STFDF与STARIMA对比：R语言时空异常分析实现及关联问询

Hey there! I've worked with both STARIMA and variogram-based kriging for spatiotemporal analysis in R, so let's walk through your questions clearly.

The go-to package for STARIMA in R is STARMA—it’s purpose-built for spatiotemporal autoregressive moving average models. Here’s a step-by-step workflow that ties into your existing STFDF data:

1. Load required packages
   First, make sure you have the necessary libraries installed and loaded:

   install.packages("STARMA") # Only run this once
   library(spacetime)
   library(STARMA)
   

2. Convert your STFDF to STARMA-compatible format
   STARMA works with a spatiotemporal matrix where rows represent time points and columns represent spatial units. You can extract this from your STFDF easily:

   # Assume your STFDF is named `stfdf_data`, and your target variable is `value`
   # Get spatial unit identifiers and time points
   spatial_ids <- stfdf_data@sp@data$ID # Replace with your spatial ID column name
   time_steps <- stfdf_data@time
   
   # Build the spatiotemporal matrix
   st_matrix <- t(sapply(stfdf_data@data, function(x) x$value))
   rownames(st_matrix) <- as.character(time_steps)
   colnames(st_matrix) <- spatial_ids
   

3. Fit the STARIMA model
   Use the starima() function to specify your model. The key parameters are:

   • p: Spatial autoregressive order
   • P: Temporal autoregressive order
   • q: Spatial moving average order
   • Q: Temporal moving average order

   Here’s an example fitting a STARIMA(1,1,1,1) model:

   starima_fit <- starima(st_matrix, p=1, P=1, q=1, Q=1)
   summary(starima_fit) # Inspect model diagnostics
   

4. Generate predictions
   Once your model is fitted, you can predict future spatiotemporal values with:

   # Predict the next 3 time steps for all spatial units
   starima_preds <- predict(starima_fit, n.ahead=3)
   print(starima_preds)
   

   If you need more flexibility, you can also combine spacetime with the forecast package for custom time series extensions, but STARMA is the most straightforward tool for pure STARIMA implementations.

Both methods tackle spatiotemporal correlation, but they approach the problem from different angles—here’s how they connect and diverge:

Core Connections #

• Both model spatiotemporal dependence: Neither assumes your data is independent. They both account for how values at nearby locations and adjacent time points are correlated, just using different mathematical frameworks.
• Mutually informative workflows: You can use variogram analysis (from your kriging work) to guide STARIMA model selection. For example, the range of your spatial variogram can hint at the appropriate p/q order, while temporal autocorrelation plots can inform P/Q choices.

Key Differences #

• Modeling philosophy:
  • Variogram-based kriging (like stkrige) is a geostatistical method. It focuses on describing spatial (and spatiotemporal) heterogeneity via variograms, then interpolating missing values based on that spatial structure. It’s great for filling gaps or predicting unobserved spatial locations at a given time.
  • STARIMA is a time series extension. It builds on ARIMA by adding spatial autoregressive/moving average terms, making it ideal for forecasting how values across all spatial units evolve over time.
• Data requirements:
  • Kriging handles irregular spatiotemporal sampling (e.g., different locations sampled at different times) well.
  • STARIMA requires balanced spatiotemporal panels—every spatial unit must have observations at every time point.
• Use case focus:
  • Use kriging if your priority is spatial interpolation (e.g., mapping a variable across an area at a single time point).
  • Use STARIMA if you need dynamic spatiotemporal forecasting (e.g., predicting temperature across multiple cities for the next week).

内容的提问来源于stack exchange，提问作者Reta