Geographically Optimal Similarity (GOS) and the Third Law of Geography in R

2.1 Characterizing geographical configurations

The geosimilarity package contains two spatial datasets:

• zn: Spatial samples of Zn concentrations and explanatory variables at sample locations

• grid: Spatial grid data of explanatory variables used for the prediction

install.packages("geosimilarity")

library("geosimilarity")

## This is `geosimilarity` 2.2.
##                         
## To cite `geosimilarity` in publications, please use:
##                         
## Song, Y. (2022). Geographically Optimal Similarity. Mathematical Geosciences. doi: 10.1007/s11004-022-10036-8.
## 

data("zn")
head(zn)

## # A tibble: 6 × 12
##     Lon   Lat    Zn Elevat…¹ Slope Aspect  Water  NDVI   SOC    pH    Road  Mine
##   <dbl> <dbl> <dbl>    <dbl> <dbl>  <dbl>  <dbl> <dbl> <dbl> <dbl>   <dbl> <dbl>
## 1  120. -28.5    10     455. 0.236  306.  0.014  0.184 0.909  5.95 49.4     55.6
## 2  120. -28.4    30     451. 0.207  293.  2.20   0.202 0.906  6.05 49.0     51.1
## 3  120. -28.4    30     443. 0.285  325.  0.0119 0.163 0.848  5.76 45.1     45.0
## 4  120. -27.4    30     509. 0.236   98.4 3.06   0.204 0.851  5.82  0.0774  49.0
## 5  120. -28.3    33     427. 0.191  329.  3.53   0.179 0.933  5.85 39.9     39.8
## 6  120. -27.3    27     510. 0.211  105.  3.38   0.191 0.868  6.07  0.0468  48.7
## # … with abbreviated variable name ¹​Elevation

Data pre-processing and variable selection:

# log-transformation
hist(zn$Zn)
zn$Zn <- log(zn$Zn)
hist(zn$Zn)

# remove outliers
library(SecDim)
k <- rmvoutlier(zn$Zn, coef = 2.5)
dt <- zn[-k,]

# correlation  
library("PerformanceAnalytics")
cor_dt <- dt[, c(3:12)]
chart.Correlation(cor_dt, histogram = TRUE, pch = 19)

# multicollinearity
library(car)
m1 <- lm(Zn ~ Slope + Water + NDVI + SOC + pH + Road + Mine, data = dt)
car::vif(m1)

In this step, the selected variables include Slope, Water, NDVI, SOC, pH, Road, and Mine.

2.2 Determining the optimal similarity

In the bestkappa function, if you set more optional numbers to the kappa vector and a higher value of the cross-validation repeat times nrepeat, a $\kappa$ value enabling more accurate prediction will be selected, but the computation time will be increased.

For instance, if the optional kappa is (0.01, 0.05, 0.1, 0.2, 0.5, 1) and nrepeat is 2, the $\lambda$ (the optimal $\kappa$) is 0.10, and its corresponding mean cross-validation RMSE is 0.6591. The computation time is 5.9s.

system.time({ # 5.9s
b1 <- bestkappa(Zn ~ Slope + Water + NDVI  + SOC + pH + Road + Mine,
                data = dt,
                kappa = c(0.01, 0.05, 0.1, 0.2, 0.5, 1),
                nrepeat = 2)
})
b1$bestkappa
b1$cvmean
b1$plot

If the optional kappa is (0.01, 0.02, …, 0.09, 0.1, 0.2, …, 1) and nrepeat is 10, the $\lambda$ (the optimal $\kappa$) is 0.08, and its corresponding mean cross-validation RMSE is 0.6668, which are more stable and reliable. The computation time is 97.7s.

system.time({ # 97.7ss
b2 <- bestkappa(Zn ~ Slope + Water + NDVI  + SOC + pH + Road + Mine,
                data = dt,
                kappa = c(seq(0.01, 0.1, 0.01), seq(0.2, 1, 0.1)),
                nrepeat = 10)
})
b2$bestkappa
b2$cvmean
b2$plot

Figure 1. Processes of determining the optimal similarity. (a) The optional kappa is (0.01, 0.05, 0.1, 0.2, 0.5, 1) and nrepeat is 2. (b) The optional kappa is (0.01, 0.02, …, 0.09, 0.1, 0.2, …, 1) and nrepeat is 10.

2.3 Spatial prediction

system.time({ # 26s
g2 <- gos(Zn ~ Slope + Water + NDVI  + SOC + pH + Road + Mine,
           data = dt, newdata = grid, kappa = 0.08)
})
grid$pred <- exp(g2$pred)
grid$uc99 <- g2$uncertainty$`0.99`
library(ggplot2)
library(viridis)
ggplot(grid, aes(x = Lon, y = Lat, fill = pred)) +
  geom_tile() +
  scale_fill_viridis(option="magma", direction = -1) + 
  coord_equal() +
  labs(fill='Prediction') +
  theme_bw() 
ggplot(grid, aes(x = Lon, y = Lat, fill = uc99)) +
  geom_tile() +
  scale_fill_viridis(option="mako", direction = -1) + 
  coord_equal() +
  labs(fill=bquote(Uncertainty~(zeta==0.99))) +
  theme_bw() 

Figure 2. Geographially optimal similarity (GOS)-based prediction (a) and uncertainty (b).

In addition, the following codes can be used to plot uncertainty under different $\zeta$ values.

library(reshape2)
uc <- cbind(grid[,2:3], g2$uncertainty)
uc <- melt(uc, id = c("Lon", "Lat"))
ggplot(uc, aes(x = Lon, y = Lat, fill = value)) +
  geom_tile() +
  scale_fill_viridis(option="mako", direction = -1) + 
  coord_equal() +
  facet_wrap(~ variable) +
  labs(fill='Uncertainty') +
  theme_bw() 

Figure 3. Geographially optimal similarity (GOS)-based spatial prediction uncertainties under different $\zeta$ values.

2.4 Model evaluation

We can compare model accuracy of GOS with various models, such as kriging, multivariate regression, regression kriging, random forest, BCS, etc., as shown in Song (2022). Here is a simple example of comparing modeling accuracy between BCS and GOS.

set.seed(99)
# split data for validation: 50% training; 50% testing
split <- sample(1:nrow(dt), round(nrow(dt)*0.5))
train <- dt[split,]
test <- dt[-split,]

library(DescTools)
# BCS
h1 <- gos(Zn ~ Slope + Water + NDVI  + SOC + pH + Road + Mine,
          data = train, newdata = test, kappa = 1)
MAE(test$Zn, h1$pred)
RMSE(test$Zn, h1$pred)

# GOS
h2 <- gos(Zn ~ Slope + Water + NDVI  + SOC + pH + Road + Mine,
          data = train, newdata = test, kappa = 0.08)
MAE(test$Zn, h2$pred)
RMSE(test$Zn, h2$pred)

As a result, the MAE of BCS is 0.5158 and the MAE of GOS is 0.5089, the RMSE of BCS is 0.6599 and the RMSE of GOS is 0.6523. Compared with BCS, GOS reduced 1.34% of MAE and 1.15% of RMSE.