The hackathon dataset contains ~50.000 quality screened GEDI L2A shots distributed across southern Netherlands and northern Belgium. For each 25m footprint we provide a single response, *rh98*, in centimeters (cm). The data extraction and stratification are documented and presented in the lecture [***Introduction of cloud-native vector format: hands-on in Python environment  by Yu-Feng Ho***](https://github.com/yu-feng-ho/OGH2025).The acronym *rh* stands for relative height and measures the height difference between the detected ground return and the point at which 98% (hence *rh98*) of the cumulative lidar-waveform energy is reached. This metric is widely used as a proxy for dominant tree height because it captures the upper canopy while trimming most noise.

![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F22197053%2Fa891bcaa2d715e8c2bd09dbe823ac01a%2Fphoto_5872729262128220782_y.jpg?generation=1756384822427141&alt=media) *Fig 1: Distribution of GEDI Sample points in Benelux*


![
](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F22197053%2F422cc7d890580facdff853315f786a7e%2Fphoto_5872729262128220781_y.jpg?generation=1756384559077736&alt=media) *Fig 2: Vanilla model CHM over Wageningen*

The predictor variables you will be using to measure the canopy height are the 64 Google DeepMind AlphaEarth Foundations (AEF) remote-sensing embeddings; values are in the [-1, +1] interval. They capture multi-season Sentinel-1, Sentinel-2 + Landsat spectral-texture context at a spatial resolution of 10m. Attached is a [user-friendly](https://book.ml4rs.com/alpha-earth-foundations.html) description of the predictors, their production and usage.