A Scalable Deep Learning Pipeline for Mapping Forest Disturbances

EasyChair Preprint no. 11077

Abstract

We present a scalable and flexible approach for mapping forest disturbances over large areas. The approach harnesses the power of a spatio-temporal deep learning model, enabling it to capitalize on complex patterns while demanding minimal data pre-processing efforts. Our methodology enables rapid map production, offering a streamlined workflow. The approach was demonstrated over central Europe, covering an area of approximately 900,000 km2. Utilizing a modest cluster, the processing time amounted to just 36 hours. The results, produced at 20 m spatial resolution, exhibit coherent patterns and promising accuracy values, with an overall accuracy of 92%. We classify disturbances into four categories, marking a significant stride towards forest disturbance attribution. Identifying areas for improvement, we aim to reduce residual artifacts and enhance accuracy by incorporating higher quality training data. Future work will focus on refining the model architecture and expanding the dataset coverage to further optimize the approach's performance and accuracy.

Keyphrases: deep learning, forest, Sentinel-2, time series

@Booklet{EasyChair:11077,
  author = {Loïc Dutrieux and Keith A. Araño and Nelly Gaillard and Pieter S.A. Beck},
  title = {A Scalable Deep Learning Pipeline for Mapping Forest Disturbances},
  howpublished = {EasyChair Preprint no. 11077},

  year = {EasyChair, 2023}}