was then interpolated to a 30 km buffer region surrounding the gate and sampled at 5 km intervals across. Considering that ice rapidly deforms as it is being funneled through Nares Strait we placed our gate farther north of Nares Strait than has been done previously² to facilitate improved motion detection. The ice area flux (F) was calculated using: $F=\sum c_i{u}_i\mathrm{\Delta }x$ where c_i is the ice concentration obtained from the closest Canadian Ice Service ice chart⁴⁰ to the Sentinel-1 image date, u_i is the ice speed normal to the flux gate at the ith location and Δx is the spacing along the gate (5 km). If we assume that the errors of the sea ice motion samples are additive, unbiased, uncorrelated, and normally distributed, then the uncertainty in ice area flux across the gate (σ_f) can be determined using the following equation: ${\sigma }_f={\sigma }_{e}L{\left(\sqrt{N_s}\right)}^{-1},$ where, σ_e is the error in ice motion of 0.43 km/day determined previously^{39 (link)}, L is the width of the gate and Ns is the number of samples across the gate. For L=139 km and Ns=27 the uncertainty in ice area flux at our gate is ~±12 km²/day. On monthly or annual timescales, the uncertainty is close to zero.

As covariate layers for producing SoilGrids250m predictions we used an extensive stack of covariates, which are primarily based on remote sensing data. These include (see e.g. Fig 4):
These covariates were selected to represent factors of soil formation according to Jenny [40 ]: climate, relief, living organisms, water dynamics and parent material. Out of the five main factors, water dynamics and living organisms (especially vegetation dynamics) are not trivial to represent as these operate over long periods of time and often exhibit chaotic behaviour. Using reflectance bands such as the mid-infrared MODIS bands from a single day, would have little use to soil mapping for areas with dynamic vegetation, i.e. with strong seasonal changes in vegetation cover. To account for seasonal fluctuation and for inter-annual variations in surface reflectance, we instead used long-term temporal signatures of the soil surface derived as monthly averages from long-term MODIS imagery (15 years of data). We assume here that, for each location in the world, long-term average seasonal signatures of surface reflectance or vegetation index provide a better indication of soil characteristics than only a single snapshot of surface reflectance. Computing temporal signatures of the land surface requires a considerable investment of time (comparable to the generation of climatic images vs temporary weather maps), but it is possibly the only way to represent the cumulative influence of living organisms on soil formation.
For processing the covariates we used a combination of Open Source GIS software, primarily SAGA GIS [28 (link)], R packages raster [41 ], sp [42 ], GSIF and GDAL [43 ] for reprojecting, mosaicking and merging tiles. SAGA GIS and GDAL were found to be highly suitable for processing large data as parallelization of computing was relatively easy to implement.
We updated the 1 km global soil mask map using the most detailed 30 m resolution global land cover map from 2010. This was combined with the global water mask [44 (link)] and the global sea mask map based on the SRTM DEM [45 (link)] to produce one consistent global soil mask that includes all land areas, expect for: (a) fresh water bodies such as lakes and rivers, and (b) permanent ice.

The input datasets selected for this analysis were MODIS (1) MOD11A2 Land Surface Temperature (LST) 8-day composite data (Wan et al., 2002 ), and (2) MCD43B4 Bidirectional Reflectance Distribution Function (BRDF) – corrected 16-day composite data (Schaaf et al., 2002 ), from which Enhanced Vegetation Index (EVI) was derived using the equation defined in Huete et al. (1999) . The MODIS LST dataset consists of both daytime and nighttime average temperatures aggregated, respectively, from the descending and ascending paths of the NASA Terra Satellite. The BRDF dataset contains 16-day products, with overlapping temporal windows that result in an 8-day temporal resolution, which were derived from data collected by the MODIS sensors on both the Aqua or Terra satellites.
The MODIS data were collected on a per-tile basis and then merged using the MODIS reprojection tool (Dwyer and Schmidt, 2006 ) to create seamless mosaics for all of Africa. A total of 42 tiles were required to cover the continent for each image date (i.e., the day of the year corresponding to the center of the composite temporal window). The BRDF mosaics each consisted of seven spectral bands, three of which were needed to derive the EVI, and mosaics were created for each of these bands prior to deriving the EVI for each image date. The resulting data archives consisted of 594 EVI mosaics (from day 049, 2000 to 361, 2012), and 590 LST-day and LST-night mosaics (from day 065, 2000 to 361, 2012). Temporal mean and standard deviation images were derived on a per-pixel basis from the full mosaic archives for each of the three variables for subsequent use in the gap filling algorithms. Producing images of summary statistics was also useful for identifying pixels that never contain usable data (e.g., ocean pixels) that could be ignored in the gap-filling procedures, thus reducing run-time.
The initial step in the gap filling process was to identify gap pixels in need of filling through the use of a despeckling algorithm, which is a processing step that need only be used if corresponding datasets describing pixel-level data quality do not exist. While MODIS products have associated quality assurance datasets useful for identify potential gaps, we developed a generic gap-finding approach to demonstrate the potential utility of our gap filling approach for a wide range of remotely sensed products. Gaps were identified by finding all pixels that contained a no-data or otherwise unacceptable value within the input mosaic that corresponded to usable pixels within the mean image, thus indicating that the pixel in question contained usable data on other dates. Unacceptable pixel values were identified by calculating a z-score for each pixel based on the mean and standard deviation images, and then searching for any pixel with an absolute z-score exceeding a user-defined threshold (we used 2.58, which corresponds to the 0.99 confidence interval, see supplemental information for more details). When such a pixel was found we examined neighboring pixels (we used a neighborhood size of 40 to 80 pixels) to determine if they were similarly unusual with respect to the mean value of the pixel. If the original z-score was beyond a second user-defined threshold (we used ±0.2) from the median neighborhood z-score, or if too few neighboring pixels were found within a user-defined search radius (we used 10 km), the original pixel was reclassified as a gap. In practice, pixels removed by the despeckling algorithm typically represent approximately 5% of gap pixels or 0.5% of all usable pixels present in the final output images.
Based on the results of the gap identification process the flag image was modified to indicate whether pixels were (1) a no-data pixel that should be ignored in subsequent processing, (2) a usable raw value that could be passed directly through to the final output (and is suitable for use in the gap-filling models), or (3) a gap to be filled. A preliminary analysis of the raw imagery mosaics indicated that, on average, approximately 5–15% of the pixels within an image were gaps in need of filling (Table 1).Table 1

The mean and standard deviation percentages of gap pixels within the full Africa mosaics as calculated from the full imagery time-series (e.g., approximately 15% of a typical EVI mosaic consists of gap pixels).

Dataset	Proportion of missing pixels per image (%)
	Mean	Standard deviation
EVI	14.77	5.93
LST day	5.25	2.28
LST night	8.51	3.28

Images were retrieved from the proprietary Scanco microCT file type format (.ISQ), which contained single-pixel width two-dimensional (2D) trans-axial projections, or ‘slices’, that together formed stacks depicting an entire pod as three-dimensional (3D) volumes. Thirty-two distinct 3D volumes were included in the experiment dataset, each containing a single entire B. napus pod. All 2D trans-axial (XY) slices were 512 × 512 pixels, therefore the height and width of all 3D volumes was also 512 pixels. Individual 3D volumes varied in length from 505 to 1397 slices, with a total of 29,871 slices in the experimental dataset. The total dataset contained 471 seeds.
The total dataset was split into a model training and validation dataset comprised of 13 3D volumes, 12,475 2D slices and 262 seeds and a model testing dataset containing comprised of 19 3D volumes, 17,396 2D slices and 209 seeds. This split was decided upon due to the uneven number of seeds in each seed pod, with the training and validation dataset containing 262(56%) of seeds and the testing dataset containing 209(44%) of seed. Another factor impacting the split of data was that intact 3D volumes of entire seedpods needed to be used for testing, to demonstrate that reliable seed detection and segmentation could be achieved on the original imagery without any pre-processing. Conversely, Weigert et al. (2020) demonstrated that more accurate results could be obtained in a computationally efficient manner by training a StarDist-3D on smaller sub-volumes of the original 3D volume data containing objects of interest, in this case sub-volumes containing at least one entire seed. The 3D volumes in the model training and validation dataset were therefore comprised of 138 small sub-volumes of stacked 2D slices containing a single seed, or multiple seeds in instances where seeds occupied some of the same 2D slices. These sub-volumes ranged in size between 24 to 84 2D slices depending on the size of the single seed or multiple overlapping seeds contained within. This sub-division was carried out in order to ensure a mixture of seeds from different seed pods could be used for model training and validation. 117 sub-volumes containing 220 seeds were randomly sorted into the final ‘training’ dataset, and 21 volumes containing 42 seeds were sorted into the final ‘validation’ dataset.

There is a variety of methods to filter noise, determine stimulation, decode, and predict neural activity. Some authors have attempted neural decoding through training sessions, while others have determined arbitrary thresholds which can be used as a neural ‘switch’ to enable neural plasticity to activate desired movements. When EEG is used in the setting of neural bypasses, recording thresholds are determined which then translate to effector stimulation, often with FES. EEG thresholds to stimulate FES are typically obtained through motor imagery as measured by attention with sensorimotor rhythm and beta/theta oscillation ratios, or Common Spatial Patterns based on Event De/Synchronization [21 (link)–41 ]. Others have used steady-state visual evoked potentials (SSVEP) to trigger FES stimulation [42 (link)]. Furthermore, some studies have used alpha rhythms as a deactivating signal following a stimulation event [43 (link)]. When single-cell recordings are used, the threshold for stimulating FES has typically been cell firing rate [5 (link)]. When ECoG is used, the rate of high-gamma oscillations has been selected as a threshold for effector muscle stimulation [3 (link)]. In studies with microelectrode arrays, neuronal action potential rate, or average spectral high-frequency power, have typically been used as thresholds for stimulation of effector muscles [7 (link)]. Microelectrode arrays have also been used to record mean wavelet power after artifact removal during trials of imagined movements in paralyzed patients [4 (link), 44 (link)]. Microelectrode arrays are often used during training sessions prior to paralysis or with simulated motor tasks to create predictive models of neuronal control of muscle activity [19 (link), 45 (link), 46 (link)].
In some studies, daily calibration is required to train neural decoding algorithms which presents a limitation for the translation of these technologies to real-world environments. One possible approach to this problem is a neural network capable of decoding without daily training sessions [18 (link)]. Other decoding methods consist of gradient boosted trees, support vector machines, and linear methods [12 (link), 47 ]. A drawback to any decoding method is the group of associated assumptions. For example, assumptions made with a regularized linear regression are that outputs are proportional to input changes, additional noise is assumed to be Gaussian noise, and that the regression coefficients are from a Gaussian distribution [47 ]. Bouton et al. suggest that nonlinear methods of decoding may help to increase robustness and accuracy of specific decoders [48 ]. As methods increase in complexity, so do their associated assumptions. Glaser et al. states that a crucial assumption built into decoders is the form of the relation between the input and output. With machine learning methods, multiple decoding models can be organized in ensembles [47 ]. Bouton et al. demonstrates this in their finding that Long Short-Term Memory-based deep learning networks, used in tandem with repeatability-based feature selection based on temporal correlation, results in positive outcomes and accurate decoding [12 (link)].
Developing devices to maximize spatial resolution, temporal resolution, and biocompatibility will be crucial in developing robust neural interfaces. Additionally, a number of unidirectional recording or stimulation devices exist that have not yet been combined into neural bypasses, including novel spinal cord stimulators or neurotransmitter-sensing electrodes [49 (link), 50 (link)]. There exists a lack of comparative analysis across studies with different methodologies to determine the most efficacious methods of achieving neural bypass.