Back to publications

Journal Article

Identifying climate and environmental determinants of spatial disparities in wheat production using a geospatial machine learning model

AuthorsKai Ren, Yongze Song*, Linchao Li, Francesco Mancini, Zhuoyao Xiao, Xueyuan Zhang, Rui Qu, Qiang Yu*

JournalGIScience & Remote Sensing,

Keywords Agricultural remote sensing Spatial heterogeneity Spatial statistics Geographical detector Geospatial intelligence
https://doi.org/10.1080/15481603.2025.2533487

Overview

This paper builds a geospatial machine learning workflow that combines hotspot detection, geographically optimal zones-based heterogeneity (GOZH), and determinant interaction analysis to explain why wheat production varies across Australia and how the dominant controls changed between 2016 and 2021.

Abstract

Wheat production is crucial in global food security and sustainable development, especially under severe global climate change, frequent extreme weather events, and significant population growth worldwide. A deeper understanding of spatial variation in wheat production and its determining factors is essential for implementing different cultivation practices, water and fertilizer management, and adaptive variety selection across different regions. However, existing methods primarily focus on identifying single-variable factors while lacking geographical spatial characteristics, which may lead to an incomplete exploration of spatial disparities in wheat production, predictions, and responses to changes in determining factors. This study develops a geospatial machine learning model by integrating spatial autocorrelation, spatial stratified heterogeneity, and decision tree learning to identify spatial disparities of wheat production and their determinants. The model is applied to wheat production analysis in Australia, the world's fifth wheat-producing country in 2022. First, a spatial autocorrelation method is employed to identify hotspot areas of wheat production in Australia. Next, the geographically optimal zones-based heterogeneity (GOZH) model, which integrates spatial stratified heterogeneity and decision tree learning, is used to identify determinants and their interactions that drive spatial disparities in wheat production. Finally, the developed geospatial machine learning model is evaluated by comparing its effectiveness with the commonly used geographical detector model. The results demonstrate pronounced spatial heterogeneity in Australian wheat production driven by environmental, climatic, and soil factors and their interactions. Identifying these spatial determinants enables more efficient crop management - such as targeted sub-regional practices, climate-adaptive variety selection, and soil health strategies - thereby supporting food security and sustainable agricultural systems.

Method Implementation

The implementation combines spatial autocorrelation diagnostics, geographically optimal zones-based heterogeneity (GOZH), and a geographical detector comparison to explain spatial disparities in Australian wheat production.

  1. Preprocess wheat production and explanatory variables. Wheat production and 20 explanatory variables are aligned to Local Government Areas (LGAs). Climate, soil, geographic, and environmental variables are standardized by LGA means; temperature is converted from Kelvin to Celsius; soil outliers are removed using a 2.5 standard deviation threshold; and PCA is used to reduce layered soil attributes.
  2. Identify spatial clustering in the response variable. Global Moran's I is used to test whether wheat production is spatially clustered, while LISA/local Moran's I identifies hotspot, cold spot, and non-significant regions:
    \[ I = \frac{ n\sum_{i=1}^{n}\sum_{j=1}^{n}w_{ij}(x_i-\bar{x})(x_j-\bar{x}) }{ \sum_{i=1}^{n}\sum_{j=1}^{n}w_{ij}\sum_{i=1}^{n}(x_i-\bar{x})^2 } \]
    \[ I_i = \frac{x_i-\bar{x}}{s^2} \sum_{j=1}^{n}w_{ij}(x_j-\bar{x}) \]
  3. Measure the explanatory power of individual determinants. The geographical detector framework evaluates how strongly each stratified explanatory variable explains wheat production using the power of determinants (PD):
    \[ PD = 1 - \frac{\sum_{z=1}^{h}N_z\sigma_z^2}{N\sigma^2} = 1 - \frac{SSW}{SST} \]
    A larger PD indicates stronger explanatory power for the spatial heterogeneity of wheat production.
  4. Delineate geographically optimal zones with GOZH. GOZH uses decision-tree-based spatial discretization to search for zones that maximize between-zone differences while minimizing within-zone variation. Its optimal power of determinants is written as:
    \[ \Omega = \max(PD) = \gamma(X,D) = 1 - \frac{\min(SSW_{X,D})}{SST} \]
    \[ \min(SSW_{X,D}) = \min\left\{ \sum_{z=1}^{h}\sum_{j=1}^{N_z}(y_{zj}-\bar{c}_z)^2 \right\} \]
    In the wheat case study, LGAs are recursively split with the 20 explanatory variables, using a minimum node size of 10 LGAs to avoid excessive partitioning.
  5. Evaluate interaction effects at multiple spatial scales. The GOZH model is run for Australia as a whole, hotspot and non-hotspot regions, and individual states. The resulting \(\Omega\) values quantify the combined influence of geographic, climatic, soil, and environmental determinants under different spatial contexts.
  6. Validate the model against OPGD. The same response and explanatory variables are analyzed with the optimal parameters-based geographical detector (OPGD). GOZH and OPGD are compared through individual-variable PD values, principal-variable stratification curves, and overall PD values to assess whether GOZH better explains wheat-production disparities.

Figures and Tables

Figure 1. Distribution of Australian total wheat production and mean wheat production in 2016 and 2021.
Figure 1. Distribution of Australian total wheat production and mean wheat production in 2016 and 2021. 图 1. 2016 年和 2021 年澳大利亚小麦总产量与平均小麦产量分布。
Table 1. The wheat growing seasons in various states of Australia.
Table 1. The wheat growing seasons in various states of Australia. 表 1. 澳大利亚各州的小麦生长季。
Figure 2. Spatial distributions of explanatory variables in 2021. (a)-(c) geographical variable, (d)-(g) climate variables, (h)-(p) soil attributes, (q)-(t) environmental variables.
Figure 2. Spatial distributions of explanatory variables in 2021. (a)-(c) geographical variable, (d)-(g) climate variables, (h)-(p) soil attributes, (q)-(t) environmental variables. 图 2. 2021 年解释变量的空间分布。(a)-(c) 为地理变量，(d)-(g) 为气候变量，(h)-(p) 为土壤属性，(q)-(t) 为环境变量。
Table 2. A summary of explanatory variables that potentially affect spatial disparities of wheat production.
Table 2. A summary of explanatory variables that potentially affect spatial disparities of wheat production. 表 2. 可能影响小麦产量空间差异的解释变量汇总。
Figure 3. Schematic overview of identifying the spatial disparities and determinants of wheat production in Australia based on the geospatial machine learning model.
Figure 3. Schematic overview of identifying the spatial disparities and determinants of wheat production in Australia based on the geospatial machine learning model. 图 3. 基于地理空间机器学习模型识别澳大利亚小麦产量空间差异及其决定因素的示意图。
Table 3. Methods and procedures for pre-processing explanatory variables.
Table 3. Methods and procedures for pre-processing explanatory variables. 表 3. 解释变量预处理的方法与流程。
Figure 4. Results of Moran's I and LISA analysis. Value of Moran's I in 2016 (a) and 2021 (d), hotspot analysis based on LISA in 2016 (b) and 2021 (e), significance analysis based on LISA in 2016 (c) and 2021 (f).
Figure 4. Results of Moran's I and LISA analysis. Value of Moran's I in 2016 (a) and 2021 (d), hotspot analysis based on LISA in 2016 (b) and 2021 (e), significance analysis based on LISA in 2016 (c) and 2021 (f). 图 4. Moran's I 和 LISA 分析结果。2016 年（a）和 2021 年（d）的 Moran's I 值，2016 年（b）和 2021 年（e）基于 LISA 的热点分析，以及 2016 年（c）和 2021 年（f）基于 LISA 的显著性分析。
Figure 5. Mean wheat production of hotspot and non-hotspot areas based on the results of LISA analysis.
Figure 5. Mean wheat production of hotspot and non-hotspot areas based on the results of LISA analysis. 图 5. 基于 LISA 分析结果的热点区与非热点区平均小麦产量。
Table 4. Statistical details on the number of LGAs, total average production, and average production characteristics across hotspots, cold spots, and other regions.
Table 4. Statistical details on the number of LGAs, total average production, and average production characteristics across hotspots, cold spots, and other regions. 表 4. 热点区、冷点区及其他区域在 LGA 数量、总平均产量和平均产量特征方面的统计信息。
Figure 6. Results of the geographical optimal zones-based heterogeneity (GOZH) model for assessing power of determinants (PD) of wheat production.
Figure 6. Results of the geographical optimal zones-based heterogeneity (GOZH) model for assessing power of determinants (PD) of wheat production. 图 6. 用于评估小麦产量决定因素解释力（PD）的地理最优分区异质性（GOZH）模型结果。
Figure 7. Changes in the top three determinants affecting spatial disparities in wheat production from 2016 to 2021, as well as their change rates and coefficients of variation.
Figure 7. Changes in the top three determinants affecting spatial disparities in wheat production from 2016 to 2021, as well as their change rates and coefficients of variation. 图 7. 2016 年至 2021 年影响小麦产量空间差异的前三位决定因素及其变化率和变异系数变化。
Figure 8. Decision tree of identifying optimal zones (a, d), geographically optimal zones of wheat production at the LGAs identified using the GOZH model (b, e), and statistical summaries of explanatory variables within zones for explaining characteristics of zones (c, f).
Figure 8. Decision tree of identifying optimal zones (a, d), geographically optimal zones of wheat production at the LGAs identified using the GOZH model (b, e), and statistical summaries of explanatory variables within zones for explaining characteristics of zones (c, f). 图 8. 最优分区识别决策树（a, d）、利用 GOZH 模型识别的 LGA 尺度小麦产量地理最优分区（b, e），以及用于解释各分区特征的分区内解释变量统计汇总（c, f）。
Figure 9. Contributions of explanatory variables for each pair of geographically optimal zones of mean wheat production in 2016 (a) and 2021 (b).
Figure 9. Contributions of explanatory variables for each pair of geographically optimal zones of mean wheat production in 2016 (a) and 2021 (b). 图 9. 2016 年（a）和 2021 年（b）平均小麦产量各对地理最优分区之间解释变量的贡献。
Table 5. Contributions of explanatory variables to dividing optimal zones.
Table 5. Contributions of explanatory variables to dividing optimal zones. 表 5. 解释变量对最优分区划分的贡献。
Figure 10. Omega values in Australia, hotspot and non-hotspot regions, and across individual states using the GOZH model.
Figure 10. Ω values in Australia, hotspot and non-hotspot regions, and across individual states using the GOZH model. 图 10. 基于 GOZH 模型得到的澳大利亚整体、热点区与非热点区以及各州的 Ω 值。
Figure 11. PD values of individual explanatory variables in the OPGD model in 2016 (a) and 2021 (c), differences in individual variable PD values between the OPGD and GOZH models in 2016 (b) and 2021 (d).
Figure 11. PD values of individual explanatory variables in the OPGD model in 2016 (a) and 2021 (c), differences in individual variable PD values between the OPGD and GOZH models in 2016 (b) and 2021 (d). 图 11. 2016 年（a）和 2021 年（c）OPGD 模型中各解释变量的 PD 值，以及 2016 年（b）和 2021 年（d）OPGD 与 GOZH 模型之间各变量 PD 值的差异。
Figure 12. Variation of PD values for the four principal variables across different strata in the GOZH and OPGD models in 2016 ((a) AWC, (b) Nto, (c) ASP, (d) PTO) and 2021 ((e) EVI, (f) TP, (g) SLT, (h) NTO).
Figure 12. Variation of PD values for the four principal variables across different strata in the GOZH and OPGD models in 2016 ((a) AWC, (b) Nto, (c) ASP, (d) PTO) and 2021 ((e) EVI, (f) TP, (g) SLT, (h) NTO). 图 12. 2016 年与 2021 年 GOZH 和 OPGD 模型中四个主要变量在不同分层下的 PD 值变化：2016 年为 (a) AWC、(b) Nto、(c) ASP、(d) PTO，2021 年为 (e) EVI、(f) TP、(g) SLT、(h) NTO。

Data Availability Statement

Data and codes supporting the findings of this study are available on Figshare.

Funding

This research was funded by the National Key Research and Development Program of China (2023YFB3907403) and supported by the China Scholarship Council (Grant No. 202206300058).