DIST-S1-Calibration

We assume that change maps are functionally represented as a thresholded distance function between pre-event and post-event images. That is to say, we have some "metric" that we quantify disturbance in a "post-image" scene relative to a set of "pre-images": $$ \textrm{dist}({\textrm{pre-images}}, \textrm{post-image})\rightarrow \mathbb R_{\geq 0} $$ Using $\textrm{dist}$, we can find a suitable $T \in \mathbb R$ such that a change map is given by $\textrm{dist} > T$. Often $T$ is determined empirically at given site, but as we are determining these change maps operationally, we need to calibrate $T$ and additional metric parameters. These notebooks are determining provide some insight into this calibration activity.

Install

Install dist-s1 environment (see the environment.yml from the research repository) and install distmetrics (which is currently private repo). Also needed is einops for the transformer model (see the distmetrics repository and it's environment.yml). Make sure to install the dist-s1 kernel too.

Usage

python run.py --event chile_fire_2024 --distmetric_name 'mahalanobis_2d mahalanobis_vh mahalanobis_1d_max log_ratio_vh transformer'