A Spatio-Temporal Neural Network Forecasting Approach for Emulation of Firefront Models
Highlights
-
Figures 2-5 are pretty cool
-
Definitely a paper that could use a lot more work. They have an interesting direction, but there's way too much lacking. I would give it a strong reject.
Summary
Computational simulations are typically used to model wildfire spread under conditions such as terrain, fuel type, and weather. However, small perturbations mean that one must run a large set of computationally expensive simulations to quantify uncertainty. Model emulation provides an alternative, more efficient representation using machine learning when used in ensemble to quantify uncertainty (Annoyingly, the authors don't actually do this though). This paper introduces an framework for wildfire model emulation. The authors show good agreement between simulated and their emulated firefronts on synthetic dataset.
Key Contributions
-
A machine learning based model emulation framework for fire spread models.
-
Claimed data augmentations, but they aren't significant contributions (see weaknesses for details).
Strengths
-
The authors acknowledge one of the biggest flaws of their system, which is that minor differences between emulation and simulation grow more exaggerated over time. Thus, using emulation may only be correct up to a certain timeframe.
-
Figures 2-5 are actually pretty interesting. I especially like the difference maps of Figure 3 and figure 5. The use of a river and more complex topology in figures 2 and 4 are also a nice touch.
Weaknesses / Questions
-
I was a little skeptical of why they chose their model at first, but the argument for linear transformations to preserve the temporal relationships is fine. U-Net with skip connections to learn just the residual is fine too, given they are approaching this as an iterative prediction task with small, incremental changes. I wish there were comparisons between nonlinear autoencoders and at least an ablation study somewhere.
-
Their emulation iterates whenever new weather information is input, which is every 30 minutes. They use 4 time slices, so updates are every 7.5 minutes. Part of their justification is centered around traditional simulation models being slow and computationally heavy. It's hard to see clear justification of this when there isn't even a solid evaluation for comparing compute times. They could obviously show it (and their abstract even points to it) if they used ensemble emulation as that can clearly show better performance than ensemble simulation after N simulations.
-
They further present their data augmentation method as novel, however, the novelty is overstated. The cropping and patching method along with rotation, reflection, and transposition is standard practice in computer vision. Targeting around the fire parameter, arguably the most unique part of their approach, aligns with the standard CV problem of selecting "regions of interest". Here, the region of interest is the fire parameter. Perhaps if they targetted the problem of adapting the "region of interest" issue to the domain-specific fire parameter problem exclusively, that would have landed better than trying to claim a bunch of standard practice methods as novel and getting away with it because the reviewers have probably never read an article in CVPR.
-
The writing often switches from lackluster simple writing to overly convoluted writing that really should be more concise. There could be a little more polish to tie the different writing styles together.
-
The technique is only evaluated against simulated data. The paper would have been stronger with at least one real fire (or at least an attempt at it).
Related Work
- SPARK Fire simulation platform.