Abstract: Over the last few decades, deforestation and climate change have caused an increasing number of forest fires. In Southeast Asia, Indonesia has been the most affected country by tropical peatland forest fires. These fires have a significant impact on the climate, resulting in extensive health, social, and economic issues. Existing forest fire prediction systems, such as the Canadian Forest Fire Danger Rating System, are based on handcrafted features and require the installation and maintenance of expensive ground instruments, which can be a challenge for developing countries. We propose a novel, cost-effective, machine-learning-based approach named Agni that uses remote sensing data to predict forest fires in Indonesia. Our prediction model achieves more than 0.81 area under the receiver operator characteristic (ROC) curve, performing significantly better than the baseline approach. The results support our claim that machine-learning-based approaches can lead to reliable and cost-effective forest fire prediction systems.
Introduction
Peatland forest fires in Southeast Asia, particularly in Indonesia, represent a massive environmental and economic burden. These fires release gigatons of stored carbon and create hazardous haze affecting millions of people across the region. Traditional prediction systems like the Canadian Forest Fire Danger Rating System (CFFDRS) are difficult to implement in developing nations due to high costs and technical complexity. They require dense networks of ground-based weather stations and rely on mathematical equations that often fail to capture the non-linear relationships between weather, soil, and fire behavior.
To address these challenges, we developed Agni, a deep learning architecture designed to provide a "satellite-only" prediction model that is both accurate and affordable.
Methodology
The Agni pipeline utilizes a CNN-LSTM neural network to process spatial and temporal data. The model employs a 2D Convolutional Neural Network (CNN) to extract spatial features from satellite spectral bands, which are then passed to a Long Short-Term Memory (LSTM) network to capture temporal patterns over a one-year historical window.
Data Representation and Efficiency
To manage the massive storage requirements of high-resolution satellite imagery—which would exceed 1.6 TB for the study area—we implemented a histogram-based downsampling method. Each spectral band is converted into a 32-bin histogram, providing a storage-efficient representation of an 8km x 8km area while preserving the distribution of spectral signatures.
Custom Loss Function
Fire events are rare, representing less than 1% of the dataset. Standard loss functions like Mean Squared Error (MSE) often lead models to simply predict "no fire" to achieve high accuracy. We designed a custom loss function that applies a high penalty for false negatives (mispredicting actual fire events), forcing the model to prioritize the detection of hotspots. The model is trained to predict the presence of a fire 5 weeks into the future, providing a critical lead time for authorities.
Datasets
The study utilizes two primary open-source datasets: NASA's FIRMS (Fire Information for Resource Management System) for ground truth labels and USGS Landsat 7 for input features. Landsat 7 provides eight spectral bands, as detailed below:
A significant challenge with Landsat 7 data is the "SLC-off" issue, where a hardware failure causes missing data strips in images. Agni was designed to be robust against these artifacts, learning to extract meaningful patterns despite the missing data.
Key Results
Agni significantly outperformed the spatial logistic regression baseline, which never exceeded an AUC of 0.70. In contrast, Agni achieved an Area Under the ROC Curve (AUC) between 0.81 and 0.85 across various test months in 2019.
Our ablation studies confirmed that the custom loss function was essential; performance dropped below 0.75 AUC when using standard MSE. Furthermore, the model proved robust to data reduction; even when the historical look-back window was reduced from 12 months to just 3 months, the AUC remained above 0.81.
Conclusion and Future Work
The results suggest that remote sensing data can act as a viable proxy for ground-based sensors, offering a game-changing solution for environmental monitoring in resource-constrained regions. Future work will focus on integrating additional data sources, such as precipitation and soil moisture, and improving computational efficiency by addressing batch size constraints caused by varying satellite data dimensions.
Summary updated