HLWATER V1.0 - Trained Mask R-CNN for surface water mapping in boreal forest-tundra

The HLWATER V1.0 is a Mask R-CNN supervised model trained over PlanetScope (including Dove and Dove-R) satellite imagery (Analytic MS – Surface Reflectance) with focus on the automated detection and delineation of small water bodies in the circumpolar North. The model was trained over a dataset of very-high resolution 23,432 manually delineated lakes, ponds, rivers, streams, creeks and coastal sectors, located in diverse landscapes from the sporadic to the continuous permafrost zones of Canada. The training dataset is freely available upon request to pedro-freitas@edu.uslisboa.pt. 

Most training features (97%) were water bodies smaller than 1 hectare located in diverse environmental and hydrological settings. HLWATER V1.0 accuracy was tested for Eastern Hudson Bay (Nunavik, Subarctic Canada), a region that comprises a variety of water bodies, environments and permafrost in different stages of degradation. For the evaluation of model performance, a two-scale testing approach was used by comparing with manually delineated water bodies from: i) PlanetScope imagery (≈ 3 m pixel size, Very-high resolution - VHR), and ii) Unmanned Aerial System orthomosaics (0.05 – 0.15 m pixel size, Ultra-high resolution - UHR). The accuracy varied with landscape units, with mean Intersection over Union (IoU) 0.5 F1 Scores of 0.53 to 0.71 and mean F1 Scores of 0.62 to 0.95 at VHR. Setting 166 m2 as the minimum size detection threshold of the model, the IoU 0.5 F1 Scores were 0.7 to 0.91 and F1 Scores were 0.76 to 0.83 at UHR. 

The manuscript detailing this model is published as should be cited when using the model: Freitas, P., Vieira, G., Canário, J., Vincent, W. F., Pina, P., & Mora, C. (2024). A trained Mask R-CNN model over PlanetScope imagery for very-high resolution surface water mapping in boreal forest-tundra. Remote Sens. Environ., 304(114047). https://doi.org/10.1016/j.rse.2024.114047.

The folder contains the following options for using the model:

* .dlpk - Esri Deep Learning Package file

* .emd - Training file for running in ArcGIS Pro

* .pth - Training file for running in PyTorch deep learning framework

Folder naming meaning:

A - ID of the image chips

B - ID of the training

E - Number of epochs

B - Batch size + pre-trained model (ResNet)