Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Designing a Classifier for Active Fire Detection from Multispectral Satellite Imagery Using Neural Architecture Search

Amber Cassimon, Phil Reiter, Siegfried Mercelis, Kevin Mets

TL;DR

This work demonstrates a reinforcement learning–based Neural Architecture Search workflow to automatically design compact neural networks for on-board active fire detection from multispectral Sentinel-2 imagery under nanosatellite power constraints. A performance predictor (gradient boosted trees) estimates post-quantization F1 from architectural features, enabling rapid evaluation within an RL loop that optimizes both accuracy and model size. The best-found architecture contains 1,716 parameters, achieves a median post-quantization F1 of 99.884% on validation, and runs on a Google Coral Micro Dev Board with about 0.984 ms latency and ~780 mW power, confirming feasibility for low-power on-board processing. The study highlights the potential and challenges of NAS in space systems, including predictor-induced adversarial samples and the importance of hardware-aware utility shaping for multi-objective optimization.

Abstract

This paper showcases the use of a reinforcement learning-based Neural Architecture Search (NAS) agent to design a small neural network to perform active fire detection on multispectral satellite imagery. Specifically, we aim to design a neural network that can determine if a single multispectral pixel is a part of a fire, and do so within the constraints of a Low Earth Orbit (LEO) nanosatellite with a limited power budget, to facilitate on-board processing of sensor data. In order to use reinforcement learning, a reward function is needed. We supply this reward function in the shape of a regression model that predicts the F1 score obtained by a particular architecture, following quantization to INT8 precision, from purely architectural features. This model is trained by collecting a random sample of neural network architectures, training these architectures, and collecting their classification performance statistics. Besides the F1 score, we also include the total number of trainable parameters in our reward function to limit the size of the designed model and ensure it fits within the resource constraints imposed by nanosatellite platforms. Finally, we deployed the best neural network to the Google Coral Micro Dev Board and evaluated its inference latency and power consumption. This neural network consists of 1,716 trainable parameters, takes on average 984μs to inference, and consumes around 800mW to perform inference. These results show that our reinforcement learning-based NAS approach can be successfully applied to novel problems not tackled before.

Designing a Classifier for Active Fire Detection from Multispectral Satellite Imagery Using Neural Architecture Search

TL;DR

This work demonstrates a reinforcement learning–based Neural Architecture Search workflow to automatically design compact neural networks for on-board active fire detection from multispectral Sentinel-2 imagery under nanosatellite power constraints. A performance predictor (gradient boosted trees) estimates post-quantization F1 from architectural features, enabling rapid evaluation within an RL loop that optimizes both accuracy and model size. The best-found architecture contains 1,716 parameters, achieves a median post-quantization F1 of 99.884% on validation, and runs on a Google Coral Micro Dev Board with about 0.984 ms latency and ~780 mW power, confirming feasibility for low-power on-board processing. The study highlights the potential and challenges of NAS in space systems, including predictor-induced adversarial samples and the importance of hardware-aware utility shaping for multi-objective optimization.

Abstract

This paper showcases the use of a reinforcement learning-based Neural Architecture Search (NAS) agent to design a small neural network to perform active fire detection on multispectral satellite imagery. Specifically, we aim to design a neural network that can determine if a single multispectral pixel is a part of a fire, and do so within the constraints of a Low Earth Orbit (LEO) nanosatellite with a limited power budget, to facilitate on-board processing of sensor data. In order to use reinforcement learning, a reward function is needed. We supply this reward function in the shape of a regression model that predicts the F1 score obtained by a particular architecture, following quantization to INT8 precision, from purely architectural features. This model is trained by collecting a random sample of neural network architectures, training these architectures, and collecting their classification performance statistics. Besides the F1 score, we also include the total number of trainable parameters in our reward function to limit the size of the designed model and ensure it fits within the resource constraints imposed by nanosatellite platforms. Finally, we deployed the best neural network to the Google Coral Micro Dev Board and evaluated its inference latency and power consumption. This neural network consists of 1,716 trainable parameters, takes on average 984μs to inference, and consumes around 800mW to perform inference. These results show that our reinforcement learning-based NAS approach can be successfully applied to novel problems not tackled before.
Paper Structure (16 sections, 1 equation, 19 figures, 2 tables)

This paper contains 16 sections, 1 equation, 19 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. NAS
  4. Multispectral Image Processing
  5. Active Fire Detection
  6. Methods
  7. Target Task
  8. Performance Data Gathering
  9. Performance Predictor Training
  10. NAS Agent Training
  11. Experiments
  12. Statistical Analysis of Performance Data
  13. Performance Predictor
  14. NAS Agent
  15. Deployment
  16. ...and 1 more sections

Figures (19)

  • Figure 1: A schematic overview of the procedure used to generate the dataset.
  • Figure 2: A histogram of the geographic distribution of our dataset. Note the logarithmic Y-axis.
  • Figure 3: An example network involving multiple inputs to the output node.
  • Figure 4: An example network involving multiple inputs to an intermediate node.
  • Figure 5: A histogram showing the obtained post-quantization F1 scores on the validation set. The median is indicated with a solid black vertical line.
  • ...and 14 more figures