Quantifying uncertainty and confusion in land change analyses: a case study from central Mexico using MODIS data

Abstract

Land cover classifications of coarse-resolution data can aid the identification and quantification of natural variability and anthropogenic change at regional scales, but true landscape change can be distorted by misrepresentation of map classes. The Lerma-Chapala-Santiago (LCS) is biophysically diverse and heavily modified by urbanization and agricultural expansion. Land cover maps classified with a Mahalanobis distance algorithm and possibilistic metrics of class membership were used to quantify uncertainty (potential error in class assignment) and change confusion (potential error in land change identification). While land change analysis suggests that 33% of the landscape underwent a change in class, led by changes from or to the mosaic class (~19% of landscape), classification uncertainty values for 2001 and 2007 were 0.59 and 0.62, respectively, with highest uncertainty among bare soil classes, and an average confusion index value of 0.65, with pixels experiencing change at 0.67 and pixels experiencing persistence at 0.61 on average. These results indicate that uncertainty and potential error in land cover classifications estimates may inhibit accurate assessments of land change. Estimates of land change may be refined using these metrics to more confidently identify true landscape change and to find classes and locations that are contributing to errors in land change assessments.

Distinguishing Land Change from Natural Variability and Uncertainty in Central Mexico with MODIS EVI, TRMM Precipitation, and MODIS LST Data

“…Data for this project included vegetation index and land surface temperature products from MODIS, average daily precipitation from TRMM, a DEM from SRTM, and maps of land cover persistence and change [1]. Methods explicated below included the extraction of summary statistics from the time series of dependent and independent variables and a multiple linear regression, performed over the seven-year series and each annual subseries independently (summarized patterns of EVI, precipitation, and temperature shown in Figure 2).…”

“…Methods explicated below included the extraction of summary statistics from the time series of dependent and independent variables and a multiple linear regression, performed over the seven-year series and each annual subseries independently (summarized patterns of EVI, precipitation, and temperature shown in Figure 2). Results of the regression model were compared to a previous land cover and uncertainty analysis to explain the effect of variability upon land change error and confusion [1]. …”

“…Maps of eight natural and anthropogenic land cover categories for 2001 and 2007 were crosstabulated to identify regions of persistent and changing land cover over the seven-year period [1] ( Figure 3). Land cover classes were based on a systematic roadside survey of the Lerma-ChapalaSantiago during January and July 2007 and included the following thematic categories: forest; shrub; grass; sparse vegetation; built (including urban and rural settlements); crop (including mechanized and smallholder agriculture); and mosaic (defined by a sub-pixel mixture of cropland and natural vegetation).…”

“…Land cover classes were based on a systematic roadside survey of the Lerma-ChapalaSantiago during January and July 2007 and included the following thematic categories: forest; shrub; grass; sparse vegetation; built (including urban and rural settlements); crop (including mechanized and smallholder agriculture); and mosaic (defined by a sub-pixel mixture of cropland and natural vegetation). Land cover maps were generated for 2001 and 2007 using a combined pseudo-invariant calibration database and the Mahalanobis distance classifier in the Idrisi Taiga software package [1]. Overall map accuracy for the both maps was approximately 90%.…”

“…In the heterogeneous and rapidly changing Lerma-Chapala-Santiago watershed of Mexico [1][2][3][4], natural and anthropogenic factors, including droughts, the El Niño Southern Oscillation, agricultural expansion, and forest loss enact changes in the remotely sensed detection of vegetative vigor. This study assessed variability using a time series of Enhanced Vegetation Index (EVI) composites from the Moderate Resolution Imaging Spectoradiometer (MODIS) for 2001-2007 and per-pixel multiple linear regressions in order to model changes in EVI as a function of precipitation, temperature, and elevation.…”