Abstract
Predicting forest fires is a critical task for environmental preservation and economic stability, particularly in the peatlands of Southeast Asia. Traditional prediction systems rely on expensive ground-based weather stations, which are often unavailable in developing nations. This paper introduces Agni, a novel neural network-based approach that utilizes exclusively public satellite imagery. By combining convolutional layers with LSTMs and a specialized loss function, Agni achieves an Area Under the Curve (AUC) exceeding 0.81, significantly outperforming traditional spatial logistic regression baselines.
Introduction
Forest fires in Southeast Asia, especially within carbon-rich peatlands, pose severe environmental and economic threats. Gold-standard prediction systems, such as the Canadian Forest Fire Danger Rating System, require extensive networks of ground-based sensors—often exceeding 750 stations. For developing countries like Indonesia, the cost of installing and maintaining such hardware is prohibitive. Furthermore, traditional mathematical models often fail to capture the complex, non-linear relationships inherent in fire risk factors. This research explores whether a reliable, low-cost prediction system can be built using only remote sensing data.
Methodology
The proposed system, Agni, employs a architecture designed to handle the spatial and temporal complexities of satellite data.
Data Representation via Histograms
To manage the massive storage requirements of high-resolution satellite imagery, 8km x 8km image slices are converted into histograms. Each of the eight spectral bands is represented by a 32-bin histogram of pixel values. This approach preserves the spectral distribution while drastically reducing the data size, enabling more efficient processing.
Hybrid CNN-LSTM Architecture
The model utilizes a hybrid architecture:
- Conv2D Layers: Process the histograms for each timestep to learn relationships between different spectral bands.
- LSTM Layers: Capture temporal patterns over a one-year sliding window of historical data.
- Dense Layers: Output a final confidence score for fire presence five weeks into the future.
Custom Loss Function
Fire hotspots are rare, appearing in less than 1% of the dataset. To prevent the model from defaulting to a "no fire" prediction, a custom loss function was implemented. It applies a high penalty when the model predicts a low probability for an actual fire event, forcing the model to prioritize sensitivity.
Results and Discussion
Agni consistently achieved an AUC between 0.81 and 0.85 across various evaluation periods in 2019. This performance is a marked improvement over spatial logistic regression baselines, which did not exceed an AUC of 0.70.
Robustness to "Broken" Data
A significant finding was the model's robustness to Landsat 7 hardware failures. Despite the Scan-Line Calibrator (SLC) failure resulting in missing data strips, the neural network learned directly from the incomplete imagery without requiring complex statistical imputation.
Efficiency and Temporal Depth
The use of histograms allowed the researchers to bypass the need for massive GPU clusters. Additionally, while a full year of history is optimal, the model maintained an AUC > 0.81 with only three months of data, suggesting high potential for rapid deployment in new regions.
Conclusion
Agni demonstrates that high-accuracy forest fire prediction is possible using only remote sensing data and machine learning. By addressing the data imbalance and high computational costs of traditional methods, this approach offers a viable, low-cost solution for fire monitoring in developing nations. Future work will focus on integrating additional data sources, such as soil moisture and human activity, to further refine predictive accuracy.
Summary updated