SATPC0031551 Tab 04 4 SOW.pdf

SERVIR-Amazonia TensorFlow and Machine Learning Operations SME SoW

SME Activity: SERVIR Subject Matter Expert (SME) on Deep Learning and TensorFlow and Machine Learning Operations to support land cover monitoring services in the Peruvian Amazon

Background:

SERVIR is a joint development initiative of NASA and the US Agency for International Development (USAID), working in partnership with leading regional organizations around the globe, to help developing countries use information provided by Earth observing satellite data and geospatial technologies for managing natural resources and environmental risks. SERVIR is a project within NASA’s Applied Sciences Program (ASP) that focuses on specific geographic regions to develop thematically focused science products, tools and services that meet the needs of decision-makers in those regions. The four Thematic Service Areas are: Agriculture and Food Security, Water & Water-Related Disasters, Land Cover and Land Use Change and Ecosystems, and Weather and Climate. SERVIR is currently active in five regional Hubs: (1) SERVIR-Eastern and Southern Africa (E&SA), in Nairobi, Kenya; (2) SERVIR-Hindu Kush Himalaya (HKH) in Kathmandu, Nepal; (3) SERVIR-Mekong in Bangkok, Thailand;

(4) SERVIR-West Africa (WA) in Niamey, Niger; and (5) SERVIR-Amazonia (SAMZ) in Cali, Colombia.

The SERVIR Science Coordination Office (SCO), which is located at Marshall Space Flight Center (MSFC), participates in the work planning processes for hubs to provide guidance on existing and new NASA applied research. The SCO has the responsibility to provide science and Geospatial Information Technology (GIT) support as needed to enhance hub capacity and the scientific value of SERVIR products, tools, and services.

To support this responsibility, the SCO periodically needs subject matter expert (SME) support to address specific issues in the Thematic Service Areas and GIT. SME activities are established as needed to provide guidance and assist in the development of workshops, applications, products, tools, services and/or activities needed by SERVIR Hubs, SCO, SERVIR Global Network, and Applied Sciences Team (AST).

SME Task Description:

TensorFlow (TF) is a freely available platform for machine-learning (ML), provided by Google, that allows developers to choose from a variety of tools and libraries to develop ML applications. Over the past few years, neural network ML approaches have increasingly been applied to land-use change classification, deforestation assessment, and other image change analysis in the Earth Observation sciences, where among

SCO has many resources and pre-requisites material to help the SME craft training material and these materials will be provided to the SME near the start of the task. There is flexibility regarding the format of the lessons and exercises, with a preference for interactive sessions.

● Workshop materials (including slide decks, example scripts, practice data, etc.)

should be made freely and openly available. All materials will be hosted on a public-facing SERVIR training website.

Science Support*

● The SME will provide approximately 60 hours of direct science support during and after the workshop for 1) Detection of Selective Logging in the Peruvian Amazon, 2) Monitoring Informal Forest Roads (MOCAF) and 3) Gold Mining Monitoring with Deep Learning (a new version of RAMI). ACCA is already testing applications of neural networks and TF to determine the best algorithm. However, SAMZ needs support to implement techniques that will improve the accuracy of ACCA’s existing models, or if it’s viable, create new models with different architectures and workflow optimizations based on the application of DNN (pixel-based classifications), CNN (object-based classifications), and RNN (Time-Series analysis) for EO.

● The SME will provide recommendations for the optimal transfer and operationalization of the Monitoring Informal Roads, Selective Logging Activity in the Peruvian Amazon, and Gold Mining Monitoring (RAMI with Deep Learning) service given infrastructure (optimal use of SERVIR specific servers), data storage, Cloud integration, and computing need considerations.

*Science support tasks will be accomplished using freely available cloud computing resources.

The SME shall also deliver a virtual, one-hour post-workshop and consultation wrap-up meeting for the SERVIR SCO to discuss and identify further challenges and opportunities in operationalizing TF for land cover monitoring. As part of this wrap-up meeting, the SME will include an overview of the science support and workshops provided to SERVIR-Amazonia.

SME Deliverables:

As a result of the training tasks described above, the SME will provide:

● Lectures and hands-on exercises (in the form of slide decks, test scripts, example data, and interactive exercises) for the following workshop topics, all prepared by the week of August 14th, 2023:

Introduction to ML, Deep Learning, and AI

~1 hr

Intro to Google Colab, TF 2.0 and Keras Library

~1 hr

Overview of building blocks (tensors) of TF main methods, functions, and datasets

~1 hr

Integrating TF with Google Earth Engine and Google Cloud Platform

~1 hr

Building, compiling, and training a deep learning segmentation model to process datasets

~1.5 hr

Advanced techniques for Deep Learning, metrics evaluation, data augmentation, etc.

~1 hr

Testing difference architectures and transfer learning with Convolutional Neural Networks

~1.5 hr

TensorFlow advanced applications with deep learning object detection, regression, time-series analysis, hypertuning

~3 hrs (1.5 hrs of object detection and regression + 1.5 hrs of time-series analysis hypertuning)

● A 1-hour post-activity wrap-up meeting for the SERVIR SCO, in the form of a powerpoint

As a result of the science support tasks described above, the SME will provide:

● An updated NN-TF workflow for the Selective Logging Monitoring application that reduces errors of commission and improves the accuracy and computational efficiency of the logging detection model with SkySat, following requirements provided by SAMZ and SERVIR SCO

● An updated NN-TF workflow for the Monitoring Informal Roads application that reduces errors of commission and improves the accuracy and computational efficiency of the informal road detection model with Sentinel-2, following requirements provided by SAMZ and SERVIR SCO

● An updated NN-TF workflow for the Gold Mining Monitoring Service with Deep Learning application that reduces errors of commission and improves the accuracy and computational efficiency of the gold mining detection model with Sentinel-1, following requirements provided by SAMZ and SERVIR SCO.

● Specific recommendations for the service and service transfer/Machine learning operations in the form of a report following the provided template. Scripts and all related documentation should be made available to the SAMZ and SERVIR

SCO.

Period of Performance:

The period of performance will be from June 2023 - October 2023.

Travel:

This SME activity will be conducted both virtually and in person. In person components to the SME will include travel to Lima, Perú for the TF/ML Training in Earth Science Applications with ACCA/SERVIR Amazonia. The majority of the time, the SME will communicate virtually with SCO staff located in Huntsville, AL and SAMZ staff located in Perú.