The Forestry Department of the United Nations Food and Agriculture Organization (FAO) has introduced the first version of the System for Earth Observation Data Access, Processing and Analysis for Land Monitoring (SEPAL) in 2013 with the support of Norway. SEPAL and recently launched SEPAL version 2 emerged from concrete and practical needs based on years of experience and expertise in capacity building in developing countries using remote sensing in forestry and land monitoring applications. SEPAL is a cloud-based computing platform for fast access and processing of remotely sensed data sources designed which was built in the first place to assist national forest monitoring and reporting for the Reducing Emissions from Deforestation and Forest Degradation, Forest Conservation, Sustainable Management of Forests and Enhancement of Carbon Stocks in Developing Countries (REDD+) mechanism.
FAO, as a technical implementing agency of the UN, is responsible for assisting countries in development of national forest monitoring systems (NFMS) and operational satellite land monitoring systems (SLMS) to follow the implementation of national REDD+ policies and measures, and to monitor the results of these efforts. SEPAL has succesfully supported countries to establish and maintain a NFMS SLMS capable of producing the information required to make consequential decisions about forest management; decisions that promote sustainable forest management and can potentially mitigate the effects of global climate change. Specifically, SEPAL addresses the existing challenges countries face when developing forest monitoring systems due to difficulties accessing and processing remotely sensed data; a key source of information for monitoring forest area and change over large, often remote forest areas. After the Paris agreement, the use of SEPAL in-country is moreover expanded to also monitoring the implementation of the NDCs and SDGs.
SEPAL functionality depends on four main factors, (i) a powerful, cloud-based computing platform for big data processing and storage, (ii) a direct link to earth observation data repositories, (iii) an open-source set of software tools capable of efficient data processing and (iv) project funding allocated to the purchase of high resolution datasets for participating countries. The computing platform enables national/local technicians and the national governments to process data quickly without locally maintained high performance computing infrastructures. The direct link to data repositories (e.g. Landsat, Sentinels, radar data) allows fast access to satellite imagery and other earth observation data for processing. The software tools, initially based on the FAO’s Open Foris Geospatial Toolkit (www.openforis.org) perform powerful image processing, are completely customizable and function similarly ‘on the cloud’ or on the desktop. The Bfast algorithm of WUR (NL) was deployed on SEPAL in order to help advance countries in their monitoring of actual changes and degradation. Some practical examples will be shown (e.g. Uganda).
SEPAL has been deployed initially in selected UN-REDD participating countries as well as FCPF countries. The project is led and run by a Project Team in the Forestry Department of FAO, with the assistance of an Advisory Group with representatives from key user and provider institutions such as the Global Forest Observations Initiative (GFOI), World Bank Forest Carbon Partnership Facility, space agencies and user countries. The SEPAL system is actually being deployed as a country system in Ecuador. Other applications of the SEPAL cloud has been succesfully implemented, namely in the context of SAFE (Safe Access to Fuelwood and Energy), as well as other monitoring options (e.g. near real-time monitoring, peatland monitoring and hazards). Finally, given there is interest to test the platform also for marine applications, this presentation will touch upon more possible applications and the way forward towards tackling climate change beyond the forest sector.

In Airbus Defense & Space Intelligence we have an Open Innovation initiative.
We are supporting various events (hackathon/challenge...) such as Copernicus Masters, GeoChallenge, ActInSpace, HelloTomorrow, VivaTech...
Fort that we have a tool named sandbox where the users can find Airbus satellite images dataset for free during the competition:
https://sandbox.intelligence-airbusds.com/web/

It would be great to collaborate during your event in order to propose our services.
We can also give prizes through Voucher Prizes on our OneAtlas Data web portal for the winners.

More info about our initiative here:
https://www.youtube.com/embed/uL_MpdGlcjc?rel=0&showinfo=0

In addition to Earth observation data NASA and ESA have been collecting for over 40 years, upcoming missions will deliver exponential growth in these agencies’ archives. NISAR alone will collect 80-90TB of data a day. Given these data volumes, storing and processing on the cloud is not just a feature but a requirement for performing scientific analysis. Existing tools and infrastructure do not enable scientists to take full advantage of all available datasets.

Will the cloud present unprecedented opportunities in scientific research and discovery? Can the cloud provide collaborative spaces for scientists, so they can stop working in isolation, likely duplicating work?

The Multi-Mission Algorithm and Analysis Platform (MAAP), developed jointly by NASA and ESA, will deliver cloud resources for scientists to share and collaborate on data and algorithms. This talk will cover practical and technical concerns of developing a cloud-based open science platform to foster scientific collaboration on data and algorithms to support above-ground biomass estimation. The MAAP team includes biomass research scientists who drive development through consulting on datasets, libraries and interfaces required to support biomass analyses.

The MAAP will host interfaces with access to over 80 datasets including biomass data from field, airborne and satellite LiDAR (e.g. GEDI) and airborne and satellite SAR data (e.g. NISAR), supplemented by data from a broader range of instruments. NASA recognizes the value of reusing and open sourcing software, so MAAP architecture is comprised of existing NASA open-source software components. These include the Common Metadata Repository (CMR) for metadata catalog, Earthdata Search Client (EDSC) for data search and discovery, HySDS for data processing, and Cumulus for data ingest, archive, and processing. MAAP is contributing additional features to these software components based on biomass user requirements.

However, there is no one-size fits all platform for processing and analysis of the remote sensing-data. The MAAP teams are adopting and developing OGC standards which extend beyond the platform itself, such as standards for dataset metadata, algorithms and web processing systems. The adoption and development of common standards will answer questions on how to run an algorithm developed on a separate platform and how a user can add and publish new datasets.

At a minimum, MAAP will save scientists from the burden of downloading datasets, and installing libraries supported by the platform. Further, MAAP will feature cross-agency data and algorithm sharing. At the most, it will advance the research fields of open science and biomass estimation through greater scientific collaboration, sharing and reproducibility.

Resource management of image-processing workflows with Deep Learning

Keywords: Reinforcement Learning, Image processing workflows, Reinforce

When producing a useful and readable satellite images, a complex and resource intense process transforms the initial onboard satellite sensor information into the final satellite product.
In this workflow, each calculation process is defined as a job that is scheduled into an HPC or Cloud. The resources such as memory usage in GB, CPU usage and WallTime (time of calculation) are given for a single job as well as for the HPC. The problem is to optimize the resources of the HPC for a given workflow of job.
We first analyze the real resources consumed by each single job by monitoring the output logs from the HPC resource manager application (PBS).
Thereafter the optimization problem is to find the best job schedule in order to reduce a target function. For a HPC, with a fixed set of resources, the target function is the total time to finish a given workload; for the cloud, with an elastic set of resources, the target function is the total cost.

We show that this problem can be formulated within the framework of Reinforcement Learning (RL). RL is very useful methodology to solve this kind of NP-complex problems because the dynamics of the problems are unknown, and the model finds the best job arrangement by learning and exploring different solutions by simulation.

We use a modify version of the REINFORCE algorithm developed by Sutton [1] and a deep learning algorithm to describe the possible set of actions that the model can perform.

We apply this model in the context of a real workflow of satellite products in a HPC, and a workflow schedule for a given application based on COPA, a Capgemini’s internal orchestrator in the cloud.

[1] Sutton Learning to Predict by the Methods of Temporal Differences: Machine Learning 1988
.