GlobalWheatYield4km: a global wheat yield dataset at 4-km resolution during 1982–2020 based on deep learning approaches

Abstract

Abstract. Accurate and spatially explicit information on crop yield over large areas is paramount for ensuring global food security and guiding policy-making. However, most public datasets are coarse resolution in both space and time. Here, we used data-driven models to develop a 4-km dataset of global wheat yield (GlobalWheatYield4km) from 1982 to 2020. First, we proposed a phenology-based approach to map spatial distribution. Then we determined the optimal grid-scale yield estimation model by comparing the performance of two data-driven models (i.e., Random Forest (RF) and Long Short-Term Memory (LSTM)), with publicly available data (i.e., satellite and climatic data from the Google Earth Engine (GEE) platform, soil properties, and subnational statistics covering ~11000 political units). The results showed that GlobalWheatYield4km captured 82 % of yield variations with RMSE of 619.8 kg/ha. In addition, our dataset had a higher accuracy (R2 ~0.73) as compared with Spatial Production Allocation Model (R2 ~ 0.49) across all regions and years. The GlobalWheatYield4km dataset will play important roles in modelling crop system and assessing climate impact over larger areas ((DOI of the referenced dataset: https://doi.org/10.6084/m9.figshare.10025006; Luo et al., 2022b).

ChinaWheatYield30m: a 30 m annual winter wheat yield dataset from 2016 to 2021 in China

“…Random forest (RF) is a model with predictive performance commonly used in the current yield estimation literature (Li et al, 2020;Cheng et al, 2022;Luo et al, 2022). RF regression is a classic ensemble machine learning model that establishes multiple unrelated decision trees by randomly extracting samples and features and obtains the prediction results in parallel.…”

“…The number of decision trees was set to 200, and the maximum depth of the tree and the number of features optimized the models' hyperparameters through a pretuned procedure (Li et al, 2021;Cheng et al, 2022). We compared our yield production (Chi-naWheatYield30m) with an existing 4 km dataset of global wheat yield (GlobalWheatYield4km) (Luo et al, 2022) using in situ data to validate the reliability of our dataset. More specifically, we calculated the correlation coefficient (r) and relative root mean square error (rRMSE) between the in situ measurement yields and the estimates of GlobalWheatYield4km or ChinaWheatYield30m from 2016 to 2021.…”

“…To our knowledge, to date there is no fine resolution (30 m) winter wheat yield distribution map. Previous research has generated the winter wheat yield distribution map of some major production areas in China at moderate resolution, e.g., 10 km, 5 arcmin grid, 5 min grid, 4 km and 1 km (Monfreda et al, 2008;Grogan et al, 2022;Luo et al, 2022;Cheng et al, 2022). Moderate-resolution yield maps have a mixed-pixel problem, which may lead to great uncertainties, as mentioned in comparison with the 4 km yield dataset.…”

“…Generally, a semi-mechanistic model has great potential in yield estimation, but its application is often limited to a relatively small area, e.g., farm, county or city scale, rather than a larger scale. National crop yield datasets, which are of great significance for large-scale agricultural resource allocation, agricultural system model construction and climate change impact assessment, are produced at coarse spatial resolutions (Table 1), e.g., 0.5 • , 10 km, 4 km or 1 km resolution (Monfreda et al, 2008;You et al, 2014;Iizumi and Sakai, 2020;Grogan et al, 2022;Luo et al, 2022;Cheng et al, 2022), and are mostly downscaled based on the statistical yield datasets and other datasets (Monfreda et al, 2008;You et al, 2014;Iizumi and Sakai, 2020;Grogan et al, 2022). This method of yield downscaling may lead to inaccurate yield estimates and incorrect assessments of the impact of climate change.…”

“…In addition, yield predictions cannot rely on statistical data alone. Luo et al (2022) and Cheng et al (2022) developed yield datasets combining coarse-resolution real-time remote sensing data with agricultural statistics, but because 1 km × 1 km plots or 4 km × 4 km farmlands are rare in China, their field application is limited. Although these datasets are useful for analyzing larger-scale temporal and spatial changes in yield, they cannot deal with small-scale spatial heterogeneity, which happens to be the most significant characteristic of the Chinese farmers' economy.…”