AutoML-based Almond Yield Prediction and Projection in California

California is the most significant almond-producing region in the world. Eighty percent of the almonds consumed globally and all commercially sold in the United States are produced in California. Almonds were California’s most profitable agricultural export product and the second-largest commodity overall in 2020 [1]. Previous studies have shown that climate factors can influence almond growth [2, 3], and when these climate factors change due to global warming, the yield of almonds will inevitably suffer. Thus, a precise projection is necessary for both scientific research and climate adaptations.

Because of the difficulty in simulating the growth process of perennial crops, very few process-based crop models can be utilized to investigate the relationship between climatic variables and almond yields. Data-driven approaches, including machine learning (ML) and deep learning (DL) models, have gained popularity in Earth system modeling since the emergence of artificial intelligence due to their capacity to fit nonlinear functions. Many studies have applied ML and DL models for various tasks, such as hydrology projection, air quality analysis and severe weather forecasting [4, 5, 6]. The ML and DL models can be designed in an end-to-end manner to examine the relationship between input and output variables. While this strategy is often effective, a proper model architecture is required, which often involves specialized knowledge. Automated machine learning (AutoML) frameworks are designed to save these efforts and to facilitate the use of ML models by researchers without a background in computer science, as is frequently the case for climate studies.

The goal of this study is to build an AutoML framework that creates pipelines from California’s climatic variables to estimate almond yields. Section 2 introduces the data sources and AutoML model. In section 3, the model is trained and tested with observational records, and future projections are in section 4.

The coefficient of determination ($R^{2}$ score) and root mean squared error (RMSE) are used to quantify the model predictability, and AutoGluon is compared against random forest and linear regression. A 5-fold cross-validation approach is used together with a train-test split method (70%-30%) to benchmark the ML models. There are various presets in AutoGluon that specify the hyperparameters. In general, high accuracy corresponds to longer training time and larger models for disk consumption. The choice of presets may affect the model performance, but to maintain simplicity and automation, we only select one preset (‘high-quality’) in this study.

As listed in Table 1, both the AutoGluon and random forest have relatively lower skill with the train-test split, with the exception of linear regression. This is probably due to the decrease in the number of training samples, since 80% samples are used as the training set in cross-validation, whereas only 70% are used for the train-test split. Notably, the AutoGluon model achieves the best performance for both the cross-validation and train-test settings.

Projections under various climate change scenarios are produced using the trained AutoGluon model and MACA forcing data. Figure 2 shows the comparison of the historical simulation with observations. The results from 17 MACA models are generally consistent with one another, with a small inter-quantile range, and are effective at capturing the trend in almond yield. Comparing the ensemble mean simulations, RCP4.5 is slightly closer to observations ($R^{2}=0.5205$) than RCP8.5 ($R^{2}=0.5199$). Despite the fact that these values are lower than those in Table 1, it is to be expected given that these climate models are only forced by aerosol and greenhouse-gas concentrations and so are only climatologically consistent with observed history. Shorter-term phenomena that can affect yield (i.e., heat waves or wildfires) are averaged out.

AutoGluon projections under RCP4.5 and RCP8.5 scenarios are shown in Figure 3. In addition to climate change scenarios, we also include two technology scenarios: ‘WOTech’ denotes agricultural technology remaining at its current level as of 2020, whereas ‘WTech’ denotes agricultural technology continuing improving through 2100. Compared with historical simulations, the future projections display higher uncertainty, with larger inter-quantile ranges, likely due to the discrepancy among climate models. It is evident by comparing the various technology scenarios within the same climatic scenario that technological advancement is necessary to maintain historical trends in annual almond yield growth. This finding highlights the important role of technological advancement in climate adaptation. On the other hand, almond yield will decline in both the RCP4.5 and RCP8.5 scenarios after peak technological advancement. As for the climate change scenarios, the RCP8.5 shows a slightly greater decline in yields compared with RCP4.5, likely due to high temperatures under this scenario.

The use of deep learning and machine learning techniques to answer climate change-related problems is growing in popularity. Not all tasks, however, can offer sufficient data samples for complex neural networks. In this study, we demonstrate how an AutoML framework can predict and project California’s almond yield. As an ensemble ML model, AutoGluon achieves better performance compared with single ML models, without requiring more data samples or computational resources. It also enables us to analyze the relative importance of input variables using, i.e., a permutation-based method, from which the total precipitation during the bloom period (February-Mid March) is shown as the most important variable. These results are not shown here, as they are unrelated to our projection analysis.

Our projection results highlight the critical role that technology plays in climate adaptation. Increases in almond yield require further technological developments; in the absence of further development, climatological factors will lead to a decrease in yields. However, the cost of developing new techniques is not taken into consideration in our analysis. From the perspective of stakeholders, it could be promising to incorporate our projection results within a socioeconomic model to gain insight into the balance between investment cost and expected output.

We would like to thank computational resources funded by the Walter and Margaret Milton Graduate Atmospheric Science Award at University of California, Davis. The workshop presentation is funded by the Lawrence Livermore National Laboratory Postdoc Development Program. Part of this work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.