About

Introduction

New catalogues of near-daily temporal data will soon dominate the global archives, offering unprecedented possibilities for rapid and continuous land monitoring. However, there has been little exploration of artificial intelligence (AI) techniques to leverage the spatiotemporal dimension at scale. Training data remain rare relative to the spatiotemporal sampling that is necessary to adequately capture natural and man-made phenomenology latent in these large volumes of high cadence data.

The project aims to establish the foundations for the next generation of continuous land monitoring applications. The project’s ambition is to enable new and better ways of measuring and understanding the human footprint on our planet, which is a key challenge of the UN Sustainable Development Goals. Together, the project team will develop improved AI processes, with focus on high resolution and cadence, and provide critical training data to establish the foundations for the next generation of Copernicus Land Monitoring Service (CLMS) products. By fusing satellite data from Copernicus and third party high-resolution sources, we will provide intensified monitoring of Land Cover (LC) and Land Use (LU), Land-Use Change and Forestry (LULUCF) at a much higher level of detail and temporal cadence than is possible today.

Objectives

  1. Establish the foundations for the next generation of rapid cadence land monitoring applications and in particular for the Copernicus Land Use Land Cover (LULC) suite of products.
  2. Demonstrate the fusion of Copernicus high resolution satellite imagery and third party very high-resolution imagery.
  3. Provide intensified monitoring of Land Use, Land-Use Change and Forestry (LULUCF), at a much higher level of detail and cadence than it is possible today.
  4. Enable new and better ways of measuring and understanding the human footprint on our planet, which is a key challenge of the UN Sustainable Development Goals.

Approach

The project will explore and insert groundbreaking new experimental modules on top of existing platforms and production pipelines. It follows a two-fold strategy involving:

  1. Creation of the most complete and dense remote sensing spatiotemporal Machine Learning training corpus to-date, combining Sentinel-2 with high cadence, very high resolution, harmonized multispectral 3m Planet imagery at 500,000 patch locations over Europe, and open sourcing these datasets for the benefit of the entire remote sensing community.
  2. Development and benchmarking of alternative ways of detecting and classifying change from very high cadence observations by training state-of-the-earth multiscale supervised and unsupervised Deep Learning (DL) classifiers on these unique data sources.

Highlights of the approach and relation to the project’s objectives and deliverables

Specific Objectives

Deliverables

Approach

Stimulate the development of new spatiotemporal modelling applications with the release of the most comprehensive spatiotemporal datasets ever created

Revolutionary dense spatiotemporal dataset to drive innovation in AI4IO beyond the scope of this project.

  • Planet’s Dove Satellites and relevance to the Copernicus Mission
  • CESTEM: Groundbreaking approach to harmonization and daily-gap filling of high resolution time series
  • Spatiotemporal sampling methodology for large training corpus of EO data

Implement and benchmark novel DL architectures for change detection that can realize the full potential of the new combined data sources

DL architectures for patch-based change detection that leverage the uniqueness of the above datasets

  • Patch classifier and supervised change detection approach
  • Weakly or fully unsupervised models, which can learn directly from the data what is change

Demonstrate improved LULC classification using daily harmonized multisensor, multiscale imagery

Practical demonstration on how this benefit Copernicus: prioritization and update of CLC in selected AOI(s)

  • Methodology for an improved LULC product

Demonstrate large scale thematic change detection with high temporal cadence using daily harmonized multisensor, multiscale imagery

Monthly thematic, high resolution change detection over Europe

  • Multi-resolution, multi-stage ensembles
  • Validation of change maps and calibration of change detection models

Next generation monitoring via sensor fusion

Planet Fusion Monitoring will be heavily used throughout the RapidAI4EO project. It is Planet’s first in a series of game-changing Fusion products that combine multiple data types and refine them into a single information stream. The CubeSat-Enabled Spatio-Temporal Enhancement Method (CESTEM, Houborg and McCabe, 2018) is a groundbreaking approach to enhance, harmonize, inter-calibrate, and fuse cross-sensor data streams leveraging rigorously calibrated ‘gold standard’ satellites (i.e., Sentinel, Landsat, MODIS) in synergy with superior resolution CubeSats from Planet.

CESTEM – approach to harmonization and daily gap filling of high resolution time series

NDVI values for an example site; comparison between the Sentinel-2, Landsat and PlanetScope data and the fusion dataset CESTEM

Expected Results

  1. Open source remote sensing spatiotemporal Machine Learning training corpus, combining Sentinel-2 with 3m Planet imagery at 500,000 patch locations over Europe.
  2. New land monitoring prototype to demonstrate:

a.   monthly high resolution thematic change detection over large areas of the EU (June 2022)

b.   highly effective end-to-end process to monitor and update the CORINE land cover product, with emphasis on improved understanding of land use with faster update cycles and reduced maintenance costs (March 2023)

Workflow showing relationship of the main engineering work

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 101004356.