Search for: All records

This paper assesses trending AI foundation models, especially emerging computer vision foundation models and their performance in natural landscape feature segmentation. While the term foundation model has quickly garnered interest from the geospatial domain, its definition remains vague. Hence, this paper will first introduce AI foundation models and their defining characteristics. Built upon the tremendous success achieved by Large Language Models (LLMs) as the foundation models for language tasks, this paper discusses the challenges of building foundation models for geospatial artificial intelligence (GeoAI) vision tasks. To evaluate the performance of large AI vision models, especially Meta’s Segment Anything Model (SAM), we implemented different instance segmentation pipelines that minimize the changes to SAM to leverage its power as a foundation model. A series of prompt strategies were developed to test SAM’s performance regarding its theoretical upper bound of predictive accuracy, zero-shot performance, and domain adaptability through fine-tuning. The analysis used two permafrost feature datasets, ice-wedge polygons and retrogressive thaw slumps because (1) these landform features are more challenging to segment than man-made features due to their complicated formation mechanisms, diverse forms, and vague boundaries; (2) their presence and changes are important indicators for Arctic warming and climate change. The results show that although promising, SAM still has room for improvement to support AI-augmented terrain mapping. The spatial and domain generalizability of this finding is further validated using a more general dataset EuroCrops for agricultural field mapping. Finally, we discuss future research directions that strengthen SAM’s applicability in challenging geospatial domains.

 

Accurate representation of permafrost carbon emissions is crucial for climate projections, yet current Earth system models inadequately represent permafrost carbon. Sustained funding opportunities are needed from government and private sectors for prioritized model development. 

Deciduous tree cover is expected to increase in North American boreal forests with climate warming and wildfire. This shift in composition has the potential to generate biophysical cooling via increased land surface albedo. Here we use Landsat-derived maps of continuous tree canopy cover and deciduous fractional composition to assess albedo change over recent decades. We find, on average, a small net decrease in deciduous fraction from 2000 to 2015 across boreal North America and from 1992 to 2015 across Canada, despite extensive fire disturbance that locally increased deciduous vegetation. We further find near-neutral net biophysical change in radiative forcing associated with albedo when aggregated across the domain. Thus, while there have been widespread changes in forest composition over the past several decades, the net changes in composition and associated post-fire radiative forcing have not induced systematic negative feedbacks to climate warming over the spatial and temporal scope of our study.

 

Abstract. Tundra environments are experiencing elevated levels of wildfire, and thefrequency is expected to keep increasing due to rapid climate change in theArctic. Tundra wildfires can release globally significant amounts ofgreenhouse gasses that influence the Earth's radiative balance. Here wedevelop a novel method for estimating carbon loss and the resultingradiative forcings of gaseous and aerosol emissions from the 2015 tundrawildfires in the Yukon–Kuskokwim Delta (YKD), Alaska. We paired burn depthmeasurements using two vegetative reference points that survived the fireevent – Sphagnum fuscum and Dicranum spp. – with measurements of local organic matter and soil carbonproperties to estimate total ecosystem organic matter and carbon loss. Weused remotely sensed data on fire severity from Landsat 8 to scale ourmeasured losses to the entire fire-affected area, with an estimated totalloss of 2.04 Tg of organic matter and 0.91 Tg of carbon and an average lossof 3.76 kg m−2 of organic matter and 1.68 kg m−2 of carbon in the2015 YKD wildfires. To demonstrate the impact of these fires on the Earth'sradiation budget, we developed a simple but comprehensive framework toestimate the radiative forcing from Arctic wildfires. We synthesizedexisting research on the lifetime and radiative forcings of gaseous andaerosol emissions of CO2, N2O, CH4, O3 and itsprecursors, and fire aerosols. The model shows a net positive cumulativemean radiative forcing of 3.67 W m−2 using representative concentration pathway (RCP) 4.5 and 3.37 W m−2using RCP 8.5 at 80 years post-fire, which was dominated by CO2emissions. Our results highlight the climate impact of tundra wildfires,which positively reinforce climate warming and increased fire frequencythrough the radiative forcings of their gaseous emissions. 

Abstract. Fire is the dominant disturbance agent in Alaskan and Canadianboreal ecosystems and releases large amounts of carbon into the atmosphere.Burned area and carbon emissions have been increasing with climate change,which have the potential to alter the carbon balance and shift the regionfrom a historic sink to a source. It is therefore critically important totrack the spatiotemporal changes in burned area and fire carbon emissionsover time. Here we developed a new burned-area detection algorithm between2001–2019 across Alaska and Canada at 500 m (meters) resolution thatutilizes finer-scale 30 m Landsat imagery to account for land coverunsuitable for burning. This method strictly balances omission andcommission errors at 500 m to derive accurate landscape- and regional-scaleburned-area estimates. Using this new burned-area product, we developedstatistical models to predict burn depth and carbon combustion for the sameperiod within the NASA Arctic–Boreal Vulnerability Experiment (ABoVE) coreand extended domain. Statistical models were constrained using a database offield observations across the domain and were related to a variety ofresponse variables including remotely sensed indicators of fire severity,fire weather indices, local climate, soils, and topographic indicators. Theburn depth and aboveground combustion models performed best, with poorerperformance for belowground combustion. We estimate 2.37×106 ha (2.37 Mha) burned annually between 2001–2019 over the ABoVE domain (2.87 Mhaacross all of Alaska and Canada), emitting 79.3 ± 27.96 Tg (±1standard deviation) of carbon (C) per year, with a mean combustionrate of 3.13 ± 1.17 kg C m−2. Mean combustion and burn depthdisplayed a general gradient of higher severity in the northwestern portionof the domain to lower severity in the south and east. We also found larger-fire years and later-season burning were generally associated with greatermean combustion. Our estimates are generally consistent with previousefforts to quantify burned area, fire carbon emissions, and their drivers inregions within boreal North America; however, we generally estimate higherburned area and carbon emissions due to our use of Landsat imagery, greateravailability of field observations, and improvements in modeling. The burnedarea and combustion datasets described here (the ABoVE Fire EmissionsDatabase, or ABoVE-FED) can be used for local- to continental-scaleapplications of boreal fire science. 

Abstract Forest characteristics, structure, and dynamics within the North American boreal region are heavily influenced by wildfire intensity, severity, and frequency. Increasing temperatures are likely to result in drier conditions and longer fire seasons, potentially leading to more intense and frequent fires. However, an increase in deciduous forest cover is also predicted across the region, potentially decreasing flammability. In this study, we use an individual tree-based forest model to test bottom-up (i.e. fuels) vs top-down (i.e. climate) controls on fire activity and project future forest and wildfire dynamics. The University of Virginia Forest Model Enhanced is an individual tree-based forest model that has been successfully updated and validated within the North American boreal zone. We updated the model to better characterize fire ignition and behavior in relation to litter and fire weather conditions, allowing for further interactions between vegetation, soils, fire, and climate. Model output following updates showed good agreement with combustion observations at individual sites within boreal Alaska and western Canada. We then applied the updated model at sites within interior Alaska and the Northwest Territories to simulate wildfire and forest response to climate change under moderate (RCP 4.5) and extreme (RCP 8.5) scenarios. Results suggest that changing climate will act to decrease biomass and increase deciduous fraction in many regions of boreal North America. These changes are accompanied by decreases in fire probability and average fire intensity, despite fuel drying, indicating a negative feedback of fuel loading on wildfire. These simulations demonstrate the importance of dynamic fuels and dynamic vegetation in predicting future forest and wildfire conditions. The vegetation and wildfire changes predicted here have implications for large-scale changes in vegetation composition, biomass, and wildfire severity across boreal North America, potentially resulting in further feedbacks to regional and even global climate and carbon cycling. 

Wildfire frequency and extent is increasing throughout the boreal forest-tundra ecotone as climate warms. Understanding the impacts of wildfire throughout this ecotone is required to make predictions of the rate and magnitude of changes in boreal-tundra landcover, its future flammability, and associated feedbacks to the global carbon (C) cycle and climate. We studied 48 sites spanning a gradient from tundra to low-density spruce stands that were burned in an extensive 2013 wildfire on the north slope of the Alaska Range in Denali National Park and Preserve, central Alaska. We assessed wildfire severity and C emissions, and determined the impacts of severity on understory vegetation composition, conifer tree recruitment, and active layer thickness (ALT). We also assessed conifer seed rain and used a seeding experiment to determine factors controlling post-fire tree regeneration. We found that an average of 2.18 ± 1.13 Kg C m -2 was emitted from this fire, almost 95% of which came from burning of the organic soil. On average, burn depth of the organic soil was 10.6 ± 4.5 cm and both burn depth and total C combusted increased with pre-fire conifer density. Sites with higher pre-fire conifer density were also located at warmer and drier landscape positions and associated with increased ALT post-fire, greater changes in pre- and post-fire understory vegetation communities, and higher post-fire boreal tree recruitment. Our seed rain observations and seeding experiment indicate that the recruitment potential of conifer trees is limited by seed availability in this forest-tundra ecotone. We conclude that the expected climate-induced forest infilling (i.e. increased density) at the forest-tundra ecotone could increase fire severity, but this infilling is unlikely to occur without increases in the availability of viable seed.