[Feature Request] Native Temporal Reducers for Climate Layers Aligned to 14-Day EO Windows

[robmarkcole]

I commonly configure Sentinel-2 with period_duration: "14d" so rslearn produces one Sentinel-2 mosaic per 14-day period within a window. I want climate layers to be reduced over those exact same 14-day windows.

For example, if a window spans several months and Sentinel-2 materialization yields one mosaic per 14-day interval, I want an ERA5-based climate layer to produce matching outputs for each of those same intervals. Typical reductions would be:

• TEMPORAL_MEAN
• TEMPORAL_MAX
• TEMPORAL_MIN

Example use cases:

• mean / max / min 2 m temperature over the same 14-day interval as each Sentinel-2 mosaic
• mean soil moisture over the same 14-day interval
• total or mean precipitation over the same 14-day interval

Current behavior does not support this directly: period_duration splits matching into sub-periods, but does not reduce the raster time dimension. Existing MEAN / MEDIAN compositing methods operate across overlapping items, not across timesteps within a multi-temporal raster

Requested capability

Add native temporal reducers for raster layers that operate over the T dimension within each matched item group / request time range:

• TEMPORAL_MEAN
• TEMPORAL_MAX
• TEMPORAL_MIN

These reducers should work naturally with period_duration, so that if a layer is matched into 14-day groups, each group can be reduced to a single raster representing that same 14-day interval.

Concrete example

Sentinel-2 configuration:

"data_source": {
  "class_path": "rslearn.data_sources.planetary_computer.Sentinel2",
  "init_args": {
    "cache_dir": "cache/planetary_computer",
    "harmonize": true,
    "query": {
      "eo:cloud_cover": {
        "lte": 100
      }
    },
    "sort_by": "eo:cloud_cover",
    "sort_ascending": true
  },
  "query_config": {
    "max_matches": 12,
    "period_duration": "14d",
    "space_mode": "MOSAIC",
    "time_mode": "WITHIN"
  }
}

Desired climate layer behavior:

{
  "layers": {
    "era5_14d_mean": {
      "type": "raster",
      "band_sets": [{
        "bands": ["temperature-2m", "total-precipitation"],
        "dtype": "float32"
      }],
      "data_source": {
        "class_path": "rslearn.data_sources.climate_data_store.ERA5LandHourly",
        "init_args": {
          "bounds": [-122.4, 47.6, -122.3, 47.7]
        },
        "query_config": {
          "max_matches": 12,
          "period_duration": "14d",
          "space_mode": "MOSAIC",
          "time_mode": "WITHIN"
        }
      },
      "compositing_method": "TEMPORAL_MEAN"
    }
  }
}

Expected result:

• sentinel2, sentinel2.1, sentinel2.2, ...
• era5_14d_mean, era5_14d_mean.1, era5_14d_mean.2, ...

With each climate layer aligned to the same 14-day interval as the corresponding Sentinel-2 group.

[robmarkcole]

close as done