Progress 09/15/22 to 09/14/23

Outputs
Target Audience:Our target audiences during the first year of the grant include the following organizations: Agencies NASA, USDA, USFS, USGS, NOAA State/Local NY DEC, CalFire, NYC, GA State Forestry, ND State Forestry, AZ State Forestry, MD State Forestry Insurance Chubb Insurance, TravelersInsurance Utilities Duke Energy, Excel, EEI (Trade Association for Utilities) Changes/Problems:Given our work so far, our key successes are a fast, accurate, and stable platform; demonstrated technical and scientific capability in ML/AI, platform engineering, GIS technology, and fire behavior fundamentals; and an ability to pivot across facets of the fire space. However, we have not seen the level of client engagement expected. Our assessment here is the demand for the SDI and PCL algorithms is lower than expected. Our client interactions have indicated that they put more focus on standard fire behavior metrics (rate of spread, heat per unit area, etc.) and their own intuition and experience for setting fire lines based on that information. We believe we are in a good position to pivot to a product that meets the needs of our stakeholders. We are looking for additional field partners and clients to ensure that we can work with to build a meaningful product. Additionally, we have made corresponding changes in our product and research roadmap. First, we have shifted our platform to give more focus to standard fire behavior metrics. NASA's SMD Wildfire Stakeholder Engagement Workshop (2022) identified a set of key firetech needs for pre-fire, active fire and post fire activities, including a need for more high fidelity fire behavior models. Research in these areas is active and our collaborating partners at the state level have echoed these needs. Traditional fire spread models used in tools like FlamMap are not tuned well for regions outside of the western U.S. or particular use cases (prescribed burn, active fire, etc.). This is an active concern and impacts trust level in fire behavior modeling results which could be used to extend controlled burn opportunities. We propose layering on predictive analytics on top of traditional spread models, essentially leveraging AI to tune traditional fire behavior models for particular regions and use cases. The Workshop also identified the need for up-to-date, accurate fuel maps aimed at facilitating risk assessment and planning, active fire response and fire behavior modeling, and damage assessment. Precision in real-time (RT) mapping for fuel models, loadings, and moisture can help planners handle both RT wildfires and prescribed burns. RT maps are also needed to efficiently address change monitoring concerns for issues such as insect infestations, tree stand monitoring, invasive species detection, and forest composition change. Finally, existing public landscape maps are insufficient in their largely agricultural areas as they use native fuel models and fire risk shifts based on the crop and its lifecycle stage. We are proposing a turnkey solution for the acquisition of custom, high-resolution, near RT landscape maps for fuel models, loadings, and moisture. These fuel maps can be used for forest ecology needs, layered into public LANDFIRE data for improved general maps, and used in FlamMap for better fire behavior modeling. To date and as noted here, we are near completion on a Land Mapping pilot that classifies 20+ crops and land types at over 80% accuracy. Per our review of the literature, we believe this to beat existing models which tend to cover smaller areas and constrain the problem to a smaller set of classes. We plan to improve on our algorithm and expand upon it to show generalizability over geography and time. At this point, we know of no end-to-end solution for pulling remote sensed imagery, processing it, and classifying the resulting image. We have developed this in conjunction with a client partner and plan more customer discovery along these lines to ensure we are building an in-demand product. We have also put development towards our Wildfire Explorer to share up-to-date wildfire data for use by the public. It provides top level fire statistics and details active large wild and prescribed fires. This tool is intended for public use and to increase Cornea's visibility to the public. It will be expanded to showcase Cornea's predictive capabilities to generate insights about where wildfires might go and how they might behave under different conditions. An alpha launch is planned for this spring, with a wider release expected for the fire season. Additionally, based on this turn, the last major change is that we have decided to make a shift away from developing a simulation infrastructure for the time being, as noted in the table above. Our current execution infrastructure can execute most desired artifacts for fire behavior within 30 minutes and allows for custom weather and landscape, by far beating most existing options. A larger simulation infrastructure is a big undertaking, and we made the decision to hold on this pending client demand. What opportunities for training and professional development has the project provided?The buildout of the Beacon platform provided opportunities for development of our engineering capabilities. We developed expertise in efficient and scalable ML, employing efficient packages for distance calculations (SciPy pdist, SciPy KDTree, PyKDTree, BallTree), handling sparse matrices, and distributed computation (Ray.io). We also developed a core competency in GIS platform design and development. These skills can be leveraged across any large scale ML and GIS initiatives. With the continued PCL development and the work on the Land Mapping Pilot, we advanced our breadth of ML techniques (DL and CNNs, Linear SVMs), and developed additional competencies in satellites (public satellite instrumentation, operations, data retrieval, image processing), image classification (fuel models, crop detection, object detection), and relevant Python packages (polars for large data frames, GDAL for image processing). How have the results been disseminated to communities of interest?The Beacon platform is Cornea's flagship product. We share the details with customers and potential clients on an ongoing basis through sales calls, events, and product deployments. We have a publication in progress for the land mapping pilot, as this represents novel research advancing the quality of image classification in the area of land mapping. We also have grant applications in progress for further development. Additionally, we are planning analpha release of Cornea's Wildfire Explorer in the early spring, with a wider public release for the fire season. What do you plan to do during the next reporting period to accomplish the goals?As noted in the table above, we have several projects planned to finish out our SBIR Phase II goals: Metrics: Proposal in progress to use ML/AI with traditional fire behavior models to tune for regions and use cases. Metrics: Continued development in our near real-time land mapping (fuel models, moisture, agriculture) algorithm and inclusion into Beacon. Data and UI: Addition of relevant data layers and customizable inputs for landscape, roads/trails, and water into Beacon. Data and UI: Further development of the Wildfire Explorer to showcase Cornea's predictive capabilities to generate insights about where wildfires might go and how they might behave under different conditions. All of this is pending client traction and demand.

Impacts
What was accomplished under these goals? Objective Status Detail Metrics: Accuracy and generalizability of PCLS Complete Burn probability, crown fire and wildfire potential features added for PCL AUC of 84% (up from 81%) PCL model generalizability proven across regions in Great Basin, Nor Cal, Idaho, Arizona and Central Washington (grass, shrub, timber, forest, etc.) Metrics: Fire behavior metric inclusion (Expand metrics/models based on client demand) Complete & Ongoing Standard fire behavior metrics based on Flammap included in Beacon Proposal in progress to use ML/AI with traditional spread models to tune for regions and use cases Metrics: Land Mapping (Expand metrics/models based on client demand) Pilot in Dev Real-time land mapping (20+ crops and land types) using satellite imagery in development for North Dakota Proposal in progress to expand to fuel models and moisture Metrics: Shared feature and modeling library In Progress Internal libraries developed for features, modeling and evaluation per application Development ongoing per application Execution Infra: Optimization of execution time Complete Current product can execute most desired artifacts within 30 minutes and allows for custom weather and landscape. Execution Infra: Infra for large scale simulations TBD Hold for client demand. Data and UI: Data Warehouse In Progress Resource Catalog of data sources developed with updates ongoing for weather, land cover, satellites, fire occurrence, fire behavior, fire risk and more Data warehouse in place for fire behavior modeling needs (landscape, roads, water, etc.) and resulting artifacts, as well as GIS data layers Data and UI: Beacon - GIS platform for predictive analytics (Sophisticated and complete UI) Complete & Ongoing Initial Beacon application completed and in production. Iteration ongoing to include data layers in GIS and more customizable inputs Data and UI: Wildfire Explorer (Sophisticated and complete UI) Complete & Ongoing Dashboard and map for ongoing wild and prescribed fires as a public information and marketing tool ready for alpha launch Additional work planned to detail fires and predicted behavior Detailed Description of Progress Towards Technical Objectives: 1. Metrics: Accuracy and generalizability of PCLs For this objective, we aimed to complete the work we started in our Phase I grant. Namely, conduct SDI/PCL metric iteration to improve accuracy with a target of 85% AUC over a generalized US landscape. Original models focused on an area dominated by grass and shrub fuels, we aimed to show that our model is generalizable across other environments. The addition of features to the model based on annualized burn probability (FSPro) product such as burn probability, crown fire potential, and large wildfire potential brought the PCL AUC in the original Great Basin area from 80.8% -> 83.8%. We added areas with forest fuel models: NorCal, Idaho, Central Washington, Arizona. The below tables break down model performance by flame length class (very low > extreme), fuel type (shrub, grass, timber, non-burnable), climate type (dry/humid). In general, we see consistent performance across these factors. Given our expansion beyond SDI/PCL models, we consider this objective complete. Region AUC Great Basin 83.75% NorCal 77.39% Idaho 81.31% Central WA 77.49% Arizona 87.29% Table. PCL generalizability across landscapes Region Fuel Type % Of Total AUC NorCal Nonburnable 11.9% 89.61% NorCal Grass 23.7% 77.83% NorCal Grass + Shrub 17.6% 73.07% NorCal Shrub 10.8% 68.04% NorCal Timber Shrub 18.7% 80.16% NorCal Timber Litter 17.2% 73.34% Table. PCL generalizability across fuel types. Region Flame Length % Of Total AUC Arizona very low 15.64% 88.07% Arizona low 27.64% 86.02% Arizona moderate 50.69% 86.47% Arizona high 5.14% 84.12% Arizona very high 0.86% 86.85% Arizona extreme 0.03% 62.21% Table. PCL generalizability across flame lengths. 2. Metrics: Fire behavior metric inclusion (Expand metrics/models based on client demand) A second objective within our metrics objective was to improve and expand to predictive metrics for ignition likelihood, fire spread, fire perimeter, and risk analytics through additional/enhanced inputs and evaluation data and with improved fire modeling features and simulations. We have added standard fire behavior metrics from Flammap to our Beacon platform. These include rate of spread, heat per unit area, and flame length, with additional metrics to be determined. 3. Metrics: Land Mapping (Expand metrics/models based on client demand) Our second objective within Metrics is to expand on predictive metrics and risk analytics. We are near completion on a Land Mapping pilot which classifies 20+ crops and land types at over 80% accuracy. We are training on a dataset of 6M pixels of 30x30m resolution coming from the Landsat 9, Sentinel-1 and Sentinel-2 satellites, with a complementary test data set of 4M pixels. The figures below show satellite image data from Sentinel-1 and Landsat-9. We have trained the model using standard machine learning algorithms such as Random Forest, Extreme Gradient Boosting and Linear SVM. Figure. Landsat-9 Satellite image swaths, overlayed onto geo. Figure. Sentinel-1 Satellite images, before and after pre-processing. 4. Metrics: Shared feature and modeling library A third goal within our Metrics objective is to develop a shared feature and modeling library, and evaluation frameworks to support the efficient development and productionization. Accordingly, we have developed Internal libraries for features, modeling and evaluation per application. Work in this area is ongoing. 5. Execution Infra: Optimization of execution time Here, we aim to achieve a target service level agreement (SLA) of 30 minutes generation time per metric for existing and upcoming metrics. The current product can execute most desired artifacts within 30 minutes and allows for custom weather and landscape, so we regard this objective as complete. 6. Data and UI: Data Warehouse We aim to develop a sophisticated and complete data warehouse to support fire prediction and analysis. We have developed an internal data warehouse to support our fire behavior modeling needs and current GIS data layers, as well as capture resulting artifacts. This includes data sources for historical fires and perimeters, roads/transportation, rivers/lakes, fire scars and disturbances, and fuels. We have also developed a Resource Catalog with over 80 data sources to track the plethora of relevant external data sources for weather, land cover, satellites and other remote-sensed data, fire occurrence, fire behavior, fire risk, fire/energy behavior, and other general tools. 8. Data and UI: Beacon - GIS platform for predictive analytics (Sophisticated, complete UI) A key aspect of this objective is a sophisticated and complete UI and platform to support fire prediction and analysis. The Initial Beacon application is completed and in production, with iteration ongoing to include data layers in GIS and more customizable inputs. Highlights of the interface include: an expansive set of data overlays for in-depth analysis including our best-in-class SDI and PCL algorithms; customizable inputs for landscape, weather, roads/trails, water to capture changing real-time conditions; a web-based platform for tablets, mobile, etc.; and a flexible microservices architecture for iterative and efficient continued development. The below figure shows the SDI/PCL platform view. 9. Data and UI: Wildfire Explorer (Sophisticated, complete UI) As part of our goal to provide a UI to support fire prediction and analysis, we have built an active fire dashboard to support public interest and inquiries into wildfires. Our Wildfire Explorer provides a map and detail of active wild and prescribed fire activity, as well as top-line historical statistics. The below figure shows the interface layout. ?

Publications