How GeoSutra handles evidence
Environmental data is only useful if you can see where a number came from and how far to trust it. These are the rules the Observatory is built on.
Environmental data is only useful if you can see where a number came from and how far to trust it. These are the rules the Observatory is built on.
Every number links to the dataset, provider, version and licence it came from. The full list is on Data & Sources.
Each figure carries an evidence badge: Observation (measured by a sensor), Model / reanalysis (ERA5-Land, CMIP6, WorldCover, WorldPop, Hansen tree cover), Reported (BIPAD incidents, GBIF records, OpenStreetMap) or GeoSutra-derived (a calculation we ran). ERA5-Land is a ~9 km reanalysis, not a weather station; a FIRMS detection is a thermal anomaly, not a confirmed wildfire; Hansen “tree cover” is canopy taller than ~5 m, not Nepal’s legal forest area.
Where a method produces an interval, a p-value or a model spread, it is shown next to the number. Trends come with their test and p-value; a trend that is not statistically detectable is reported that way, never as “no change”. Where no uncertainty estimate exists, the result says so instead of inventing a confidence score.
If a dataset is unavailable, too cloudy or does not cover the area, the result is “not available” with the reason. Missing data is never shown as zero, and one dataset is never silently substituted for another.
A dataset producer’s published accuracy is not GeoSutra’s validation. For example, ESA reports about 76.7 % global overall accuracy for WorldCover 2021 v200; GeoSutra has not established Nepal-specific accuracy. Today no Observatory method is GeoSutra field-validated, and every method says so.
Every method in the method library has a version. A result records the method version and dataset versions it was produced with, so a later change of method cannot silently rewrite an old result.
Every analysis records its area, period, parameters, datasets and processing back-end, and can be exported as a reproducible JSON request and cited.
Every method lists its limitations and what it is not suitable for — for example, NDVI trends are not biomass, active-fire detections are not burned area, and incident points placed at a district default are not incident locations.
A traditional GIS shows layers. An AI chatbot gives an answer. The Observatory answers a question with a registered method, the data it used, the calculation step by step, the uncertainty and the limitations — so a researcher can reproduce it, a municipality can use it and a journalist can check it. No AI model computes any environmental number here.