Error analysis and correction of spatialization of crop yield in China – Different variables scales, partitioning schemes and error correction methods

Highlights

• We explored the influence of variables scales on precision of crop yield spatialization.

• We detected the relationship between partitioning schemes and precision of crop yield spatialization.

• We compared the pros and cons of seven different error correction methods.

Abstract

Spatialization of crop yield is beneficial to comprehensive analysis between interdisciplinary data. Multivariable linear regression models are often applied to spatialization of attribute data. The variables scales and the partitioning of China should be considered when the model is constructed. Different variables scales and partitioning schemes will inevitably results in different spatialization errors. Spatialization errors can be reduced by error correction methods. Different methods have different influence on the accuracy of crop yield spatialization. In this study, three variables scales were selected including prefectural scale, county scale and grid cell (1 km × 1 km). Five partitioning schemes (no partition of China, 7 regions of China, 9 regions of China, 10 regions of China, partitions of China by province) were considered. A total of 28 kinds of multivariable linear regression models were constructed with area of different types of farmland as independent variables, crop yields as dependent variables. Then, seven kinds of error correction methods were used to correct crop yield spatialization results. Three error evaluation indicators were selected to investigate the influence of different variables scales, partitioning schemes and error correction methods on the precision of spatialization results. The conclusions can be drawn as follows: (a) Nine models with intercept based on variables at regional scale could not be used to spatialize crop yield, while the others can be used for spatialization of crop yield. (b) The precision of the spatialization result based on the model without intercept is higher than that based on the model with intercept. (c) For models without intercept, precision of spatialization results increased first and then decreased with the refinement of partitioning scheme. (d) For models without intercept, the precision of spatialization results improved with scaling down of the variables scale from prefectural scale to county scale and grid scale. (e) Among the seven kinds of error correction methods, average correction method, weight coefficient correction methodⅡ and weight coefficient correction method III can’t be used to correct initial spatialization results. (f) Proportional coefficient correction method, weight coefficient correction methodⅠ, weight coefficient correction method Ⅳ and weight coefficient correction method Ⅴ can be used to correct initial results of spatialization. (g) The precisions of corrected spatialization products based on error correction methods, which can improve the precision of initial spatialization products, are very closely. This research made up for the deficiency of spatial error analysis of crop yield, explored the relationship between different sample scales and partitioning schemes and spatial error, compared the pros and cons of different error correction methods. Meanwhile, it also provided valuable information for other types of social and economic statistical data.

Introduction

Given a backdrop of global environmental dynamism and climate change, traditional geo-ecological processes have undergone drastic changes over the past few decades. The geographical processes are no longer simple natural processes, and the researches of ecological processes also are no longer confined to the dynamics and development in ecosystem. The integration and intersection of multiple disciplines is becoming an important characteristic of modern geo-ecological processes (Fu et al., 2006).

It is an important symbol of the combination of human activities and geo-ecological processes to apply statistics to the study of geo-ecological processes. Socio-economic statistics are collected and published based on administrative division. So they have low spatial resolution and lack of the description to spatial distribution characteristics of socioeconomic statistics. It is difficult to use them for comprehensive analysis of socio-economic data and other data in practical application, which limits their application to geographical research to a great extent. There are three major problems. First, the contradiction between the spatial heterogeneity of geographical elements and the homogeneity of statistics in the same administrative division; Second, the disagreement between landscape scale and statistical scale; Third, the statistical indicators in different regions are inconsistent (Liu and Li, 2012). The spatialization of socio-economic statistics can solve the above problems effectively (Liao and Zhang, 2009).

Numerous studies focused on the spatialization of socio-economic statistics, including spatialization of population (Tobler et al., 1995, Tobler et al., 1997, Sutton et al., 2001, Tian et al., 2005) and gross domestic product (GDP) statistics (Ebener et al., 2005, Doll et al., 2006, Sutton et al., 2007, Elvidge et al., 1997, Elvidge et al., 2009a, Elvidge et al., 2009b and Ghosh et al., 2009). With the rapid development of Remote Sensing (RS) and Geographic Information System (GIS) technology, the spatialization of agricultural production data are frequently studied, mainly including spatialization of crop acreage (Qiu et al., 2003, Leff et al., 2004, You and Wood, 2006, You et al., 2009, Monfreda et al., 2008, Khan et al., 2010, Zhang et al., 2013, Jin et al., 2015, Salmon et al., 2015, Liu et al., 2017) and agricultural production inputs (Potter et al., 2010, Sun et al., 2010, Yan and Pan, 2014). However, there are fewer researches on crop yield spatialization. For instance, Shi et al. used the cultivated land data to spatialize maize yield per unit area statistics by multivariable linear regression model, and got a spatial distribution map of maize yield per unit area in Jilin province (Shi et al., 2011). Liu et al. took population density as the dependent variables and crop yield as independent variables to construct a regression model with the support of land use data. The model was then applied to spatialize provincial-level crop yield statistics, resulting in a distribution map of crop yield of China at 1 km by 1 km in 2000 and the precision of crop yield spatialization results were analyzed from provincial scale down to prefectural scale and county scale (Liu and Li, 2012). But few studies explored the influence of variables scales and partitioning schemes on precision of crop yield spatialization.

As one of frequently-used geo-data processing methods, spatialization of attribute data inevitably results in errors during data processing. Spatialization errors can be reduced by correcting initial spatialization results. Many error modifying methods have been used to correct spatialization errors, such as average correction method (Wu et al., 2015), proportional coefficient correction method (Shi et al., 2016), weight coefficient correction method based on the basic idea that different farmland types have the same weight (Liao and Qin, 2014). However, there are few researches about comparing the pros and cons of different error correction methods. So, in this study we will discuss the influence of some new error correction methods on crop output spatialization and compare them with the existing error correction methods to improve spatialization precision.

This study attempts to simulate the spatial distribution of crop yield in China using land use data with the following objectives: (1) exploring the influence of variables scales on precision of crop yield spatialization; (2) detecting the influence of partitioning schemes on precision of crop yield spatialization; and (3) comparing the pros and cons of different error correction methods.