High Resolution Near Real Time Land Use and Land Use Change

OUSD (R&E) MODERNIZATION PRIORITY: Artificial intelligence/machine learning TECHNOLOGY AREA(S): Information systems, modeling and simulation technology OBJECTIVE: Develop a high-resolution fully automated land use and land use change (LULUC) map of the globe, updated daily, using commercial or publicly available satellite imagery. Identify mission-specific types of change in near real-time across broad areas. DESCRIPTION: NGA produces timely, accurate, and actionable geospatial intelligence (GEOINT) to support national policymakers on matters of national security and to support federal agencies responding to humanitarian and disaster relief efforts. Many of NGA's GEOINT products begin with LULUC maps detailing environmental conditions and changes relating to human activities and natural phenomena. Time series of LULUC maps enable deeper analysis and the development of follow-on predictive analytics relating to broad topics in environmental security and national security. Recent advances in deep learning have dramatically improved the state-of-the-art (SoTA) for techniques such as large-scale semantic segmentation and change detection, which may be applied to accurately and efficiently produce LULUC maps [1]. Concurrently, the volume of available satellite imagery has grown tremendously, including commercial imagery products that image the entirety of the earth every day at high resolution. Together, these advancements in deep learning and imagery availability may be used to produce highly accurate LULUC maps of the globe, enhancing NGA's GEOINT capabilities (e.g., [2]). Only direct to Phase II proposals are being accepted under this topic. A direct to Phase II proposal must demonstrate the proposer's possession of an existing prototype LULUC capability that is at a minimum equivalent to the Phase I deliverables below. Performers should improve upon the SoTA for LULUC mapping by (1) increasing the resolution and accuracy of LULUC segmentation maps and (2) decreasing the time required to produce LULUC maps and associated GEOINT products to at least weekly and ideally daily (weather conditions and imagery collection allowing). PHASE I: A successful Phase 1 will result in a 10-30 m resolution, 6+ land use class LULUC mapping capability covering at least 60% of the landmass of the globe, which can be updated automatically on demand with <3 days of combined human effort and compute time. A >1500 km2 LULUC example should be made available for demonstration covering at least two separate dates at the same areas. PHASE II: In addition to specifying the performer's existing Phase I capability, the performer must identify the SoTA for LULUC mapping and its plan for surpassing the SoTA supported by sound scientific and engineering principals. A successful Phase II will result in a <10 m resolution, 10+ land use class LULUC mapping capability, which can be updated daily if weather conditions and imagery collection allow. 6+ significant change types, at least three of which are directly anthropogenic, must be automatically identified. Performers will be expected to provide comprehensive reports detailing technical advancements and performance metrics, which will be provided to NGA and submitted to an academic journal or conference. LULUC maps and associated products produced during the period of performance shall be delivered to NGA without further use restrictions. Collaboration with a program of record at NGA (e.g., SAFFIRE) for potential integration at the end of Phase II is preferred. PHASE III DUAL USE APPLICATIONS: Accurate, timely, and high-resolution LULUC products are a critical source of monitoring global change caused by environmental factors and human activities. GIS analysts across a variety of Government and commercial sectors rely on these mapping products to improve understanding on topics such as land use planning, hydrology, food and environmental security, and resource allocation and management. REFERENCES: Khan S., Alarabi L., and Basalamah S., Deep hybrid network for land cover semantic segmentation in high-spatial resolution satellite images, Information 2021, 12, 230. doi.org/10.3390/info12060230. A new land cover map of the world, ArcGIS StoryMaps, storymaps.arcgis.com/stories/486cd2ae2016454f951c97f802f125b3, accessed 24 September 2021. KEYWORDS: Land use, land cover, land use change, remote sensing, computer vision, machine learning, deep learning, segmentation