Yield Prediction Through Integration of Genetic, Environment, and Management Data Through Deep Learning

May combine two claims.

Recent deep learning yield prediction studies used relatively large datasets, with or without genetic data, and did little feature engineering.

“Several recent studies have applied these methods with a relatively large dataset either with (Washburn et al. 2021) or without (Khaki et al. 2020) a genetic component into the model, with little feature engineering performed.”

• cites Washburn et al. 2021 sourceNot checked — no full text of the source
• cites Khaki et al. 2020 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

A recent study using a hybrid machine learning and crop growth model prediction system found water-related features important, although not as important as the trend in genetic and management improvements over time.

“Additionally, in a recent study using a hybrid machine learning and crop growth model prediction system, the authors found that water-related features (e.g.average drought stress, average water table in season) were important, although not as important as the trend in genetic and management improvements over time (Shahhosseini et al. 2021).”

• cites Shahhosseini et al. 2021 sourceSupported

May combine two claims.

BLUP models had the smallest average error, whereas deep learning models were more consistent and had similar average error.

“BLUP models achieved the lowest average error, but deep learning models performed more consistently with similar average error.”

May combine two claims.

In the best model, adding interactions changed how sensitive the model was to weather and management features, including lower importance scores for late-season timepoints around 200 days after planting.

“Examining the effect of interactions in the best-performing model revealed that including interactions altered the model's sensitivity to weather and management features, including a reduction of the importance scores for timepoints expected to have a limited physiological basis for influencing yield-those at the extreme end of the season, nearly 200 days post planting.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The reported weather data showed evidence of equipment malfunction.

“Within the reported weather data, we observed evidence of equipment malfunction and imputed or adjusted values using linear models.”

May combine two claims.

Many sites lacked measurements for many soil properties across the seasons considered, and the timing of fertilizer applications was absent in some cases.

“For example, many sites lack measurements for many soil properties across the seasons considered here, and the timing of fertilizer applications was absent in some cases.”

May combine two claims.

Salience of PCs differed by as much as 0.432 (PC 24), with differences for the first 8 PCs (31% variance explained) ranging from 0.200 (PC1) to 0.309 (PC7).

“Salience of PCs differs by as much as 0.432 (PC 24), with the differ- ence in the salience of the first 8 PCs (31% variance explained) ran- ging from 0.200 (PC1) to 0.309 (PC7).”

May combine two claims.

The soil subnetwork from CO was notably deeper than the one from SO (7 vs 2 dense layers) but narrowed more by the last processing layer (2 vs 27 units).

“The soil subnetwork resulting from CO is notably deeper than the one from SO (7 and 2 dense layers, respectively) but also narrows more by the last processing layer (2 vs 27 units).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Following training and test set definition, 41,513 observations remained: 37,273 in training and 4,240 in test.

“Following the training and test set definition, 41,513 observations remained: 37,273 in the training set and 4,240 in the test set.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The disagreement between epochs minimizing the two validation-loss criteria ranged from 2 epochs for CO Genomic and CO interaction models to 404 epochs for CO weather and management model.

“The disagreement between the epochs which minimized these values ranged from 2 epochs in the case of the CO Genomic model and CO interaction model up to 404 epochs for the CO weather and management model.”

May combine two claims.

The genomic subnetworks from CO and SO both had 2 layers, but CO widened somewhat (layer 1 = 83 units, 16% dropout; layer 2 = 133 units, 23% dropout) while SO began over twice as wide and constricted more (layer 1 = 196 units, 15% dropout; layer 2 = 47 units, 6% dropout).

“Genomic subnetworks resulting from CO and SO are both 2 layers, but the CO model widens somewhat (layer 1 = 83 units, 16% dropout, layer 2 = 133 units 23% dropout) while the SO model begins over twice as wide and constricts more (layer 1 = 196 units, 15% dropout, layer 2 = 47 units 6% dropout).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Observations per site in the full dataset ranged from 156 at GAH2 in 2016 to 3,589 at MNH1 in 2018, with median 498 and mean 608.5 observations per year.

“Observations per site ranged from 156 at”

May combine two claims.

28 of 41 total sites are exclusively in training data and account for 23,758 observations; shared sites account for 13,515 observations; no sites are exclusive to test.

“28 of the 41 total sites are exclusively found in the training data and account for 23,758 observations with the shared sites accounting for 13,515 observations.No sites are exclusive to the test set.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Except for the SO DNN, the alternate heuristic also resulted in longer training durations.

“With the exception of the SO DNN, this also resulted in longer training durations.”

May combine two claims.

Of 3,006 genotypes, 1,559 occur in both sets accounting for 27,131 training and 4,221 test observations; 1,435 genotypes (10,142 observations) are only in training; 12 genotypes (19 observations) are exclusive to test.

“Of the 3,006 genotypes present, 1,559 occur in both sets accounting for 27,131 observations in the training set and 4,221 in the test set.1,435 genotypes (10,142 observations) are only found in the training set whereas 12 genotypes (19 observations) are exclusive to the test set.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Within a location, 1-2 replicates per genotype was typical (median 2, mean 1.62), but could reach 46.

“Within a location 1-2 replicates per genotype is typical (median of 2, mean of 1.62) but ranges as high as 46 replicates (2369/LH123HT at”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The top 5 most represented genotypes had 503, 232, 184, 173, and 172 observations; median and mean genotype had 9 and 13.8 observations.

“The number of replicates for each genotype varies with the top 5 most represented genotypes being 2369/LH123HT (503), PHW52/PHN82 (232), PHW52/PHM49 (184), B73/PHN82 (173), and B73/MO17 (172) and the median and mean genotype having 9 and 13.8 observations, respectively.”

May combine two claims.

The two training durations closest to optimum were the CO genomic model (2 epochs over) and the SO model (77 epochs over), performing at 100.2% and 101% of minimum.

“The 2 training durations closest to the optimum were the CO genomic model (2 epochs over) and the SO model (77 epochs over).These models performed at just 100.2 and 101% minimum.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Replicate numbers per genotype varied from 1 to several hundred, with median 18 and mean 26.2 observations.

“The number of replicates for each genotype varies widely from 1 to several hundred (e.g.2369/LH123HT: 882, PHW52/ PHN82: 421, B37/H95: 333, PHW52/PHM49: 331, B73/PHN82: 328), with a median of 18 and mean of 26.2 observations.”

May combine two claims.

The simple linear fixed effects model ranked third (RMSE 0.973, nRMSE 14.933%) and DNN-SO ranked fourth (RMSE 1.024, nRMSE 15.716%), followed by the machine learning models.

“The simple linear fixed effects model ranked third (RMSE 0.973, nRMSE 14.933%) and DNN-SO ranked fourth (RMSE 1.024, nRMSE 15.716%) followed by the machine learning models tested (Fig. 2, Table 5).”

May combine two claims.

The first 8 genomic principal components explain 31% of variance, the first 50 explain 50%, and 1,725 PCs explain over 99%.

“We find that 31% of the variance is explainable by the first 8 principal components (PCs), 50% is explainable by the first 50 PCs, and >99% of the variance is explainable by 1,725 PCs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The soil dataset has an average completion rate of 47% across all site-by-year combinations.

“The soil dataset contains many missing values, having an average completion rate of 47% across all site-by-year combinations.”

May combine two claims.

The SO and CO weather models had the largest differences between optimal and used epoch numbers (697 and 563 epochs, respectively) and achieved 121% and 110% of the minimum possible RMSE.

“The SO and CO weather models had the largest differences between optimal and used epoch numbersdifferences of 697 and 563 epochs respectively and achieved 121 and 110% of the minimum possible RMSE.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The CO soil model trained 185 epochs beyond optimal but had RMSE at 102% of minimum.

“The CO soil model trained an excess of 185 epochs but only had RMSE at 102% minimum.”

May combine two claims.

The BLUP had about four times the standard deviation in RMSE across replicates (0.058) as the CO DNN (0.013), making the CO DNN more consistent.

“However, the standard deviation in performance across replicates is about 4 times higher for the BLUP (SD RMSE 0.058) than the CO DNN (SD RMSE 0.013) making the CO DNN a more consistent performer than the BLUP model.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The balancing resulted in 3,006 genotypes represented, a reduction of 665 genotypes.

“This also resulted in fewer genotypes being represented (3,006 total, a reduction of 665 genotypes).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The data were approximately balanced: all site-year groups in the test set had 265 observations, and no training site-year group exceeded this.

“These data were approximately balanced as described above with all site-year groups in the test set having 265 observations, and no site-year group in the training set exceeding this, although some did contain fewer observations.”

May combine two claims.

The SO model overfits faster and to a greater extent than the full CO model, which did not show evidence of substantial overfitting.

“The SO model overfits faster and to a greater extent than the full CO model, which does not show evidence of substantial overfitting (Fig. 1b, d and e).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Most site-years contain a single replicate of a genotype (mean observations 1.27).

“Most site-years contain a single replicate of a genotype (mean observations: 1.27).”

May combine two claims.

The SO model achieved a lower loss than the CO model, and its accuracy worsened rapidly with further training.

“The SO model achieves a loss lower than the CO model, and the accuracy worsens rapidly with further training.”

The quote only partly states this claim.

Allowing multiple data types in the CO DNN reduced average RMSE by 0.070 (1.073% nRMSE) relative to the next best DNN, the CO weather/management model.

“Allowing the use of multiple data types in the CO DNN reduced average RMSE by 0.070 (1.073% nRMSE) relative to the next best DNN, the CO”

The quote only partly states this claim.

All models outperformed the intercept model on soil data.

“All models outperformed the intercept model.”

May combine two claims.

The RF model was an exception, with RMSE 0.373 and nRMSE 5.729% above the intercept model.

“The RF model was a clear exception to this, achieving an RMSE of 0.373, nRMSE 5.729% above the intercept model.”

May combine two claims.

The CO network had higher average R2 (0.171 vs 0.032) and more consistent R2 across replicates (SD 0.022 vs 0.065).

“Similarly, average R 2 was higher in the CO network (CO: 0.171 vs SO: 0.032) and more consistent across replicates as well (standard deviation CO: 0.022 vs SO: 0.065).”

May combine two claims.

The full dataset's 96,137 observations spanned 41 sites, 6 years (2014-2019), and 158 site-year combinations.

“In the full dataset (available at 10.5281/zenodo.6916775)the 96,137 observations were spread over 41 sites across 6 years (2014-2019) (158 site-year combinations).”

May combine two claims.

The two most salient PCs were PC 26 (0.423) and PC 24 (0.402), which accounted for 0.350% and 0.392% of total genomic variance, respectively.

“The 2 most salient PCs are PC 26 (0.423) and PC 24 (0.402) which account for 0.350 and 0.392% of the total genomic variance respectively.”

May combine two claims.

The most salient region of time was within the first few days after planting; 8 of the 10 days with highest average salience were days 2-9 following planting.

“Additionally, we find specific time points which appear to be salient broadly with the most salient region of time being within the first few days of planting, indeed 8 of the 10 days with the highest average salience are days 2-9 following planting.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Yield values had a mean of about 147.397 bushels per acre and standard deviation of about 48.169 bushels per acre.

“In the case of yield, these values were a mean of ∼147.397bushels per acre and a standard deviation of ∼48.169 bushels per acre, i.e. y = (y Original -147.397)/48.169.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The additional training durations ranged from 2 extra epochs for the CO genomic and interaction models to 404 extra epochs for the CO weather/management model.

“These ranged from an additional 2 epochs in the cases of the CO genomic and interaction models and as many as 404 epochs in the case of the CO weather/management, as shown in Table 6.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the weather and management subnetwork, CO resulted in a notably deeper network (6 pairs vs 2 pairs of convolution layers) but used a similar number of filters in the final convolution layer pairs (CO 294 vs SO 303).

“Finally, in the weather and management subnetwork CO resulted in a notably deeper network (6 pairs vs 2 pairs of convolution layers) but used a similar number of filters in the final convolution layer pairs (CO 294 vs SO 303).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Across all observations, 3,671 unique genotypes were recorded, derived from 1,681 female and 223 male parent genotypes.

“Across all observations, 3,671 unique genotypes were recorded, derived from 1,681 female and 223 male parent genotypes.”

May combine two claims.

The CO network had lower normalized errors (nRMSE 14.6% vs 15.7%) and lower standard deviation (0.197% vs 0.531%).

“Similar results were seen in the normalized errors (nRMSE CO: 14.6% SO: 15.7%, standard deviation CO: 0.197%, SO: 0.531%).”

May combine two claims.

The CO network was better at predicting yield in the testing set, with lower mean RMSE (0.948 vs 1.024) and more consistent accuracy across replicates (SD 0.013 vs 0.035).

“The CO network was better at predicting yield in the testing set.It achieved a lower mean RMSE (CO: 0.948 vs SO: 1.024) and was more consistently accurate across replicates (standard deviation CO: 0.013 vs SO: 0.035).”

Probably not a substantial claim: may not pass S3 (load-bearing).

These hybrids resulted from 1,238 female and 68 male parent genotypes.

“These hybrids resulted from 1,238 female and 68 male parent genotypes.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

94,996 (98.8%) observations were from hybrids; inbreds accounted for 1,141.

“94,996 (98.8%) of the observations are from hybrids; inbreds account for the other 1,141.”

May combine two claims. The quote only partly states this claim.

The majority of weather/management factors had average salience between 0.140 and 0.10, with six below that threshold: average soil temperature (0.095), maximum wind speed (0.084), average soil moisture (0.076), phosphorus applied (0.052), and potassium applied (0.033).

“The majority of factors had an average salience between 0.140 and 0.10 with 6 falling below this thresholdaverage soil temperature (0.095), maximum wind speed (0.084), average soil moisture (0.076), phosphorus applied (0.052), and potassium applied (0.033).”

May combine two claims.

In linear models, the improvement relative to the next best model of the same type was 0.046 RMSE (0.699% nRMSE) for fixed effects models and 0.007 RMSE (0.111% nRMSE) for BLUPs.

“In linear models, the improvement relative to the next best model of the same type is 0.046 RMSE (0.699% nRMSE) for fixed effects models and 0.007 RMSE (0.111% nRMSE) for BLUPS.”

May combine two claims.

RF and RNR performed best with soil data only; using all data reduced performance by 0.024 RMSE (0.375% nRMSE) and 0.012 RMSE (0.185% nRMSE), respectively.

“RF and RNR perform best with soil data only the use of all data reduces performance by 0.024 RMSE (0.375% nRMSE) and 0.012 RMSE (0.185% nRMSE) respectively.”

May combine two claims.

While the best performance was achieved with a BLUP incorporating genomic, soil, weather, and management data, a DNN performed similarly to the BLUP and with less variance in accuracy between model replicates.

“While the best performance was achieved with a BLUP incorporating genomic, soil, weather, and management data, a DNN performed similarly to the BLUP model and with less variance in accuracy between model replicates (Fig. 2).”

May combine two claims.

The SO DNN had a different weather submodule and a higher average RMSE than the weather/management model (by 0.006 RMSE, 0.089% nRMSE) but was more consistent (SD 0.035 vs 0.074).

“The SO DNN contains a different weather submodule and has a higher average RMSE than the weather/management model (by 0.006 RMSE, 0.089% nRMSE) but performed more consistently with a standard deviation of RMSEs 0.035 relative to 0.074.”

May combine two claims.

kNN and SVR performed best with weather and management data only; using all data reduced performance by 0.014 RMSE (0.218% nRMSE) and 0.056 RMSE (0.862% nRMSE), respectively.

“kNN and SVR perform best with weather and management data only with the use of all data reducing performance by 0.014 RMSE (0.218% nRMSE) and 0.056 RMSE (0.862% nRMSE) respectively.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the authors' hands, total time spent tuning was driven more by modules with computationally intensive components (convolution layers) rather than the number of modules to optimize.

“In our hands, the total time spent tuning was driven more by modules with computationally intensive components (convolution layers) rather than the number of modules to optimize.”

May combine two claims.

The two DNNs with interactions are more consistent, with much lower RMSE dispersion: standard deviations of 0.013 for the CO model and 0.035 for the SO model, compared with 0.074 for the CO weather model.

“However, the 2 DNNs with interactions are more consistent having a far lower dispersion in RMSE, with standard deviations of 0.013 in the CO model and 0.035 in the SO model as compared with 0.074 in the CO weather model.”

May combine two claims.

Compared with the submodel, incorporating interactions appears to raise the salience of irrigation, though irrigation is highly salient in both models.

“Relative to the submodel, incorporating interactions appears to increase the salience of irrigation, although it is highly salient in both models (relative to other time-series factors).”

May combine two claims.

With only genomic information, these models perform worse than a simple intercept-only model and improve only slightly when soil data are added.

“When restricted to exclusively genomic information these models underperform simple intercept-only models and only improve slightly with the use of soil data.”

May combine two claims.

Weather and management data reduces error substantially relative to genetic and soil data, but the variation in performance is large.

“We note that weather and management data does reduce error substantially relative to genetic and soil data, but the variation in performance is large (Fig. 2).”

May combine two claims.

The consecutively optimized DNN model yields more consistent yield estimates than the best BLUP model, despite a slightly higher average error.

“The consecutively optimized DNN model developed here shows promise for complementing existing models for crop selection and improvement, as it produces more consistent estimates of yield, despite having a slightly higher average error than the best BLUP model.”

May combine two claims.

Nitrogen's mean salience is comparable to that of relative humidity and photoperiod, whereas phosphorus and potassium have much lower mean salience.

“we observe that nitrogen does have a mean salience comparable to relative humidity and photoperiod, while phosphorus and potassium are far lower.”

May combine two claims.

The BLUP required over 100 GB of RAM to fit while the DNN required GPUs, and fitting took hours to days instead of minutes.

“This is true in terms of hardware (the BLUP required over 100 GB of RAM to fit while the DNN required the use of GPUs) and time (hours to days instead of minutes to fit).”

May combine two claims.

The daily average solar radiation is the next most salient feature of this dataset, followed by maximum temperature and average wind direction.

“The daily average of solar radiation (SolarRadiationMean) is the next most salient feature of this dataset, followed by the maximum temperature (TempMax) and the average wind direction (WindDirectionMean).”

May combine two claims.

Both strategies failed to select the epoch that gave the minimum test-set loss for all submodels and led to apparent overfitting in the Weather and Management submodel and the SO model.

“Both strategies considered failed to select the epoch resulting in the minimum loss in the test set for all submodels and resulted in apparent overfitting in the Weather and Management submodel (c.) and the SO model (e.).”

May combine two claims.

Interaction-containing BLUPs have a lower average error (RMSE 0.937, nRMSE 14.388%) than the weather and management model (RMSE 0.945, nRMSE 14.500%), but a higher standard deviation of RMSE (0.058 vs 0.040).

“Allowing for interactions further improves performance with interaction-containing BLUPs having a lower average error (RMSE 0.937, nRMSE 14.388%) than the weather and management model (RMSE 0.945, nRMSE 14.500%), albeit with a higher standard deviation of RMSE (0.058 vs 0.040).”

May combine two claims.

A simple linear fixed effects model also performed fairly well and required substantially fewer computational resources and model selection efforts.

“A simple linear fixed effects model also performed fairly well and required substantially fewer computational resources and model selection efforts.”

May combine two claims.

The simultaneously optimized model reaches a lower minimum error than the selected model and does so in far fewer epochs, but it overfits much faster.

“The simultaneously optimized model achieves a minimum error lower than our selected model (see Fig. 1b) and does so in far fewer epochs, but overfits much faster as well.”

May combine two claims.

The CO DNN with interactions (RMSE 0.948, nRMSE 14.554%) and the weather and management model (RMSE 1.018, nRMSE 15.627%) both have lower average error than the SO Model (RMSE 1.024, nRMSE 15.715%).

“A similar result is seen between the CO DNN with interactions (RMSE 0.948, nRMSE 14.554%) and weather and management model (RMSE 1.018, nRMSE 15.627%), both of which achieve lower average error than the SO Model (RMSE 1.024, nRMSE 15.715%).”

May combine two claims.

The CO and SO strategies produce substantially different hyperparameter values and therefore different network architectures, even though the training data are the same.

“Data available for training are the same but the CO and SO strategy result in substantially different hyperparameter values and thus network architecture.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The submodel has additional salience peaks before planting and near the end of the date range considered.

“The submodel contains additional peaks of salience prior to planting and near the end of the considered date range.”

May combine two claims.

Deep learning is a promising route for predicting complex traits in complex environments and may help reveal environmental and genetic influences.

“Based on these results, deep learning provides a promising avenue for the phenotypic prediction of complex traits in complex environments and a potential mechanism to better understand the influence of environmental and genetic factors.”

May combine two claims.

DNNs learn directly from data, which reduces the need for feature engineering and dimensionality reduction.

“Additionally, DNNs”

Probably not a substantial claim: may not pass S3 (load-bearing).

For distortions not randomly distributed, filtering may bias the sample and result in a model that appears to perform well but generalizes poorly.

“Note, however, that for distortions that are not randomly distributed, filtering may bias the sample and result in a model that appears to perform well but generalizes poorly (e.g. to sites similar to those with a preponderance of observations excluded).”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

If balance were not a concern, or if it could be achieved without discarding observations, one could employ more strict data filtering without producing a dataset too small for machine learning.

“Indeed, if the balance were not a concern, or if it could be effectively achieved without discarding observations in some groups, one could potentially employ more strict data filtering without producing a dataset too small to benefit from machine learning.”

• AssumptionThe authors state this as an assumption: The distribution of sites collected matches the distribution of sites of interest for prediction, and the number of observations per field site is representative of a future number of observations.

Classified as probably not a substantial claim.

Train/test splits were generated randomly with the constraint that any location-year combination could appear in only the testing or training set.

“We generated train/test splits randomly, with the constraint that any location-year combination could appear in only the testing or training set.”

Classified as probably not a substantial claim. May combine two claims.

The validation set was used to assess hyperparameter performance without compromising the test set, and this was repeated several times to prevent overfitting to a single validation set.

“The validation set is used to assess the performance of a considered set of hyperparameter values without compromising the test set.This was repeated several times to prevent overfitting to a single validation set.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The training set was split into training and validation sets stratifying by site-group-by-year groups for hyperparameter selection.

“For use in hyperparameter selection, the training set was split into a training and validation set, stratifying by site-group-by-year groups.”

Classified as probably not a substantial claim.

The 41 experimental sites were grouped into 29 site groups using GPS distance such that no two groups were within 0.5 degree.

“The 41 experimental sites were grouped into 29 experimental site groups using the distance between their GPS coordinates such that no 2 groups were within 0.5 degree from each other.”

Classified as probably not a substantial claim.

Observations within site-by-year groups were kept in the same partition of the data.

“To this end, we kept observations within site-by-year groups in the same partition of the data.”

Classified as probably not a substantial claim.

Except where soil features are static from season to season, the model will not be trained and tested on observations with identical soil features.

“Furthermore, except in cases where soil features are static from season to season, the model will not be trained and tested on observations with identical soil features.”

May combine two claims.

Data transformation before cross-validation folds may introduce a different information leak in hyperparameter selection, but this does not affect final evaluation because test set data were not used in those calculations.

“As data transformation occurs prior to determining cross-validation folds a different information leak is potentially introduced within the hyperparameter selection process but does not create an issue for the final evaluation of the model's performance because the test set data were not used in these calculations.”

May combine two claims.

Training and test sets do not overlap with respect to site-year combinations but share sites and genetics.

“The training and test sets do not overlap with respect to site-year combinations but share sites and genetics.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Dropout can reduce overfitting in fully connected layers.

“Where fully connected layers are used, overfitting can be reduced by randomly removing neurons from a layer with a certain”

• AssumptionRests on an assumption: DNNs are prone to overfitting to training data, which can cause poor performance.

Classified as probably not a substantial claim. May combine two claims.

The best-performing 4 hyperparameter sets for each model were trained for 1,000 epochs and evaluated on 10 defined testing/validation set splits.

“The best-performing 4 hyperparameter sets for each model were trained for 1,000 epochs and evaluated on 10 defined testing/validation set splits.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A custom subclassed tuner randomly selected one of the previously defined validation folds to prevent overfitting to a single validation set without increasing computational cost.

“We used a custom subclassed version of the tuner to randomly select one of the previously defined validation folds to prevent overfitting to a single validation set without increasing the computational cost.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Neural networks have been used for yield prediction since at least 2001.

“While neural networks have been applied to the problem of predicting yield since at least 2001 (Liu et al. 2001)”

• cites Liu et al. 2001 sourceNot checked — no full text of the source

Classified as probably not a substantial claim. May combine two claims.

All site-group-by-year combinations were randomly downsampled so none had more observations than the smallest number in the testing set, to prevent overrepresentation.

“To prevent over representation of certain groups, all site-group-by-year combinations were randomly down sampled so that none had more observations than the smallest number in the testing set.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Convolutional layers lower overfitting risk because they work well with fewer parameters than fully connected layers.

“For example, the use of convolutional layers minimizes the potential of overfitting because they perform well with fewer parameters relative to fully connected layers.”

• AssumptionRests on an assumption: DNNs are prone to overfitting to training data, which can cause poor performance.

Probably not a substantial claim: may not pass S3 (load-bearing).

The scale of the G2F project increases the chances of data being absent or compromised due to equipment malfunction, logistical or procedural issues, and resource constraints.

“The scale of this ambitious project increases the chances of data being absent or compromised due to equipment malfunction, logistical or procedural issues, and resource constraints.”

Probably not a substantial claim: may not pass S4 (based).

Incorporating multiple data sources and allowing interactions between data types generally appeared to improve accuracy for DNNs, linear fixed effects models, and BLUPs.

“Incorporating multiple data sources and allowing interactions between data types generally appears to improve accuracy for DNNs, linear fixed effects models, and BLUPs.”

Classified as probably not a substantial claim.

Input data were centered and scaled based on the mean and standard deviation of each parameter in the training data prior to hyperparameter selection and training.

“Prior to hyperparameter selection and training the input data were centered and scaled based on the mean and standard deviation of each parameter in the training data.”

May combine two claims.

The full CO model performed best by a small margin and allowed direct comparison of interaction effects on factor salience because the single-modality subnetworks were identical except for the prediction layer.

“The full CO model, in addition to performing best (albeit by a small margin), presents an opportunity to directly compare the influence of interactions between data modalities on the salience of factors because the single modality subnetworks are identical except for the prediction layer.”

Classified as probably not a substantial claim. May combine two claims.

The CO and SO hyperparameter selection strategies had the same range of possible networks, the same data driving network selection, and both used Bayesian optimization.

“Two hyperparameter selection strategies were employed, CO and SO, which have the same range of possible networks (hyperparameter ranges are listed in Table 1), the same data driving network selection and both use Bayesian optimization.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The 0.5 degree distance threshold is sufficient to group nearby locations while excluding distant ones.

“While this threshold is inconsistent with respect to location it is sufficient to group nearby locations while excluding distant ones.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Different network sizes (CO containing more layers) may account for the difference in overfitting.

“The different network sizes (CO containing more layers) may account for this difference.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

With selected hyperparameters and training duration, each model was fit 10 times to account for random initialization, and each replicate and training history saved.

“With the selected hyperparameters and training duration we fit each model 10 times to account for random initialization and saved each replicate and its training history.”

Probably not a substantial claim: may not pass S3 (load-bearing).

CO reduces the range of possible combinations explored by allowing only one module to vary at a time.

“CO reduces the range of possible combinations that are explored by allowing only one module to vary at a time.”

Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

Similar to their decision to minimally transform and filter the data, they elected to avoid nudging the architecture of the network in any direction in order to allow the data to inform it instead.

“Similar to our decision to minimally transform and filter the data, we elected to avoid”

May combine two claims.

Different models may be better suited to different goals such as achieving good predictive performance consistently (DNN-CO, multiple data types) or maximizing average predictive performance (BLUP, multiple data types) while potentially underperforming expectations due to variability in performance across model replicates.

“Furthermore, different models may be better suited to different goals such as achieving good predictive performance consistently (DNN-CO, multiple data types) or maximizing average predictive performance (BLUP, multiple data types) while potentially underperforming expectations due to variability in performance across model replicates.”

May combine two claims.

The SO model is capable of performing comparably to the CO model, but overfits more rapidly.

“The SO model is capable of performing comparably to the CO model, but overfits more rapidly (Fig. 1b).”

Probably not a substantial claim: may not pass S3 (load-bearing).

As a pragmatic matter, CO benefits from the capacity to tune multiple modules at once.

“As a pragmatic matter, CO benefits from the capacity to tune multiple modules at once.”

May combine two claims.

The observed performance differences suggest a substantial environmental effect, which is intensified by splitting observations from site-by-year groups exclusively into training or testing sets.

“These performance differences are suggestive of a substantial environmental effect, exacerbated by the stratifying observations from site-by-year groups into exclusively training or testing sets.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

SO represents the reverse situation: with all features available, interactions between features in different tensors can be leveraged, but the hyperparameter space to explore is larger as all hyperparameters are free to vary.

“SO represents the reverse situation.With all features available, interactions between features in different tensors can be leveraged, but the hyperparameter space to explore is larger as all the hyperparameters are free to vary.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Linear modeling results suggest a positive association between precipitation early in development and yield.

“Linear modeling results find similar results and suggest a positive association between precipitation early in development and yield (Rogers et al. 2021).”

• cites Rogers et al. 2021Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Previous DNNs developed with a subset of G2F data also identified precipitation as substantially influencing yield.

“Previous DNNs developed with a subset of G2F data also identified precipitation as substantially influencing yield (Washburn et al. 2021).”

• cites Washburn et al. 2021 sourceNot checked — no full text of the source

May combine two claims.

Only after integration of all data types do they see a relative reduction in error and consistency in this reduction in DNNs.

“Only after integration of all data types do we see a relative reduction in error and consistency in this reduction in DNNs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Although they aimed to broaden the range of possible architectures relative to previous modeling on G2F data, they constrained the overall structure to process each tensor individually then allow interactions between the final layer of each module.

“Although we aimed to broaden the range of possible architectures relative to previous modeling on G2F data (Washburn et al. 2021), we constrained the overall structure to process each tensor individually then allowing for interactions between the final layer of each module.”

• cites Washburn et al. 2021 sourceNot checked — no full text of the source

May combine two claims.

The SO DNN architecture was selected based on 40 trials whereas the CO DNN architecture was selected based on 40 trials for each module (160 trials across the whole network), which confounds the comparison.

“The SO DNN architecture was selected based on 40 trials whereas the CO DNN architecture was selected based on 40 trials for each module (160 trials across the whole network) which confounds the comparison.”

Probably not a substantial claim: may not pass S3 (load-bearing).

They used a Bayesian optimization procedure that aims to produce useful hyperparameter combinations in fewer cycles than a simpler method such as grid approximation.

“We used a Bayesian optimization procedure that aims to produce useful hyperparameter combinations in fewer cycles than a simpler method such as grid approximation.”