Spatio-temporal deep learning with daily medium-res satellite imagery

Authors: Jesus Martinez-Manso*, Planet Labs
Topics: Land Use and Land Cover Change, Temporal GIS, Spatial Analysis & Modeling
Keywords: satellite, deep learning, time series, spatio temporal, deforestation
Session Type: Paper
Day: 4/4/2019
Start / End Time: 3:05 PM / 4:45 PM
Room: Capitol Room, Omni, East
Presentation File: No File Uploaded


The Dove constellation of satellites from Planet Labs image nearly the entire Earth's landmass on a daily basis at 3.5 meter ground resolution. Every day, almost 2 million B-G-R-NIR images are captured, geometrically rectified, radiometrically calibrated and made available for use in analytic applications. This constitutes an unprecedented dataset and enables dense time series analytics on a global scale. In this work, we present methods applied by the Analytics Engineering team at Planet labs to perform land cover change characterization in the spatio-temporal domain using deep learning. In particular, we will cover the use case of deforestation, describing the framework to perform semantic segmentation of imagery and subsequent modeling of the temporal stack. We will discuss and compare two temporal models: a probabilistic inference model and a trainable recurrent neural network with spatial convolutions. To train the latter, a large quantity of change labels is required, which is very expensive to generate manually. Instead, we use an unsupervised method to automatically create change labels with associated transition times. Together, these methods allow us to produce deforestation alerts with sub-weekly latency.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login