ANALYSIS AND 3D VISUALISATION OF WILDFIRE USING EARTH OBSERVATION DATA

Introduction

Understanding the pattern and impact of these wildfires is necessary for rebuilding and rehabilitation to occur. 2023 saw an estimated 4 billion dollars’ worth of global economic damage from wildfires, with the cost of rebuilding often being more significant than this. For the 2023 fires in Maui alone, an estimated 5.5 billion dollars is required for rebuilding efforts (PDC, FEMA, 2023). To undertake the process of rebuilding, it is therefore necessary to visually depict the spread and pattern of destruction.

For this purpose, Earth observation offers an invaluable resource for holistic analysis of wildfire indicators before, during, and after an extreme wildfire event. Factors such as burn area (Long Fournel et al., 2014), smoke plumes, wind direction, and active fire points as visible using satellites and sensors and may be measured to better understand the contributing factors and damage from such events. Significant research has gone into modeling fire risk to initiate anticipatory action against the onset of wildfires and improve emergency preparedness. Others have also investigated accurately depicting post-fire impact using remotely sensed attributes.

This current study aims to retrospectively illustrate the pattern of occurrence of an extreme wildfire event in a 3D environment using such accurately measured EO-derived indicators as the Aerosol Optical Depth (AOD) to depict smoke plumes and sub-daily observations of active fire locations from MODIS. Through this, we hope to recreate the commencement and spread of the Maui Wildfires of 2023. For better understanding and relatability, this will be done using advanced 3D modeling software to better support rebuilding and rehabilitation efforts for this and similar fires.

Objectives

This diploma thesis aims to develop and document a workflow for modeling and rendering remote sensing derived wildfire parameters in an advanced 3D modeling platform. Furthermore, the suitability of representing smoke plumes in a 3D model of a wildfire, with satellite-derived aerosol optical depth data will be evaluated. The objective of this study therefore is as follows.

• Develop the workflow for modeling and visualizing real-time and remote-sensing derived parameters in a 3D environment
  for retrospective monitoring of a wildfire event.
• Identify potential issues with applying remote-sensing-derived parameters of a wildfire in a 3D modeling environment.
• Evaluate the viability of using satellite-derived Aerosol Optical Depth in representing smoke plumes in a 3D modeling environment.
• Document processes and workflows developed for easy replication and utilization.

Methodology

Two scales of the fire are modelled in this study. Maui Island and Lahiana. For each scale, the steps taken in the workflow span Data Preprocessing, Model Creation, Animation and Rendering, and Post Processing (Video Editing)

Data Preprocessing

Maui Island Fire Animation

For this level, the datasets required were the 10m DEM, satellite imagery for texture, active fire data, and administrative boundary data.
Issues addressed here include finetuning the Island Boundary data, merging and ectracting the DEM for the terrain background, dealing with insufficient timesteps in active fire data and exploring the limitation of availabe Aerosol data intended for smoke plume depiction

Lahaina Fire and Smoke Animation

For this close-up view of the Lahaina Fire, the datasets used are Google Photorealistic 3D tiles, a coded timeline of the fire as observed from news sources and remote sensing data, and 3D fire and smoke material.
Issues addressed here include timeline extraction and syncing, Google Photorealistic tiles download using the blosm plugin in Blender, damaged/destroyed building footprints extraction and map production

Model Creation, Animation and Rendering

Maui Island Fire Animation

In Blender, the terrain background is created, the active fire point data is imported using a python script. These points are then animated and given a material (shader) influenced by their Fire Radiative Power(FRP) value. The fire points are also duplicated and animated/shaded to correspond with the burn footprints.
Other enviroment elements such as thr light and camera are also set up and animated as required. Rendering is done at 3FPS which is used because of the limited time steps available, as expressed in subchapter 3.2.1 of this study. Therefore 3 frames represent 1 second of the animation.

Lahaina Fire and Smoke Animation

Firstly, the appropriate 3D tiles are downloaded and aligned as required. These tiles are downloaded at varying levels of detail to ensure efficient resource use.
The fire and smoke model is the procedurally created within predefined fire regions. A tool that instances realistic fire/smoke material along the vertices of a bezier curve was used to instance the fire and smoke across the scene, instead of Blender's standard fluid simulations. This was done in the interest of modeling ease and resource conservation. Each created fire region was cduplicated to form its corresponding smoke region. These were then animated (prestart, start, peak, decline, end) to mimick the likely fire progression (as referenced form real-time information).

The fire model created in a field of Lahiana as an example

Predifined regions within which the fire models are created

Post Processing (Video Editing)

The individual files containing the .tiff images were imported into individual Adobe Premiere Pro projects as image sequences. Elements such as Title, Description, Callouts, Time counter Imprint, and Audio were added in this step.

Results

This section discusses the outputs from this study. Five separate video animations from separate camera angles were produced to illustrate the progression of the Maui Island and Lahaina Fires of 2023. They are described below;

Overview Animation of Maui Island Fire

This is the result of the Maui Island Animation. It shows an overview map of the island and visualizes the active fire spots as bright objects animated to appear at the appropriate time stamp and disappear afterward.
The video has 3 sections, the first of which shows the light rise of Maui Island. Here the background is the 10m DEM as terrain, textured with Sentinel 2 cloudless composite data. The title, description, inset map, scalebar, and imprint appear at the beginning and are permanent throughout the video. The animated wind icon is active from the beginning while the active fire points are burning. The second section starts with the transformation of the background to a plain grey textured terrain. This serves as a better backdrop for the active fire spots. As the fire spots appear so do their callouts, and corresponding time stamps. The legend also appears in this section. After the fire spots stop, the burn scars remain and are highlighted in red towards the third section. Finally, line charts showing the impacts (Area burned, Casualties, and Buildings Damaged) are displayed.

Animation of Lahaina Fire: South-western view

This animation shows the initial 30 hours of the fire. Starting from 06:30 hours on August 8th, 2023, to 12:34 hours on August 9th. The camera looks from the southern coast of the island just beyond where the affected buildings start. This captures the entire burn zone and illustrates the fire in Lahaina starting and progressing.
The title, description, time counter, and imprint start at the beginning and are permanent. As the animation progresses, callouts explain the progress of the fire, the area currently burning and highlight specific landmarks. At specific time steps where verification data is available in the form of video snapshots or satellite data, these are also displayed at the corresponding timestamp. The video stills are gotten from NBC News (NBC News, 2023).

Animation of Lahaina fire: Eastern hillside view

This animation also shows the initial 30 hours of the fire. Starting from 06:30 hours on August 8th, 2023, to 12:20 hours on August 9th. The camera looks from the eastern hillside and offers a closer view from the fire ignition point. This view also captures the entire burn zone, albeit with a steady camera movement facing first to the south and then to the north. It shows the same fire progression as in the above animation.
The title, description, time counter, and imprint start at the beginning and are permanent. As the animation progresses, descriptors explain the progress of the fire, and the area currently burning and highlight specific landmarks. The major road and stream in view are also labeled using callouts. The location parameters of the callouts and labels had to be manually animated every few frames because of the dynamic nature of the camera angle.

Animation of Lahaina fire: Harbor View

The purpose of this animation is to focus more on the historical district on the west coast of Lahaina and how the fire spread there. Starting from 02:50 hours and ending at 23:48 on August 8th, 2023.
The camera looks from just above the harbor and zooms in and centers on the Banyan Tree and the Old Lahaina Courthouse at the heart of the historical district. The view then shifts first to the left (towards the south), then to the west (towards the north) to see the fire burning landmarks like 505 Front Street Shopping Mall and Wohing Temple Museum. It shows the same fire progression as in the above animations. The timer gives a view into when this likely occurred in real time.
The title and time counter start at the beginning and are permanent. The description is animated as “text crawl”, such that it scrolls across the bottom of the screen as the video progresses. As the animation progresses, callouts highlight specific landmarks. Similar to the Hillside view animation, the location parameters of the callouts had to be manually animated every few frames because of the dynamic nature of the camera angle.

Lahaina Fire Animation: Aftermath (Damage and Historical Landmarks)

The purpose of this animation is to show the aftermath of the fire particularly in the historical district of Lahaina which was mostly destroyed. The animation is in 2 sections. The first section shows a side-by-side view of a fly-through from the extreme southwestern part of the burn zone in Puamana Community. In this section of the flythrough, damaged and undamaged buildings were highlighted using individually colored points. A legend is made available corresponding to the structures that are within view. The title, description, and imprint are permanently displayed. On the right side is the 3D model output of this study, while on the right is drone flight footage of the actual area captured 4 months after the fire in December 2023 made available on YouTube (Jesse G. Wald, 2023).
The second section of the flythrough starts when footage from YouTube stops. Although there is no longer real-life footage available, the remainder of the flythrough depicts the 3D model of the historical area before the fire as rendered from Blender. The purpose of this section is to showcase the quality and realism of Google 3D photorealistic tiles and to show the eventually damaged historical artifacts like the Honjwanji Mission, The Waiola Church, and prominent structures like the Aina Nalu Condo Complex. In the supporting story map to this study, images of the aftermath of these structures are showcased.
Track points are placed to identify corresponding structures, as represented in the legend. Like the two preceding animations, the location parameters of the points had to be manually animated every few frames because of the dynamic nature of the camera.

Storymap

To orientate potential users of these videos particularly the stakeholders tasked with restoring and rebuiidng after the fire, a supporting storymap was created which includes more details on the aftermath of the fires inlcluding before and after images of prominent landmarks in Lahaina, as well as a map of damaged/destroyed buildings and their respective zones (view below).

Evaluation

This section evaluates the accuracy of this study in modeling the Maui Island Fire of 2023. Here, the resulting 3D animations are assessed based on remote sensing data and the most detailed official timeline so far from the Hawaii Department of the Attorney General(Hawaii DOAG, 2024).

Evaluation was done against three separate datasources.

NASA Thermal Imagery and Shortwave Infrared

The calculated land surface temperature from the Thermal Infrared Sensor (TIRS) aboard Landsat-8 and -9, acquired on August 8, 2023, at 08:25 UTC (Local time: 10:25 pm (22:35) on August 8, 2023) was used in this evaluation. It is available on NASA’s Disaster Mapping Portal (NASA Disaster Portal, 2023). The units for this product are Kelvin (K).
The shortwave infrared imagery of the Maui Island wildfires was also captured by the Operational Land Imager (OLI) on Landsat 8 on Aug. 8, 2023 (time unknown), as seen in Figure 46. Fires are shown in yellow. The shortwave infrared data were overlaid on a natural-color mosaic image based on Landsat 8 observations for added geographic detail. This data source was simultaneously referenced for this evaluation.

The fire intensity in northern Lahaina residential (1) is visible in the surface temperature data, same as for the Lahainaluna area (2) where the fire had initially intensified and is now waning. 3 shows the still high fire intensity in the apartment complex on Aulike Street while (4) shows most of Puamana Community and Front Street still burning intensively. The depiction of the fire progress at the specified point in time is highly analogous to these data sources.

Maxar Satellite Imagery

Maxar’s publicly available 33cm satellite imagery acquired on August 9th at 11:20 local time was visually compared to evaluate the output of this study. Fire cannot be seen in the satellite imagery, but this is not to say that some parts were not still burning. The ability to see smaller fires at this scale, time of day, and sensor angle is limited. However, since this is an optical imagery that depicts what the human eye can see at 33cm spatial resolution, it was deemed adequate for this evaluation.

Judging by the smoke position in the agricultural zone to the north of Lahaina (1), we can tell that a fire is still active albeit on a smaller scale, and that the fires in most of the area have since subsided. Given the natural progression of the fire and the fact that in the north, the fire started much later, it is assumed that the fire in the residential part is still burning (2), even though it is not obvious on the satellite image.

Hawaii Attorney General Progression Timeline

Hawaii state officials released the Lahaina Fire Comprehensive Timeline Report, which is the first phase of an independent analysis conducted by the Fire Safety Research Institute (FSRI), part of UL Research Institutes. The report chronologically details the major events and response efforts related to the catastrophic fire that struck Lahaina, on August 8–9, 2023. Below is an animation presenting the findings of this report. It shows the progression of the fire from verified observations made in real time. Comparing this to the resulting animation from this study gives a good overview of the quality of the fire progression modeling done. This data set was not consulted before this evaluation.

A timestamped comparison is done to this datasoure. Please see the full text for this.

Discussion and Conclusion

While conducting this study into developing a workflow for the retrospective modeling of a wildfire progression, which resulted in the creation of 5 video outputs, each from different views of the fire progression model, some successes and challenges were identified which will now be discussed.

Evaluating Google Photorealistic 3D Tiles

Google Photorealistic 3D tiles are leveraged in this study. It is considered cutting-edge in its availability and quality and saved a significant amount of time in the workflow of this study. Given the lack of an openly available, recently acquired, high resolution Digital Surface Model (DSM) of Lahaina, in the absence of the Google 3D tiles, a workflow would have been required to create models of the area from scratch.
Although there exist city generator plugins for Blender that produce high-quality 3D city models within a few steps, they do not depict real cities. If a real model of Lahaina were to be created, it would have required modeling every building, plant, road, and other environmental features. This would not only have been labor and resource-intensive, but it would also have taken significant time.

However, there are a few challenges with its use as observed below.

• Data is not up to date: As highlighted in section 3.2.5 of this study and evidenced in the Damage and Historical Landmarks animation (section 6.5), the 3D tiles are outdated in some areas.
• Input source and methodology behind its production non-transparent: Although a few input sources were identified in subchapter 3.2.5, sources of tiles and acquisition times for specific areas are not openly available.
• It may be costly for larger projects: Due to the relatively small scale of this study area (7km by 3km), all tiles downloaded used in this study only utilized the 200 free monthly credits, gotten automatically from the creation and use of a standard Google Cloud account, plus $5 worth of credits extra. This may be significantly higher for a larger area as more credits are consumed with each download.
• Material editing can be more flexible: Unlike individually created 3D models of the environment elements, whose structure and material can be easily manipulated. This is not possible for Google 3D tiles. This is understandable, however, since they are created from 3D scans of the area.

Active fire data

This was found to be useful only on a small-scale depiction (zoomed out). Upon zooming into the Lahaina area, the fire points were not useful in explaining the fire’s progression. This is because each point is the center point of a 375m-by-375m pixel within which fire was detected (NASA FIRMS, n.d.). It is very limited in showing how the fire progressed on a submeter level as was analysed in this study. However, real-time videos, images, and commentary from social media and news outlets, verifiable through Google Maps and satellite imagery (building locations identified in updated burn Google Maps) were more useful for piecing together a realistic progression of the fire.

Modelling and Animation

While the background elements such as the terrain and 3D model of the study area are as geographically accurate as can be, the manual effort was made to animate elements like the fire and smoke model as well as the callout/labels.
While the fire/smoke path was manually interpreted from source imagery, it does not describe the exact path that the real fire took. This is because no data source currently exists, that constantly monitors such fires at the right spatial and temporal scale/resolution. Only snapshots in time are available and were leveraged and interpolated to represent a progressive spread of the fire. Therefore, a level of abstraction is considered. Furthermorem the main software used in this study, Blender for modeling and Adobe Premier Pro for Video editing had to be learned from scratch to carry out this study. Therefore, while this was highly time and resource-intensive, requiring significant effort, there is still significant room for improvement.

Upon critical review of the resulting animation, some findings were made. To create a reasonably concise video output, the model was timed at 2 frames per minute and rendered at 30 frames per second. Therefore, the resulting animation was reasonably fast-paced. While this is justified given the speed at which the fire spread, certain important information can be missed. However, being an animation, it offers opportunities for review that static media does not, such as the opportunity to slow down, fast forward, or rewatch. Also, animating the callouts/labels reduces the overall map load and the possibility of sensory overload from too many details shown at once. Throughout the animation, it is ensured that only what is viable within frame and at the appropriate time stamp is displayed

Possible Future Workflow

An additional future workflow could be leveraging wildfire modeling algorithms to statistically model the fire and then compare it in real-time, to test its efficiency and accuracy. Also, a further workflow may be developed that shows the real-time damage of 3D models of the structures as the fire consumes them. This may be done when more improvements are made to eth 3D photorealistic tiles that allow individual styling and material assignment to the buildings and structures in the model.

Limitations

Even though fires still occurred on subsequent days until August 15th, only the fires detected by the 375m spatial and limited temporal resolution VIRRS sensor could be visualized. This is a limitation of available remote sensing data to properly analyze the true extent of a fire.
The Aerosol Optical Depth (AOD) data which was originally intended to be used in modeling smoke plumes in this study could not be used also due to limitations in remote sensing data resolution.

Disclaimer

NASA Active fire points downloaded for this study contain some points observed over the sea. Because this is highly improbable, these points were manually moved to land to align with existing burn scars as observed from Maxar’s 33cm satellite imagery.