land-cover-ca v2: land cover around Lake Simcoe, Ontario, at 30 m

Labels each 30 m pixel of a Sentinel-2 L2A summer median composite as forest, cropland, built-up, water, wetland or other. Fine-tuned from Prithvi-EO-2.0-300M (revision 9eb1b1102806593963daa333bcc491b1c6f8562f) for LakePulse, an entry to the 2026 NASA Space Apps Challenge. Its job is the land within 30 m of the Lake Simcoe shoreline and its streams, where runoff enters the lake, which 500 m MODIS land cover cannot resolve.

A model estimate, not a measurement. The training labels are the 2020 Land Cover of Canada, a published map (stated overall accuracy 86.9%, 832 samples), not field-checked ground truth. Every score below measures agreement with that map.

Made by Anjani Duddukuru.

Held-out test (split by area, never by pixel)

The test block is the Holland River watershed plus the Lake Simcoe shore within 10 km of it (541.3 km², 43.2 km of shoreline, 11 chips of 224 x 224 pixels, 474,703 labelled pixels, 44,745 of them in the 30 m strip). Chips within one chip width (6.72 km) of a test chip were kept out of training and validation. All models below were scored on the same chips. Mean IoU over the six classes.

Model Whole test block 30 m shoreline and stream strip
land-cover-ca v2 (this model; class-weighted loss) 0.5145 0.4040
v2 without class weights 0.4983 0.3860
Gradient boosting, per pixel, same six bands and training data 0.4794 0.3684
v1 (2020 imagery only, 86 training chips) 0.4724 0.3558
Always the most common class (forest) 0.0293 0.0384

Per-class IoU, whole test block: forest 0.655, cropland 0.801, built-up 0.5727, water 0.9044, wetland 0.0846, other 0.069. Wetland and "other" are poorly mapped.

Adversarial probes (one chip each, mean IoU): cloudy scene without cloud masking 0.2117, winter scene 0.0484, chip at the catchment edge 0.3211, mostly water (96.5%) 0.386, half water and half shore 0.4875. Use cloud-masked summer composites only.

Training

  • Data: 636 training chips and 30 validation chips from the 2019, 2020 and 2021 summer composites, all against the 2020 labels (the label map was made from 2019 to 2021 imagery). Area: 13 watershed units around Lake Simcoe plus 22 neighbouring units within 15 km (12,529.4 km²), Ontario Watershed Boundaries. Validation: Pefferlaw River unit.
  • Loss: cross-entropy weighted by the inverse square root of each class's share of training pixels, scaled to a mean of 1: forest 0.3479, cropland 0.4308, built-up 0.8362, water 0.6651, wetland 2.5612, other 1.1588.
  • Setup: TerraTorch 1.2.14, UNet decoder on backbone layers 5, 11, 17 and 23, AdamW at 1e-4, batch 8, bf16 mixed precision, Google Colab A100. Training ran in two attempts: Colab ended the first runtime at epoch 24, and training resumed from the epoch-19 best weights with a fresh optimizer. These weights are the best by validation loss in the second attempt.

Input

Six Sentinel-2 L2A bands in this order: BLUE, GREEN, RED, NIR_NARROW (B8A), SWIR_1, SWIR_2, as surface reflectance (digital number x scale + offset, with Earth Search's offset left out where it is already applied), resampled to 30 m. No-data reflectance is set to 0, then each band is normalised with norm_stats.json (means and standard deviations of the training chips). Output classes in index order: forest, cropland, built_up, water, wetland, other.

Loading

land_cover_ca_v2_fp16.safetensors is the fp16 state dict of a TerraTorch SemanticSegmentationTask. Build the task from terratorch_config.yaml with backbone_pretrained: false (and without backbone_ckpt_path), then load the state dict with strict=True after converting tensors to float32.

Limits

  • Agreement is with a 2020 published map, not with the ground.
  • Trained and tested only around Lake Simcoe, southern Ontario. Scores elsewhere are unknown.
  • Smoke and thick haze pass the cloud mask and bias results; the project's composites drop pixels whose L2A aerosol optical thickness is above 0.4.
  • Against MODIS MCD12Q1 land cover for 2017 to 2025, this model's forest, cropland and urban shares are on average 23.80, 15.03 and 11.40 percentage points higher. It follows the 2020 federal map it learned from; MODIS is far below that map too.

Files

File What
land_cover_ca_v2_fp16.safetensors Weights (fp16)
terratorch_config.yaml Model and data module settings used in training
norm_stats.json Band means and standard deviations, class names, ignore index
config.yaml The track's full configuration (data sources, classes, split)
eval.json Test and probe scores of this model
NOTICE, LICENSE, SHA256SUMS Attribution, licence, checksums
ablations/ The two comparison models from the table above, for checking the comparison only

Ablations (for comparison, not for use)

ablations/ holds the two other models in the test table, so the comparison can be rerun:

  • ablations/v1/: 2020 imagery only, 13 watershed units, 86 training chips, no class weights. Did not beat gradient boosting (0.4724 vs 0.4689 on v1's own baseline data; 0.4724 vs 0.4794 here). Normalised with its own norm_stats.json, which differs from this model's.
  • ablations/v2_no_class_weights/: the same data as this model, without class weights. Wetland IoU 0.0114 against this model's 0.0846: the class weights account for most of the wetland gain.

Each folder has its weights, terratorch_config.yaml, norm_stats.json and eval.json. Load each with its own config: this model's state dict holds one more tensor (its loss weights) than the two ablations. In the project's MLflow registry these are land-cover-ca versions 1 (v1), 2 (this model) and 3 (v2 without class weights).

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for anjanid/land-cover-ca

Finetuned
(1)
this model