Model Evaluation and Comparison

Mamadou SOW

2026-09-26

Overview

This vignette demonstrates how to evaluate model performance using spatial cross-validation and compare different validation methods to understand the impact of spatial dependence.

Model Evaluation Metrics

Available Metrics

spatialcvR provides standard model evaluation metrics:

Calculating Metrics

library(spatialcvR)

# Example predictions vs observations
observed <- c(10, 15, 20, 25, 30, 35, 40, 45, 50, 55)
predicted <- c(11, 14, 21, 24, 31, 34, 41, 44, 51, 54)

# Calculate metrics
metrics <- spatial_metrics(observed, predicted)

print(metrics)
## Model Evaluation Metrics
## =======================
## Number of observations: 10 
## RMSE: 1 
## MAE: 1 
## R2: 0.9952 
## MAPE: 4.04 %

Understanding the Metrics

# Perfect prediction
observed_perfect <- c(10, 20, 30, 40, 50)
predicted_perfect <- c(10, 20, 30, 40, 50)
metrics_perfect <- spatial_metrics(observed_perfect, predicted_perfect)
print(metrics_perfect)
## Model Evaluation Metrics
## =======================
## Number of observations: 5 
## RMSE: 0 
## MAE: 0 
## R2: 1 
## MAPE: 0 %
# Poor prediction
observed_poor <- c(10, 20, 30, 40, 50)
predicted_poor <- c(50, 40, 30, 20, 10)
metrics_poor <- spatial_metrics(observed_poor, predicted_poor)
print(metrics_poor)
## Model Evaluation Metrics
## =======================
## Number of observations: 5 
## RMSE: 28.2843 
## MAE: 24 
## R2: -3 
## MAPE: 126 %

Spatial Cross-Validation Workflow

Step 1: Create Folds

# Load sample data
data(sample_spatial_data)

# Create spatial folds
folds_block <- spatial_folds(
  data = sample_spatial_data,
  x = "longitude",
  y = "latitude",
  k = 5,
  method = "block",
  seed = 123
)

# Create random folds for comparison
folds_random <- spatial_folds(
  data = sample_spatial_data,
  x = "longitude",
  y = "latitude",
  k = 5,
  method = "random",
  seed = 123
)

Step 2: Train and Evaluate Models

For this example, we’ll simulate model training and evaluation:

# Simulate model training and evaluation for each fold
evaluate_fold <- function(fold, data, target_var) {
  train_idx <- fold$train
  test_idx <- fold$test
  
  # Simple linear model for demonstration
  train_data <- data[train_idx, ]
  test_data <- data[test_idx, ]
  
  # Train model (using variable1 and variable2 as predictors)
  model <- lm(target ~ variable1 + variable2, data = train_data)
  
  # Make predictions
  predictions <- predict(model, newdata = test_data)
  observed <- test_data[[target_var]]
  
  # Calculate metrics
  metrics <- spatial_metrics(observed, predictions, na.rm = TRUE)
  
  return(metrics)
}

# Evaluate spatial block CV
block_results <- lapply(folds_block$folds, function(fold) {
  evaluate_fold(fold, sample_spatial_data, "target")
})

# Evaluate random CV
random_results <- lapply(folds_random$folds, function(fold) {
  evaluate_fold(fold, sample_spatial_data, "target")
})

Step 3: Aggregate Results

# Aggregate metrics across folds
aggregate_metrics <- function(results_list) {
  rmse_values <- sapply(results_list, function(x) x$RMSE)
  mae_values <- sapply(results_list, function(x) x$MAE)
  r2_values <- sapply(results_list, function(x) x$R2)
  
  list(
    RMSE_mean = mean(rmse_values),
    RMSE_sd = sd(rmse_values),
    MAE_mean = mean(mae_values),
    MAE_sd = sd(mae_values),
    R2_mean = mean(r2_values),
    R2_sd = sd(r2_values)
  )
}

block_summary <- aggregate_metrics(block_results)
random_summary <- aggregate_metrics(random_results)

cat("Spatial Block CV Summary:\n")
## Spatial Block CV Summary:
print(block_summary)
## $RMSE_mean
## [1] 6.712582
## 
## $RMSE_sd
## [1] 1.223729
## 
## $MAE_mean
## [1] 5.477149
## 
## $MAE_sd
## [1] 0.9768991
## 
## $R2_mean
## [1] 0.7782187
## 
## $R2_sd
## [1] 0.1693295
cat("\nRandom CV Summary:\n")
## 
## Random CV Summary:
print(random_summary)
## $RMSE_mean
## [1] 6.83486
## 
## $RMSE_sd
## [1] 0.5242892
## 
## $MAE_mean
## [1] 5.522551
## 
## $MAE_sd
## [1] 0.2974677
## 
## $R2_mean
## [1] 0.8758788
## 
## $R2_sd
## [1] 0.01977151

Comparing CV Methods

Using compare_cv

# Create summary metrics for comparison
block_summary_metrics <- list(
  RMSE = block_summary$RMSE_mean,
  MAE = block_summary$MAE_mean,
  R2 = block_summary$R2_mean
)

random_summary_metrics <- list(
  RMSE = random_summary$RMSE_mean,
  MAE = random_summary$MAE_mean,
  R2 = random_summary$R2_mean
)

# Compare methods
comparison <- compare_cv(
  results_list = list(
    spatial_block = block_summary_metrics,
    random = random_summary_metrics
  )
)

print(comparison)
## Cross-Validation Method Comparison
## ===================================
## Number of methods compared: 2 
## Best method (lowest RMSE): spatial_block 
## 
## Summary Table:
##         method     RMSE      MAE        R2 RMSE_diff RMSE_diff_pct
##  spatial_block 6.712582 5.477149 0.7782187 0.0000000      0.000000
##         random 6.834860 5.522551 0.8758788 0.1222781      1.821625
## 
## Performance Comparison:
##   Baseline method: spatial_block (RMSE = 6.7126)
##   Best RMSE: 6.7126
##   RMSE range: [6.7126, 6.8349]
##   MAE range: [5.4771, 5.5226]
##   R2 range: [0.7782, 0.8759]

Interpreting the Comparison

Spatial Residual Diagnostics

Calculating Spatial Residuals

# Get predictions from one fold for residual analysis
fold_idx <- 1
train_idx <- folds_block$folds[[fold_idx]]$train
test_idx <- folds_block$folds[[fold_idx]]$test

train_data <- sample_spatial_data[train_idx, ]
test_data <- sample_spatial_data[test_idx, ]

# Train model
model <- lm(target ~ variable1 + variable2, data = train_data)

# Make predictions
predictions <- predict(model, newdata = test_data)
observed <- test_data$target

# Remove NA values
valid_idx <- !is.na(predictions) & !is.na(observed)
predictions <- predictions[valid_idx]
observed <- observed[valid_idx]
coords <- cbind(
  x = test_data$longitude[valid_idx],
  y = test_data$latitude[valid_idx]
)

# Calculate spatial residuals
residuals <- spatial_residuals(observed, predictions, coords)

print(residuals)

Visualizing Residuals

# Spatial scatter plot of residuals
plot_spatial_residuals(residuals, type = "scatter",
                      main = "Spatial Residuals - Scatter Plot")

# Histogram of residuals
# (Chunk disabled due to dependency on previous disabled chunk)
# plot_spatial_residuals(residuals, type = "histogram",
#                       main = "Spatial Residuals - Histogram")

# Q-Q plot for normality
# (Chunk disabled due to dependency on previous disabled chunk)
# plot_spatial_residuals(residuals, type = "qq",
#                       main = "Spatial Residuals - Q-Q Plot")

Interpreting Residual Patterns

Practical Example: Complete Workflow

Scenario: Predicting Crop Yield

# Simulate agricultural data
set.seed(456)
n <- 100
agri_data <- data.frame(
  longitude = runif(n, 0, 1000),
  latitude = runif(n, 0, 1000),
  soil_quality = rnorm(n, 50, 10),
  rainfall = rnorm(n, 800, 100),
  temperature = rnorm(n, 20, 3),
  yield = 50 + 0.3*soil_quality + 0.02*rainfall + 0.5*temperature + 
           rnorm(n, 0, 5)
)

# Add spatial structure to yield
agri_data$yield <- agri_data$yield + 
  0.01*agri_data$longitude + 0.02*agri_data$latitude

# Create spatial folds
agri_folds <- spatial_folds(
  data = agri_data,
  x = "longitude",
  y = "latitude",
  k = 5,
  method = "block",
  seed = 456
)

# Check for spatial leakage
agri_leakage <- detect_spatial_leakage(
  data = agri_data,
  folds = agri_folds,
  x = "longitude",
  y = "latitude"
)

print(agri_leakage)

# Evaluate model
agri_results <- lapply(agri_folds$folds, function(fold) {
  evaluate_fold(fold, agri_data, "yield")
})

agri_summary <- aggregate_metrics(agri_results)
print(agri_summary)

Advanced Topics

Cross-Method Comparison

# Compare multiple spatial CV methods
methods <- c("block", "cluster", "random")
method_results <- list()

for (method in methods) {
  folds <- spatial_folds(
    data = sample_spatial_data,
    x = "longitude",
    y = "latitude",
    k = 5,
    method = method,
    seed = 123
  )
  
  results <- lapply(folds$folds, function(fold) {
    evaluate_fold(fold, sample_spatial_data, "target")
  })
  
  summary <- aggregate_metrics(results)
  method_results[[method]] <- list(
    RMSE = summary$RMSE_mean,
    MAE = summary$MAE_mean,
    R2 = summary$R2_mean
  )
}

# Compare all methods
multi_comparison <- compare_cv(method_results)
print(multi_comparison)

Performance vs. Spatial Separation Trade-off

# Test different levels of spatial separation
k_values <- c(3, 5, 7, 10)
separation_results <- data.frame(
  k = integer(),
  mean_distance = numeric(),
  RMSE = numeric()
  # (Chunks disabled for CRAN build - they work interactively but not during build)
  # For demonstration, please see the other vignettes or run the examples interactively
)

Best Practices

1. Always Compare with Random CV

# Establish baseline
random_baseline <- spatial_folds(sample_spatial_data, "longitude", "latitude",
                                k = 5, method = "random", seed = 123)

# Compare with spatial methods
spatial_methods <- spatial_folds(sample_spatial_data, "longitude", "latitude",
                                k = 5, method = "block", seed = 123)

2. Check for Spatial Leakage

# Always check leakage before trusting results
leakage_check <- detect_spatial_leakage(sample_spatial_data, spatial_methods,
                                       "longitude", "latitude")
if (leakage_check$risk_level == "high") {
  warning("High spatial leakage detected - results may be unreliable")
}

3. Analyze Residuals Spatially

# Don't just look at aggregate metrics
# Examine spatial patterns in errors
# This can reveal missing spatial predictors

4. Consider the Prediction Context

5. Document Your Approach

# Keep track of:
# - CV method used
# - Parameters (k, block size, buffer radius, etc.)
# - Spatial leakage assessment
# - Performance metrics
# - Residual analysis results

Common Pitfalls

Pitfall 1: Ignoring Spatial Dependence

Problem: Using random CV on spatial data without checking for spatial dependence.

Solution: Always compare spatial vs. random CV to assess the impact.

Pitfall 2: Overfitting to Spatial Structure

Problem: Using too much spatial information in predictors.

Solution: Keep spatial coordinates out of predictor variables unless explicitly modeling spatial effects.

Pitfall 3: Inappropriate CRS

Problem: Calculating distances on geographic coordinates.

Solution: Use projected CRS for distance-based methods, or acknowledge approximation.

Pitfall 4: Small Sample Sizes

Problem: Too few observations relative to spatial complexity.

Solution: Reduce k, use larger blocks, or collect more data.

Next Steps

Key Takeaways

  1. Use spatial CV for geospatial data to get realistic performance estimates
  2. Compare methods to understand the impact of spatial dependence
  3. Check for leakage to ensure adequate spatial separation
  4. Analyze residuals spatially to identify missing patterns
  5. Consider trade-offs between performance and spatial generalization