Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Global versus Local: Evaluating AlexNet Architectures for Tropical Cyclone Intensity Estimation

Vikas Dwivedi

TL;DR

This work addresses tropical cyclone intensity estimation from visible satellite imagery by introducing two AlexNet-based ensembles: a global model trained on all data and a distributed model with local experts per Saffir-Simpson category. The global AlexNet acts as a gating mechanism for the distributed ensemble, and their predictions are combined to estimate wind speed. Empirical results show both ensembles surpass the public Deepti benchmark, achieving RMSEs of $RMSE = 9.03$ knots and $RMSE = 9.30$ knots respectively, compared to $RMSE = 13.62$ knots for Deepti, with Grad-CAM analyses providing explainability by highlighting patterns related to Dvorak’s cloud features. The study highlights the interpretable nature of the models and suggests future work on deeper architectures and multi-channel inputs to further improve accuracy and reliability.

Abstract

Given the destructive impacts of tropical cyclones, it is critical to have a reliable system for cyclone intensity detection. Various techniques are available for this purpose, each with differing levels of accuracy. In this paper, we introduce two ensemble-based models based on AlexNet architecture to estimate tropical cyclone intensity using visible satellite images. The first model, trained on the entire dataset, is called the global AlexNet model. The second model is a distributed version of AlexNet in which multiple AlexNets are trained separately on subsets of the training data categorized according to the Saffir-Simpson wind speed scale prescribed by the meterologists. We evaluated the performance of both models against a deep learning benchmark model called \textit{Deepti} using a publicly available cyclone image dataset. Results indicate that both the global model (with a root mean square error (RMSE) of 9.03 knots) and the distributed model (with a RMSE of 9.3 knots) outperform the benchmark model (with a RMSE of 13.62 knots). We provide a thorough discussion of our solution approach, including an explanantion of the AlexNet's performance using gradient class activation maps (grad-CAM). Our proposed solution strategy allows future experimentation with various deep learning models in both single and multi-channel settings.

Global versus Local: Evaluating AlexNet Architectures for Tropical Cyclone Intensity Estimation

TL;DR

This work addresses tropical cyclone intensity estimation from visible satellite imagery by introducing two AlexNet-based ensembles: a global model trained on all data and a distributed model with local experts per Saffir-Simpson category. The global AlexNet acts as a gating mechanism for the distributed ensemble, and their predictions are combined to estimate wind speed. Empirical results show both ensembles surpass the public Deepti benchmark, achieving RMSEs of knots and knots respectively, compared to knots for Deepti, with Grad-CAM analyses providing explainability by highlighting patterns related to Dvorak’s cloud features. The study highlights the interpretable nature of the models and suggests future work on deeper architectures and multi-channel inputs to further improve accuracy and reliability.

Abstract

Given the destructive impacts of tropical cyclones, it is critical to have a reliable system for cyclone intensity detection. Various techniques are available for this purpose, each with differing levels of accuracy. In this paper, we introduce two ensemble-based models based on AlexNet architecture to estimate tropical cyclone intensity using visible satellite images. The first model, trained on the entire dataset, is called the global AlexNet model. The second model is a distributed version of AlexNet in which multiple AlexNets are trained separately on subsets of the training data categorized according to the Saffir-Simpson wind speed scale prescribed by the meterologists. We evaluated the performance of both models against a deep learning benchmark model called \textit{Deepti} using a publicly available cyclone image dataset. Results indicate that both the global model (with a root mean square error (RMSE) of 9.03 knots) and the distributed model (with a RMSE of 9.3 knots) outperform the benchmark model (with a RMSE of 13.62 knots). We provide a thorough discussion of our solution approach, including an explanantion of the AlexNet's performance using gradient class activation maps (grad-CAM). Our proposed solution strategy allows future experimentation with various deep learning models in both single and multi-channel settings.
Paper Structure (17 sections, 2 equations, 4 figures, 2 tables)

This paper contains 17 sections, 2 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Domain Knowledge-Driven Approaches:
  3. Data-Driven Approaches:
  4. Hybrid Approaches:
  5. Contributions:
  6. Organization:
  7. Methods
  8. Dataset
  9. Models
  10. Global AlexNet Model
  11. Distributed AlexNet Model
  12. Results and Discussion
  13. Evaluation Metric:
  14. Performance Evaluation
  15. Model Explanation using Grad-CAM
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Common cyclone developmental patternsdvorak1973technique.
  • Figure 2: Event bias (top) and wind speed bias (bottom) in the training dataset.
  • Figure 3: Architecture of the global Alexnet model. Plot generated using PlotNeuralNet(https://github.com/HarisIqbal88/PlotNeuralNet)
  • Figure 4: Schematic of the distributed Alexnets. The final output is the average of global and local models.