Almonds

The self-administered FFQ includes question items on the average frequency of consumption during the past year with the following eight possible responses: almost never, 1–3 times per month, 1–2 times per week, 3–4 times per week, 5–6 times per week, once per day, twice per day, and ≥ 3 times per day. These responses are then converted into intake scores of 0, 0.1, 0.2, 0.5, 0.8, 1, 2, and 3, respectively, to approximate the intake frequency per day. The consumption of foods is tabulated in grams for 20 food groups based on the Standard Tables of Food Composition in Japan (seventh revised edition) [14 ]. The food groups (shown as the number of items: description) were rice (1), bread (1), noodles (1), potatoes (1), soybean products (4: tofu in miso soup, tofu dishes, fermented soybeans [nattō], and fried tofu [ganmodoki]), green vegetables (5: pumpkin, carrot, broccoli, green-leaf vegetables, and other green-yellow vegetables), other vegetables (5: cabbage, radish, dried radish [kiriboshi-daikon], bamboo shoots, and other vegetables), fruit (2: citrus fruits and other fruits), mushrooms (1), seaweeds (1), fish (7: fish, bone-edible small fish, canned tuna, octopus/shrimp/crab, shellfish, fish eggs, and fish paste products), meat (4: chicken, beef/pork, liver, and ham/sausage/bacon), eggs (1), milk (2: milk and yogurt), oils (6: margarine, butter, mayonnaise, deep-fried dishes, light-fried dishes/sauté, and peanuts/almonds), confectionery (2: Western- and Japanese-style confectioneries), green tea (1), coffee (1), alcoholic beverages (1), and soybean paste (1).
The FFQ contains no question items on usual portion size for 43 food items, so we applied the standard portion sizes based on DRs in a population from Aichi Prefecture [3 (link)]. However, portion sizes are requested for three kinds of staple foods in Japan (rice, bread, and noodles). The daily consumption of each food item was computed by multiplying the portion size by the intake score. For alcoholic beverages, the amount and frequency per week or month were asked for the following 10 items: sake, Japanese liquor (shōchū), shōchū highball, large bottle of beer (633 mL), medium-sized bottle of beer (500 mL), 350 mL of canned beer, 250 mL of canned beer, single whiskey, double whiskey, and wine. Sugar-sweetened beverages (SSBs) were not included in the short FFQ.

The P. dulcis genome assembly was annotated by combining transcript alignments, protein alignments and ab initio gene predictions. A flowchart of the annotation process is shown in Figure S13. Scripts are available at https://github.com/jesgomez/annotation_pipeline.
First, almond RNA‐seq reads were downloaded from NCBI with the accession number SRR1251980 and aligned to the genome with STAR (v.2.5.3a) (Dobin et al., 2013). Transcript models were subsequently generated using Stringtie (v.1.0.4) (Pertea et al., 2015) and, along with the P. persica transcriptome (annotation Pp2.0a) and 4509 almond expressed sequence tags downloaded from NCBI on July 2015, were assembled into a non‐redundant set by PASA (v.2.3.3) (Haas et al., 2008). The TransDecoder program, which is part of the PASA package, was run on the PASA assemblies to detect coding regions in the transcripts. Second, the complete Rosaceae proteome was downloaded from Uniprot on July 2015 and aligned to the genome using Exonerate (v.2.4.7) (Slater and Birney, 2005). Third, ab initio gene predictions were performed on the repeat masked pdulcis26 assembly with three different programs: GeneID v.1.4 (Alioto et al., 2018), Augustus v.3.2.3 (Stanke et al., 2015) and GeneMark‐ES v.2.3e (Lomsadze et al., 2014) with and without incorporating evidence from the RNA‐seq data. Finally, all the data were combined into consensus coding sequence models using EvidenceModeler‐1.1.1 (EVM) (Haas et al., 2008). Additionally, untranslated regions and alternative splicing forms were annotated through two rounds of PASA annotation updates.
Non‐coding RNAs were annotated as follows: first, the program cmsearch v.1.1 (Cui et al., 2016) from the INFERNAL package (Nawrocki and Eddy, 2013) was run against the RFAM (Nawrocki et al., 2015) database of RNA families (v.12.0). Also, tRNAscan‐SE v.1.23 (Lowe, 1997) was run to detect the transfer RNA genes present in the genome assembly. To annotate long non‐coding RNAs (lncRNAs) we first selected PASA assemblies that had not been included in the annotation of protein‐coding genes. Those longer than 200 bp and whose length was not covered to at least 80% by a small ncRNA were incorporated into the ncRNA annotation as lncRNAs. The resulting transcripts were clustered into genes using shared splice sites or significant sequence overlap as criteria for designation as the same gene.

TensorFlow (Abadi et al., 2016 ), Keras (Chollet, 2015 ), and KerasTuner (O’Malley et al., 2019 ) libraries in Python were used for CNN model tuning and training processes. The CNN model took the image blocks, centered around each individual almond tree crown, from CERES images at 0.3 m resolution, for 4 reflectance bands (R, G, NIR, and RE) as inputs to estimate the individual tree almond yield (Figure 3). We started with the minimum block size of 21 × 21 pixels, equivalent to a 3m radius centered around each tree crown center and thus representing areas slightly bigger than one tree crown size. For each tree sample, we first identified the corresponding CERES pixel containing the tree center (as described in Section 2.3 location), and then clipped an image block extending 10 pixels towards all four directions from the center, for each band. This step resulted in 21 × 21 × 4 multi-spectral imagery associated with each individual tree crown as the input to the CNN model.
The CNN model training process is to find kernels in the convolutional layers and weights in the dense layers to minimize the differences between model estimations and ground measurements on a training dataset. The Mean Squared Error (MSE) loss function was applied for the CNN model training, which calculates the average of the squared differences between model estimations and actual values. To efficiently optimize the kernels and weights within the CNN model, the Adam optimization algorithm (Kingma and Ba, 2014 ) is used, which extends the stochastic gradient descent algorithm by calculating individual learning rates for different parameters based on the estimates of first and second moments of gradients. 5-fold cross validation (CV) is applied to randomly split the data into separate training and testing sets. The overall model performance is evaluated based on the average performance over the testing set in each fold. The Bayesian optimization algorithm is developed to select the CNN hyper-parameters automatically.
The general setup of the possible CNN structures for the Bayesian optimization algorithm are as follows: three to four convolutional blocks followed by a spatial attention module with a global average pooling layer and two fully connected dense layers. For the first dense layer, there are 30 to 100 neurons followed by a dropout layer. For each convolutional block, there are 16 to 128 convolutional layers (kernels) followed by a batch normalization and pooling layers, then another 16 to 128 convolutional layers followed by a batch normalization, pooling and ReLU activation layers. The pooling layers in each convolutional block can be either average pooling or max pooling. The overall architecture of the CNN model for the Bayesian optimization algorithm is shown in Figure S3. For model compiler, the Bayesian optimization algorithm selects learning rate varying from 10^-4 to 10^-2 with Adam optimizer. For the Bayesian optimization algorithm itself, the maximum trail number was set to 50, and for each trail, the batch size is 128 with 100 epochs.
To investigate the impact of input image block size used for the CNN model and explore how the neighboring trees potentially influence yield estimation, another two separate CNN models were built with an input image size of 41 × 41 pixels (roughly 6m radius) and 61 × 61 pixels (9m radius), respectively. To understand the contribution of the red edge band to the yield estimation, a reduced CNN model was constructed by excluding red edge reflectance as input, hereafter called “reduced CNN model”, considering that red edge band is not as widely used for aerial imaging as the other three bands. Similarly, another 14 sets of reduced CNN models were further built with all the combinations of different reflectance bands as input and compared how they influenced model’s yield estimation accuracy (Table S2).

To evaluate models’ performance in predicting almond yield, the predicted and observed individual tree yield from the reserved testing samples were compared, and the coefficient of determination (R²), Root Mean Squared Error (RMSE), and RMSE normalized by averaged yield measurement (NRMSE) were calculated. Statistics of these metrics were reported based on 5-fold cross validation.
For the model with highest accuracy, its capability to capture the within-field yield variations, such as overall spatial patterns, row to row variations, and tree to tree variations along selected transects was also evaluated. For all harvested rows, the yield distribution for all trees within each individual row was analyzed based on CNN estimations. Furthermore, three transects that are perpendicular to the row orientation of the orchard were randomly selected to examine the inter-row yield variations. The locations of the selected transects are shown in Figure 4 highlighted in blue lines.