better deep spread prediction - uploads-ssl.webflow.com

6
Better Deep Spread Prediction with Fion.tech Wildfire modeling can be broadly delineated into three principal categories: risk analysis, detection, and spread prediction. A wonderful recent review of current methods is available in [1]. Here we would like to focus on Fion’s approach and how it differs from the state of the art (SotA). Spread Prediction Recent deep progress [2] produced SotA spread prediction with relatively few input features, comparatively little data, and a strictly downsampling design. We aim to improve on that, the first (and currently only) application of deep convolutional neural networks to spread prediction, by incorporating a few key improvements: More sources of data, A superior segmentation-focused architecture, and Useful training regularizers Architecture We propose FireFlowNet, a multi-headed variant of the U-Net segmentation architecture (Figure 1): Fig. 1: FireFlowNet, a multi-headed U-Net predicting new burned area (′), optical flow (), and new fire center of mass (′). The classification heads share a convolutional body.

Upload: others

Post on 05-Dec-2021

3 views

Category:

Documents


0 download

TRANSCRIPT

Better Deep Spread Prediction with Fion.tech Wildfire modeling can be broadly delineated into three principal categories: risk analysis, detection, and spread prediction. A wonderful recent review of current methods is available in [1]. Here we would like to focus on Fion’s approach and how it differs from the state of the art (SotA).

Spread Prediction Recent deep progress [2] produced SotA spread prediction with relatively few input features, comparatively little data, and a strictly downsampling design. We aim to improve on that, the first (and currently only) application of deep convolutional neural networks to spread prediction, by incorporating a few key improvements:

• More sources of data, • A superior segmentation-focused architecture, and • Useful training regularizers

Architecture We propose FireFlowNet, a multi-headed variant of the U-Net segmentation architecture (Figure 1):

Fig. 1: FireFlowNet, a multi-headed U-Net predicting new burned area (𝐴′), optical flow (𝐹),

and new fire center of mass (𝐶𝑀′). The classification heads share a convolutional body.

FireFlowNet utilizes multiple classification heads to predict new burned area, optical flow, and new fire center of mass. Though only new fire area is of interest during inference, the other classification heads serve as regularizers during training: as the heads share a covolutional body, the weights learned must enable prediction of all three quantities. The goal is to encourage the network to learn something closer to the dynamics of the system in time. The approach in [2] was to slide a downsampling classification network over an image of a fire and predict the burn probability of a single pixel per inference. A U-Net architecture, in contrast, can predict burn probabilities of the entire input window at once. The task of predicting new burn area is intuitively one of semantic segmentation, and a U-Net is thus a more appropriate architectural choice than a binary classification network. In the future, we can move from U-Nets to other multi-resolution segmentation architectures, e.g. HRNet or its variants. We could also incorporate more explicit time-series / dynamical-systems modeling into the architecture, leveraging recent deep advancements in this field [3].

Data Input Features Fion pulls from a variety of data sources:

• Vegetation and moisture indices (e.g. NDVI, NDMI) • Topography (elevation and elevation gradients) • Imagery (RGB bands) • Current fire boundaries • Wind (interpolated from NOAA station data) • Other weather data (barometric pressure, temperature, etc.)

Fire boundaries are currently sourced from GeoMAC, which aggregates fire boundary information from a variety of sources; in the future we plan to derive our own boundaries directly from satelite data (Sentinel-2, VIIRS, LANDSAT-8) through deep learning. Geospatial indices, topography, and imagery all come from atmospherically corrected Sentinel-2 bands. We leverage some human-engineered features (NDVI, NDMI, etc.) as a way of delimiting the model’s task, but in the future will train model variants purely on raw satelite bands.

Figure 2: Example Sentinel-2 atmospherically corrected input features. Clockwise from top-left: RGB (imagery), elevation, elevation gradients (in 𝑢 and 𝑣),

NDMI (moisture index), and NDVI (vegetation index).

Output Labels Spread prediction can be tackled multiple ways: predicting a binary mask or contour of new total burned area, for example, or predicting a binary mask of just the new burn area. We choose the latter to discourage simple heuristics while enabling useful heatmap visualizations. Example labels are shown in Figure 3.

(a) (b) Figure 3: Visual overlays of training labels. (a) New burn area highlighted in orange; fire center of mass (dimmer orange) can be seen to shift slightly south. (b) Optical flow overlay, indicating

direction and speed of spread. Each pair of burn areas for a given fire can be used as an input-output pair. Crops of a specified size can be sampled randomly from the moving fire boundary, yielding examples for the training process. This can be done for historical fires ranging from 2010 to 2019. It’s important to give the model access to as many potentially relevant features as possible, even when the exact nature or degree of relevance is unclear. After all, deep learning is successful because it is successful at finding correlations we cannot.

Testing Though we designed our approach to improve upon recent work, we are still in the process of validating it. Metrics Mean intersection-over-union (mIOU) is the most common measure of semantic-segmentation accuracy, but in the context of wildfire spread it is prudent to include others, including false-positive and false-negative rates. As the model outputs a heatmap of variable confidence, it is possible false-positives could be mitigated by assuming fires will spread even to areas with a low-confidence burn prediction. Interpretability Neural networks are not often interpretable, and this becomes an issue in contexts where important (or even life/death) decisions are made. Trained models will be subject to various interpretability studies, including occlusion analysis, to glean some degree of insight into the learned “thought process”. Center-of-mass and optical flow predictions could aid in this as well. Historical vs Live Data Whenever training on one distribution of data and testing on another, models run the risk of floundering under covariate shift. There is significant work done every month on the capacity of deep models to generalize to unseen data, but it often comes down to how different the unseen data distribution is from the training distribution. In many cases they are similar, but this cannot be guaranteed. To wit: it’s possible some slight degree of covariate shift exists between fires before 2020 and those current happening. However, it is not clear this will significantly impact model performance. The issue could potentially be tackled via the introduction of mixed-in synthetic data configured to better model the dynamics of 2020 wildfire, but this is probably not necessary.

Bibliography [1] Jain, P. et al. “A review of machine learning applications in wildfire science and

management.” ArXiv abs/2003.00646 (2020). [2] Radke, et al. “FireCast: Leveraging Deep Learning to Predict Wildfire Spread”. Proceedings

of the Twenty-Eighth International Joint Conference on Artificial Intelligence, Main track (2019). Pages 4575-4581. https://doi.org/10.24963/ijcai.2019/636

[3] Vlachas, Pantelis R. et al. “Backpropagation algorithms and Reservoir Computing in

Recurrent Neural Networks for the forecasting of complex spatiotemporal dynamics.” Neural networks : the official journal of the International Neural Network Society 126 (2020): 191-217 .