如何构建以多期地理参考栅格为目标变量的预测训练数据结构

Great question—handling multi-date raster targets is super common in geospatial machine learning, and while the out-of-the-box examples for pyimpute and scikit-learn focus on single-target setups, there are totally feasible ways to adapt these tools for your multi-temporal use case. Let’s break down how to approach this, especially since you’re already digging into pyimpute’s source code.

Core Idea: Treat Multi-Date Targets as a Multi-Output Variable #

First, let’s clarify the shift from single to multi-target: instead of each pixel having a single target value (from one raster), each pixel will have a vector of target values (one per date in your time series). Most scikit-learn regression models (like RandomForestRegressor, GradientBoostingRegressor) support multi-output regression natively, which means they can predict all your date-specific targets in one go.

Here’s how to adapt pyimpute and scikit-learn step by step:

1. Prepare Your Multi-Date Target Rasters #

First, you’ll need to stack your target rasters (one per date) into a single 3D array where each pixel’s dimension holds the time series values. For example, if you have 5 date-specific target rasters, your target array will be (rows, cols, 5).

You can do this with libraries like rasterio or xarray:

import rasterio
import numpy as np

# List of paths to your date-specific target rasters
target_paths = ["target_date1.tif", "target_date2.tif", "target_date3.tif"]

# Read all target rasters and stack them
target_arrays = []
for path in target_paths:
    with rasterio.open(path) as src:
        target_arrays.append(src.read(1))
target_stack = np.stack(target_arrays, axis=-1)  # Shape: (rows, cols, n_dates)

2. Adapt Pyimpute’s Training Pipeline #

Looking at pyimpute’s core code, you’ll see that the train_model() function takes X (covariate features, shape: n_pixels x n_covariates) and y (target values, shape: n_pixels x 1 for single targets). For multi-date targets, you just need to reshape your y to be n_pixels x n_dates:

• First, flatten your target stack to a 2D array: y = target_stack.reshape(-1, len(target_paths))
• Then, filter out pixels where any target date (or covariate) has missing values (adjust this logic based on your needs—e.g., keep pixels with at least one valid date, or only pixels with all dates valid)
• Pass this 2D y to pyimpute’s train_model() along with your covariate X (same as the single-target setup)

Most scikit-learn models will handle the multi-output y automatically. If you’re using a model that doesn’t support multi-output natively, wrap it in sklearn.multioutput.MultiOutputRegressor.

3. Customize Prediction & Raster Output #

Pyimpute’s default predict() function writes a single-band raster, so you’ll need to modify the post-prediction step to output one raster per date:

After training your model and generating predictions (which will be a n_pixels x n_dates array):

# Reshape predictions back to the original raster shape
pred_stack = predictions.reshape(target_stack.shape)  # (rows, cols, n_dates)

# Write each date's prediction to a separate raster
for i, path in enumerate(target_paths):
    with rasterio.open(path) as src:
        profile = src.profile
    profile.update(count=1)
    with rasterio.open(f"prediction_date{i+1}.tif", "w", **profile) as dst:
        dst.write(pred_stack[:, :, i], 1)

4. Alternative: Build a Custom Scikit-Learn Pipeline #

If modifying pyimpute feels too restrictive, you can bypass it entirely and build a custom workflow with scikit-learn:

• Extract covariate features and multi-date target values for all valid pixels
• Use sklearn.multioutput.MultiOutputRegressor if your model requires it
• Train the model, predict on all pixels, then reshape and write the output rasters as shown above

Key Takeaway From Pyimpute’s Source #

As you noticed, pyimpute’s core logic revolves around pairing single pixel values with covariates—but the ML integration is just a wrapper around scikit-learn. The big win here is that scikit-learn’s multi-output support lets you extend this to multi-date targets without rewriting too much code. You just need to adjust how you structure your target data and how you write the final predictions back to rasters.

内容的提问来源于stack exchange，提问作者CatOfTheCanals