Geospatial Data Types and Processing for ML

Data Modalities

Tabular Data

Tabular data represents point observations, sensor readings, or aggregated socioeconomic statistics.

• Usage Patterns:
  • MLPs: Used for independent observations, such as predicting house prices based on longitude, latitude, and square footage.
  • RNNs: Essential when data is time-indexed, such as historical river gauge levels or air quality sensor logs.
  • GNNs: Critical for relational data where spatial edges represent connectivity, such as traffic flow on a road network.
• Example: A dataset of weather stations where each row contains a timestamp, station ID, temperature, and coordinates.

Multispectral Imagery

Multispectral data consists of a few discrete bands (usually 3–15) representing specific portions of the electromagnetic spectrum. $x\in {\mathbb{R}}^{C\times H\times W}$ where $C$ is the number of spectral bands (e.g., Red, Green, Blue, Near-Infrared), $H$ is height, and $W$ is width.

• Example: Sentinel-2 imagery. The Near-Infrared (NIR) band is used to calculate the NDVI (Normalized Difference Vegetation Index) to assess plant health.

Hyperspectral Imagery

Hyperspectral sensors capture hundreds of narrow, contiguous bands, creating a “spectral cube” that approximates a continuous reflectance curve.

• Dimensionality Challenge: The high number of bands causes the “curse of dimensionality,” making models prone to overfitting and high memory usage.
• Preprocessing:
  • PCA: Compressing bands into a smaller set of uncorrelated components.
  • Band Selection: Choosing specific wavelengths known to react to certain materials (e.g., moisture absorption lines).
• Example: Identifying mineral compositions in a quarry, such as distinguishing between types of clay that appear identical in standard multispectral imagery.

Radar (SAR)

Synthetic Aperture Radar is an active sensor that sends microwave pulses and measures the return signal.

• Properties: Radar is complex-valued, containing both Amplitude and Phase. It penetrates clouds, smoke, and darkness.
  • Flood Mapping: SAR detects “specular reflection” (low return) of water on the ground even when optical sensors are blocked by clouds.
  • InSAR: Uses the phase difference between two passes to detect millimeter-scale ground subsidence or earthquake deformation.

LiDAR

Light Detection and Ranging uses laser pulses to map the 3D structure of the environment.

• Representation: A sparse 3D point cloud: $D=\{(x^{(i)},y^{(i)},z^{(i)}){\}}_{i=1}^N$
• Example: Creating a Digital Terrain Model (DTM) by filtering out vegetation and buildings from the point cloud to isolate the “bare earth” topography.

---

Preprocessing & Alignment

Coordinate Reference Systems (CRS)

The Earth is an ellipsoid, but ML models require flat, Euclidean grids for convolutional kernels to be spatially consistent.

• Geographic (Lat/Lon): Units are degrees. Inconsistent distances (e.g., 1 degree of longitude is shorter at the poles than the equator) make spatial operations difficult for deep learning.
• Projected (UTM): Units are meters. Crucial for CNNs because a $3\times 3$ kernel must represent a consistent physical area (e.g., $30m\times 30m$).
• Alignment: Always reproject all datasets into a common CRS before stacking them.

Resampling and Pixel Alignment

Datasets often have mismatched spatial resolutions (e.g., 10m Sentinel-2 vs. 30m Landsat).

• Categorical Data (Labels): Use Nearest Neighbor interpolation to avoid creating non-existent “average” classes (e.g., a pixel that is 1.5 when 1 is Forest and 2 is Water).
• Continuous Data (Features): Use Bilinear or Cubic interpolation for smoother transitions in reflectance or temperature values.
• Alignment Check: Even with the same resolution, rasters may be offset by a fraction of a pixel. They must be target-aligned to the same grid origin to prevent label noise.

Evaluation Strategies

Spatiotemporal Autocorrelation

Random pixel-level splits result in massive information leakage because neighboring pixels are nearly identical.

• Spatial Holdout: Testing on an entirely different geographic tile or city than the one used for training.
• Temporal Holdout: Training on data from one year (e.g., 2024) and testing on the next (e.g., 2025).
• Significance: Failure to use spatial splits often leads to artificially high accuracy scores that do not generalize to new regions.

Change Detection

Simple image differencing is often insufficient due to atmospheric and seasonal noise.

• Seasonal Shifts: Compare images from the same time of year (e.g., June 2024 vs. June 2025) to avoid “phenological noise” where seasonal leaf-drop is flagged as permanent land clearing.
• Atmospheric Effects: Differences in haze or cloud shadow can produce false change signals.