Lumonitor is short for "land use monitor", but we are really only interested in one thing: human impacts. For this project we modeled the intensity of human impacts using a CSP dataset on Anthropogenic Impacts and Landsat 8 data from the Harmonized Landsat Sentinel dataset. Model results are viewable here. This project was generously funded by the Microsoft AI for Earth Program.
Our goal was to train a UNet-type autoencoder on the 30-m Human Modification dataset from the Disappearing West/U.S. project and Landsat 8 data from the Harmonized Landsat sentinel. For training data, we used the cloud-free annual median for each pixel in the seven available bands.
We used cloud-free annual per-pixel median values of the Harmonized Landsat Sentinel dataset and various label data to train our model. The HLS data are available on Azure Open Datasets. Upon initiation of the project there was no STAC or other catalog available (this need is now filled by the Planetary Computer for many datasets). However, loading each image within a single year even for a single 3660x3660 pixel "tile" consumed too much memory, essentially eliminating a GDAL / rasterio / numpy type approach. The approach we eventually settled on was to deploy pangeo on a daskhub instance running on Azure Kubernetes Service. Using xarray/dask for calculating the medians took approximately 2 minutes per tile (for each of approximately 1000 tiles spanning the conterminous US in each year). Code for this step is available here.
Label data were exported from Google Earth Engine and storage on Azure Blob Storage using this script.
One goal of our setup was to dynamically sample the training data directly from Blob Storage while running the model. In implementing our workflow, we determined the best performance would be gained by reading from COG format using the rasterio.windows module. Training data were initially exported in zarr format (since that is the only format which supports writing directly to Azure Blob Storage from xarray), so we downloaded and converted all files, then reuploaded to Blob Storage, using this script.
We used Pytorch and Azure ML to train our models. We spent the typical amount of time needed to test different models, etc. Some features of our process (besides dynamically sampling from COG as mentioned above) include a "dual-function" dataset which pulls random chips, but also gridded data representing all chips for a predefined area. Each can also be pulled within a specific area stored in a shapefile etc. Because the training data was heavily zero-inflated, training and test loss were typically extremely low. However, model output was excellent in most cases. But to further evaluate a running model, we developed a process to predict and symbolize the results over small pre-defined test areas at each epoch, allowing using to evaluate model efficacy in real time beyond our loss statistics. Model training was accomplished using src/model/run_training.py and src/model/train.py.
The use of a UNet-type model with many convolutions meant that for a typical 512x512 "chip", only the center 70x70 pixel area was unaffected by padding/cropping. For continental-level prediction, this meant we had to pull and evaluate millions of chips. We used Azure ML and up to 20 K80 GPU instances to produce results in a set number of "regions" (typically 100) across the area of interest. This process currently takes about 8 hours per year of data - a reasonable time. Prediction code files are src/model/run_prediction.py and src/model/predict.py.
Color ramps for each impact type were prepared "by hand" in QGIS and used to create raster tiles served from blob storage to the web application at https://cspbeta.z22.web.core.windows.net/project2.