Time Series Support

This issue is designed to track progress towards full time series support in TorchGeo. If you are interested in helping with this effort, please join the #time-series channel on Slack and attend our weekly meetings.

Datasets

Required changes

• Replace rtree with geopandas: #2747
• Replace fiona with geopandas: #2962
• Spatial-only intersection: #2837
• Xarray support: #509, #1490, #2945
• Collation functions for padding: #2921
• Convert change detection datasets: #2862, #2863, #2878, #2879, #2906, #3028
• Return time series: #2308, #3183
• Convert time series datasets: #3042, #2921, #3368, #3369, #3371, #3372
• Splitters: guarantee non-overlapping spatial and/or temporal splits: @hkristen

Desired changes

• Replace BoundingBox with slicing: #2804
• Support Polygon ROIs in all splitters: #2835
• Deprecate BoundingBox: #2847
• Store true image footprint: #2903
• Properly handle vector datasets: @isaaccorley
• Cache geopandas index: @isaaccorley
• STACDataset: see #stac channel
• ERA5
• CMIP6
• Support integer indexing?: @adamjstewart
• Unify GeoDataset and NonGeoDataset?: #1353
• Intersection/Union multiple queries?: #2408
• Tuple/MultiQuery dataset?: #877

Samplers

Required changes

• Support Polygon ROIs in all samplers: #2812
• Separate spatial and temporal samplers: #3552
• Ensure samplers do not sample outside Polygon bounds: #3552
• Visualization: visualize an entire epoch with polygons and points: #3552

Desired changes

• Support weighted samplers?: #757
• Pre-computation?: run sample_points once per epoch for debugging?: #2300

Transforms

• B x T x C x H x W <-> B x TC x H x W: #2477 (TODO: image + mask)
• Temporal embeddings: #2826
• Temporal up/downsampling: https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.interpolate.html
• UTC time <-> local time
• Data imputation: https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.interpolate.html

Models

1D

• L-TAE: #2867
• Presto: #2929
• Tessera: #3243
• Seq2Seq: #2639
• TempCNN
• Chronos v1, v2: @cnagda?
• Galileo: #3305

2.5D

• BTC: #2926
• ChangeStar: #157
• ChangeViT: #2910
• FC-Siam: #108, #345

3D

• Satlas: #1358, #1884, #2038
• ConvLSTM: #2898
• OlmoEarth: #3517
• CNN-LSTM: #2938
• AnySat
• OmniSat
• Prithvi v1, v2: @isaaccorley
• SatMAE: @isaaccorley
• SkySense v1, v2

4D

• Aurora: #2823
• ACE: #3402

Tasks

• TC -> scalar, classification: TemporalClassification
• TC -> scalar, regression: TemporalRegression (#3668)
• 2CHW -> HW, classification: ChangeDetection (#2422, #2890)
• TCHW -> scalar, classification: SpatioTemporalClassification (#3591)
• TCHW -> scalar, regression: SpatioTemporalRegression (#3591)
• TCHW -> HW, classification: SpatioTemporalSegmentation (#3671)
• TCHW -> HW, regression: SpatioTemporalPixelwiseRegression (#3086)
• TCHW -> BHW, classification: SpatioTemporalInstanceSegmentation
• TCZYX -> TCZYX, regression: WeatherForecasting

Tutorials

• 1D: air pollution tutorial (#3669)
  • Dataset: air quality (#3619)
  • Model: L-TAE (#2867)
  • Trainer: TemporalRegressionTask (#3668)
• 2.5D: Change detection tutorial (#3222)
  • Dataset: OSCD100 (#3221)
  • Model: BTC (#2926)
  • Trainer: ChangeDetectionTask (#2422)
• 3D: Crop type mapping tutorial (#3706)
  • Dataset: PASTIS100 (#3265)
  • Model: ConvLSTM (#2898)
  • Trainer: SpatioTemporalSegmentationTask (#3671)
• 4D: Weather forecasting tutorial (#3654)
  • Dataset: WeatherBench (#3649)
  • Model: Aurora (#2823)
  • Trainer: WeatherForecastingTask (not strictly required right now)

Outdated discussions

• #640
• #1544