Deep Learning-Based Decadal Land Cover Mapping and Investigation of Reindeer Mobility Patterns in the Sattasniemi District, Finland, Using Sentinel-2 and GPS Data

Keywords: Deep Learning, 3D CNN, LSTM, LULC, GPS Data, Reindeer Mobility, Arctic Boreal Forest
Abstract Type: Paper Abstract

Abstract

The increasing availability of high-resolution remote sensing data, combined with the power of deep learning techniques, offers a robust framework for understanding ecological systems and animal movement in Arctic regions. This research deploys a hybrid deep learning framework consisting of a 3D Convolutional Neural Network (3D CNN) and a Long Short-Term Memory (LSTM) model for land use and land cover (LULC) mapping and investigates reindeer mobility patterns in the Sattasniemi district of Finland, using Sentinel-2 and reindeer GPS collar data. Our hybrid deep learning model effectively captures spatial features (3D CNN) and temporal dependencies (LSTM), outperforming traditional machine learning models. We trained our model on a local wall-to-wall labeled data (pasture map) and Sentinel-2 images, and our custom 3D-CNN-LSTM model achieved satisfactory result with a high accuracy (86%). Following a successful training and fine-tuning of our deep leaning model, we produced annual and decade-long LULC maps from 2015 to 2024, which we used to analyze reindeer mobility patterns in conjunction with GPS collar data. Our high-resolution spatiotemporal land cover maps, overlaid with GPS data, reveal interesting insights into reindeer mobility across time and space. We believe that our spatiotemporal land cover maps and findings on reindeer mobility will be an asset to researchers, policymakers, and reindeer herders in the Arctic-Boreal Forest region.