The report is a great read for anyone interested in geo-spatial statistical analysis as well.

Model specifics start on page 64. It credits the Sentinel-1 for open source time series data for mountain movement and displacement. The challenge(as with all time series) is prediction(including future trends). I wish there was more of this data openly available to play with in R/python.

The early warning system in some parts of the state based on rainfall is a cool little logistic regression formulation on page 68 with three variables for thresholding based on rainfall.

f(z) = 1/(1+e^-z)) where `z = -3.817+0.077*Daily + 3DayRainfall*0.058 + 30dayRainfall*0.009`. Which seems intuitive. More rainfall in a day is likely to cause landslide than if the same amount fell over 30 days. Plugging in 742mm of rain already(June1-Sep30) for 30DayRainfall that has fallen this season it gives a crazy 17% chance of landslide (z=-3.81+0.009*742/3; f(z)=0.17). But the area(Figure 52) covered on a watch and warning list very large. 100's of people have died[4] so I am not sure how this early warning system is being used. But tourism is down by a lot[1] already.

The model for warning could likely be improved(calibrated) by studying other factors like proximity of infrastructure to rivers (Gaussian process are great for discovering such associations). Would have been nice to have some data to let other researchers get here.

[1] https://www.newslaundry.com/2023/08/29/in-himachals-gadsa-valley-villagers-point-to-impact-of-mining-felling-of-trees [2] https://www.isro.gov.in/Landslide_Atlas_India.html [3] https://en.wikipedia.org/wiki/Economy_of_Himachal_Pradesh [4] https://timesofindia.indiatimes.com/city/shimla/destruction-pours-in-himachal/articleshow/103002375.cms?from=mdr