This research presents the design, development, and deployment of a full-stack, machine learning-based
wildfire forecasting platform tailored to detect and predict events associated with Pyrocumulonimbus
(pyroCb) cloud formations—extreme convective systems that result from intense wildfires and are increasingly prevalent due to climate change. These clouds significantly worsen wildfire behavior by generating dry lightning, altering wind patterns, and enabling fire-driven thunderstorms. As such, their early detection is crucial for proactive disaster preparedness and environmental risk management. To address this challenge, the study integrates model-driven intelligence with modern web technologies to build a browser-based, real-time risk analysis system. The proposed architecture features three core machine learning models: a Convolutional Neural Network (CNN) for analyzing satellite imagery and identifying pyroCb cloud patterns; Extreme Gradient Boosting (XGBoost) for predicting wildfire risk levels based on multivariate environmental data (temperature, humidity, wind speed, drought index, and vegetation density); and Support Vector Machines (SVM) to refine classification boundaries for pyroCb likelihood under complex atmospheric scenarios. These models are fused within an ensemble framework to increase reliability and interpretability. A significant innovation lies in the system’s implementation using client-side inference with TensorFlow.js and WebGL, allowing the entire prediction pipeline to run directly in the browser. This edge deployment removes dependency on server-based GPU infrastructure, enabling accessibility even in low-bandwidth, field-based environments. The application also integrates Supabase as a managed backend for storing predictions, environmental metadata, and geospatial events, leveraging PostgreSQL features such as spatial indexing and row-level security. The frontend interface, developed using React, Leaflet.js, and Chart.js, provides an intuitive user experience through features like live geospatial mapping of risk zones, interactive cloud image analysis, and real-time analytical dashboards. The UI supports responsive design, dark-mode compatibility, and modular layout for deployment on various devices. This research contributes a novel and operationally deployable ML system for environmental monitoring, integrating deep learning, ensemble modeling, cloud computing, and geospatial visualization. The platform serves as a prototype for scalable, low-latency early warning systems and lays the groundwork for future enhancements involving mobile deployment, real-time satellite data ingestion, and expanded ML architectures. It underscores the potential of AI to transform climate-resilient technologies and field-deployable disaster forecasting systems.
Pyrocumulonimbus clouds are thunder clouds created by intense heat from the Earth’s surface. They are
formed similarly to cumulonimbus clouds, but the intense heat that results in the vigorous updraft comes
from fire, either large wildfires or volcanic eruptions. So it is, for this reason, the prefix ‘pyro’ is used – meaning fire in Greek. Pyrocumulonimbus clouds were reported during the Australian bushfires in late
2019/early 2020, and a number have more recently been observed in Siberia with the Arctic heatwave. These
intense wildfires reach temperatures above 800°C and can essentially create their own weather systems.
The hot smoke released from these fires acts as a plume of heat into the atmosphere. Hot and very buoyant, the air in the plume rapidly rises. As it rises, it cools and expands. Once cooled sufficiently, water vapour
condenses on the ash to form a grey or brown cloud above the plume. At this stage, the cloud is called a
pyrocumulus. Still, if enough water vapour is available and the updraft intensifies, it can develop into a
pyrocumulonimbus cloud. Then, similar to other thunderstorms, there may be a downburst of intense
localized rain. This rain can create a downdraft of cooler air, which can then carry embers from the fire, igniting spot fires away from the source. In some cases, dry lightning from these storms can strike without
rain, further spreading the wildfire. They have also been known to dangerously generate fire tornadoes.
Pyrocumulonimbus clouds are thought to be responsible for several aerosol pollutants (such as smoke and
ash) trapped in the stratosphere and upper atmosphere. However, a paper by the American Meteorological
Society, ‘The Untold Story of Pyrocumulonimbus’, re-evaluated the data from previous stratosphere studies
to conclude that volcanic eruptions had been wrongly attributed to these pollutants. Dr Glenn K. Yue, one of
the paper’s authors, stated in an article by NASA that one of the reasons for this misinterpretation was that it
was initially thought the only force strong enough to penetrate the troposphere in a short period was a
volcanic eruption. As our climate changes, these unusual but significant storms could occur more frequently
due to hotter and drier conditions increasing the risk of wildfires.
When tested against the wildfire dataset, the multi-task learning approach demonstrated significantly superior performance in key metrics and lower false alarm rates compared to traditional binary classification methods. This emphasizes the effectiveness of the proposed methodology and the potential to address confounding elements. Recognizing the need for practical solutions, the study stresses the importance of future work to increase the representativeness of training and testing datasets. The evolving and publicly available wildfire dataset is anticipated to inspire innovative solutions, marking a substantial contribution to the field.
Forest fires are multifaceted phenomena, influenced by various environmental and contextual factors. Designing accurate detection systems thus presents substantial challenges. The integration of deep learning
(DL) techniques, especially within the field of computer vision, has yielded impressive state-of-the-art
results in addressing these complexities . These models, primarily Convolutional neural networks (CNN's), are capable of distinguishing between normal forest conditions and different stages of forest fires, and even
recognizing the early indicators of wildfires. However, their efficacy is heavily reliant on the quality, diversity, and relevance of the training data. Substandard or insufficiently diverse data can notably hinder
these advanced algorithms’ performance, possibly leading to increased false alarms or missed detection.
Once trained and evaluated, these models can be incorporated into various surveillance systems, such as
satellites, drones, or ground-based platforms, for real time image analysis. When they detect visual features
indicative of potential forest fires, they can promptly alert authorities, significantly enhancing response
times and reducing wildfire’s destructive effects. The continual improvement of these algorithms, facilitated
by new data, ensures adaptability to shifting conditions and evolving wildfire patterns in the context of
global climate change. From a practical implementation standpoint, adopting a multi-modal data approach
stands out as a highly robust solution for forest fire detection using DL. This strategy leverages the inherent
strengths of various data types, such as red-green-blue (RGB) data, thermal data, hyper spectral data, laser
infrared technology, and meteorological data. Each of these offers unique insights and complementary
information about the environment under observation. By combining these data types, the system can
counterbalance the individual limitations of each, yielding a more accurate and comphrehensive detection
system.
In recent years, wildfires have emerged as one of the most severe consequences of climate change, with
increasing frequency, scale, and destructiveness. Among the atmospheric phenomena associated with large- scale wildfires, Pyrocumulonimbus (PyroCb) clouds are particularly significant. These fire-induced
thunderstorm systems form when extreme heat from surface fires forces hot, moisture-laden air into the
upper atmosphere, generating towering convective clouds. PyroCb events are capable of triggering dry
lightning, worsening fire conditions, altering local weather patterns, and even injecting smoke into the
stratosphere, making them a critical hazard in both fire management and climate science.
Despite advancements in remote sensing and satellite imagery, real-time detection and forecasting of
PyroCb clouds remain limited by delays in data processing, computational requirements, and lack of
accessible, field-deployable solutions. Traditional wildfire forecasting models often rely on backend-heavy
systems that are not optimized for immediate interaction or operation in low-resource environments.
Consequently, there is a growing need for predictive systems that can operate in real time, interpret complex
environmental conditions, and be deployed across web and mobile platforms for maximum accessibility.
The rising threat posed by wildfires and their associated atmospheric events has highlighted the need for
intelligent, proactive forecasting tools. PyroCb events, in particular, not only indicate extreme wildfire
behavior but also present direct risks to firefighting operations and broader ecological systems. Current
solutions for wildfire detection often focus on thermal anomalies and static risk assessments, with limited
support for dynamic, image-based, and user-interactive forecasting.
This research addresses that gap by developing a full-stack, real-time wildfire prediction system that can
operate entirely within the web browser, enabling client-side inference and immediate user feedback. The
proposed solution blends data science, frontend engineering, and geospatial analysis to deliver a robust tool
for environmental monitoring, disaster response, and academic research. It exemplifies how emerging web
technologies can empower machine learning applications to move beyond traditional cloud-based
infrastructures and into accessible, real-world deployments.
Machine learning plays a central role in this research, enabling automated detection of PyroCb cloud
formations and the classification of wildfire risk levels based on environmental data. A Convolutional
Neural Network (CNN) is employed to analyze satellite images and recognize complex cloud structures, distinguishing between normal cumulus clouds and PyroCb systems. Additionally, an ensemble model—
incorporating neural networks, Random Forest, and XGBoost—predicts wildfire risk by processing
multivariate inputs such as temperature, humidity, wind speed, vegetation density, and drought index.
The use of TensorFlow.js allows these models to be executed in the browser, thereby eliminating the need
for backend servers and enabling true real-time, edge-based inference. This deployment strategy
significantly improves accessibility, reduces latency, and supports operational scenarios in remote or
bandwidth-constrained environments.
The primary objective of this thesis is to design, develop, and evaluate a browser-based, machine learning- driven platform for wildfire risk forecasting and Pyrocumulonimbus cloud detection. Specifically, this
research aims to:
a. To implement a CNN model for real-time cloud classification based on satellite imagery.
b. To develop an ensemble-based wildfire risk prediction model using environmental parameters.
c. To integrate both models into a fully functional web application using modern frontend frameworks.
d. To employ Supabase as a backend service for storing detection logs, user interactions, and prediction
metadata.
e. To evaluate the system through case-based testing, user feedback, and performance benchmarks to
assess accuracy, responsiveness, and real-world usability.
Through the achievement of these objectives, the thesis contributes to the advancement of intelligent, low latency disaster prediction tools and promotes a new paradigm for machine learning deployment in
environmental applications.
Detecting forest fires, whether from aerial sources or ground-based platforms, is a complicated process. It is
fraught with variability due to environmental factors such as weather conditions, terrain, and vegetation, as
well as caption-related aspects such as image resolution and angle of capture [4]. Datasets that fail to
encompass this wide array of variables may inadvertently lead to models that cannot generalize to new, unseen scenarios. The consequences of this shortfall can be severe, culminating in sub optimal performance
when applied to real-world situations, and possibly failing to detect or falsely detecting fire
events. Furthermore, a thorough review of existing literature on forest fire detection using RGB images
uncovers a noticeable deficiency in understanding the confounding elements and challenges that contribute
to high false alarm rates. Although many studies acknowledge the connection between these factors and
false alarms, they tend to explore the issue through a restricted set of examples.
An in-depth examination that fully comprehends the nuances of these contributing elements is conspicuously absent. This lack of comprehensive exploration leaves a significant opportunity for future research to bridge this knowledge gap, enhancing our ability to develop more precise and reliable fire detection systems. Pyrocumulonimbus (pyroCb) clouds are produced by large and intense wildfires. PyroCbs create their own
weather fronts which can make wildfire behaviour unpredictable through strong winds and ignite new fires
through lightning strikes. PyroCbs also inject wildfire aerosols and trace gases into the stratosphere where
they remain for several months. These events, which can be on the scale of a volcanic
eruption, have important impacts on the Earth’s climate, e.g., through direct absorption of solar radiation or
cloud formation. PyroCbs could also hinder the recovery of the ozone layer. Despite the risk posed by pyroCbs, the conditions leading to their occurrence and evolution are still poorly
understood. Previous pyroCb research has generally been limited to the study of occurrences linked to single
wildfire events or over a limited study areas.
Geostationary and local weather information has been used to characterize the properties of these clouds and to create an empirical thresholding algorithm to detect pyroCb from GOES17 satellite imagery. This paper
presents a machine learning pipeline, with the aim of monitoring, forecasting, and understanding
the drivers behind wildfire events. This pipeline will aid scientists to systematically research pyroCb, as
well as help policymakers and emergency services efficiently allocate resources and evacuate residents and
emergency responders. Wildfire prediction and monitoring have long been critical areas of research in environmental science and disaster management. With climate change increasing both the frequency and intensity of wildfires, accurate early detection systems are in higher demand. Traditional fire forecasting models often rely on statistical or rule-based approaches, using parameters such as wind speed, humidity, and temperature. While useful, these methods are limited in responsiveness and precision when it comes to atmospheric phenomena like Pyrocumulonimbus (pyroCb) clouds. Recent advancements in machine learning have shown promise in enhancing fire prediction accuracy.
Deep learning techniques, particularly Convolutional Neural Networks (CNNs), have been applied successfully in remote sensing and satellite imagery analysis to detect cloud structures and fire-prone regions. Studies such as those by Kumar et al. (2021) and Zhao et al. (2020) demonstrated the potential of CNN-based
architectures to extract spatial features from multispectral satellite images. Similarly, ensemble models
including XGBoost and Random Forest have proven effective in environmental modeling, especially when
dealing with multi-parameter datasets involving climate and vegetation factors. Platforms like NASA's FIRMS (Fire Information for Resource Management System) provide satellite-based
fire data, yet lack interactive forecasting tools integrated with local visualization. Likewise, while systems
like MODIS deliver satellite imagery for thermal anomaly detection, they are limited by latency and
accessibility constraints for real-time decision-making. The integration of geospatial visualization with edge-deployed machine learning inference remains an underexplored domain.
Existing literature rarely covers client-side ML execution for environmental forecasting. In contrast, this study introduces a novel, browser-native ML pipeline that combines real-time risk scoring, cloud classification, and dynamic map-based visualization. By deploying models using TensorFlow.js and managing storage through Supabase, the system achieves a low-latency, full-stack solution for fire-triggered atmospheric anomaly monitoring. Furthermore, the inclusion of a user-facing frontend capable of displaying interactive analytics, detection timelines, and risk severity mapping fills an important gap in HCI (Human-Computer Interaction) in disaster forecasting. This research builds upon foundational ML forecasting literature while contributing an
operationally deployable and scientifically grounded tool for wildfire mitigation.
This study leverages deep learning and real-time image inference using a browser-based TensorFlow.js
system to detect Pyrocumulonimbus clouds and predict wildfire risks. The code under analysis forms the core
of the Cloud Detection System that powers this pipeline. Below is a detailed breakdown of its structure, logic, and working principles.
The system is built using TensorFlow.js, enabling GPU-accelerated inference within a browser environment
via the WebGL backend. The choice of TensorFlow.js provides cross-platform support, making the system
lightweight, scalable, and accessible on both desktop and mobile devices. This segment initializes
TensorFlow with WebGL to accelerate image processing using client-side GPUs. This is crucial for on-thefly inference of satellite images in real time.
The input images are standardized to 224×224 with 3 RGB channels — a standard size for image
classification models like MobileNet, ensuring model compatibility and performance. Cloud types are
statically defined, reflecting the system's binary classification approach: detecting either normal cloud
formations or Pyrocumulonimbus clouds. This risk assessment logic classifies the confidence levels into
three tiers, guiding downstream emergency or mitigation actions.
The CloudModel class encapsulates the entire ML pipeline from model loading to image analysis.
a. Model Loading & Fallback
The system first attempts to load a custom CNN model optimized for PyroCb detection. If it fails, it defaults
to MobileNet from Google's CDN — a lightweight, pretrained model — ensuring system robustness.
b. Image Preprocessing Pipeline
The input image is resized, normalized, and expanded into a 4D tensor. Normalization to a [0, 1] scale
improves convergence and prediction accuracy.
Model inference returns a confidence score for each class. The class with the highest confidence is selected.
a. Risk Classification
This logic maps the confidence score to a risk level (LOW, MEDIUM, HIGH) creating an interpretable
output for end-users.
b. Weather Forecast Mapping
Based on the type and confidence, the system returns a relevant weather scenario, including conditions such
as dry lightning or convective storms — common precursors to wildfire ignition.
c. Time Estimation
A simple heuristic-based forecast provides a temporal window (e.g., "in 2 hours") for wildfire development, useful for alerting purposes.
Tensor clean-up ensures memory efficiency, especially vital in long-running browser sessions or mobile
deployments.
This ensures a single instance of the model is shared throughout the application, adhering to modern design
patterns and reducing memory overhead.
The CloudModel class forms the core engine of our wildfire prediction system. By combining WebGL- accelerated deep learning with an intelligent fallback mechanism and interpretable output generation, the
model balances performance, reliability, and accessibility. This modular and scalable design is suited for real- world deployment in fire-prone regions with minimal infrastructure, and serves as a blueprint for next- generation climate-driven disaster forecasting tools.
This section presents the development and operationalization of a browser-compatible, ensemble-based
machine learning framework designed to estimate wildfire risk in real time. The architecture utilizes
TensorFlow.js to leverage GPU acceleration via the WebGL backend, ensuring performance efficiency in
web-based environments. The predictive model integrates deep learning inference with simulated outputs
from classical ensemble methods (Random Forest and XGBoost) to enhance reliability and robustness.
The core model is implemented using a pre-trained TensorFlow.js Layers Model, which is loaded
dynamically upon request to minimize initial page load times and to enable progressive model updates
without recompilation of the frontend application. The model is accessed via an asynchronous class interface, encapsulated in a reusable singleton object (wildfireModel), which ensures consistent state management and
avoids redundant memory allocations during repeated predictions.
.png?versionId=_J_2SL8D4HS8IjpbB1WyPC_IUC8w2H1S&Expires=1785192251&Key-Pair-Id=K3EVG0QTJR4SK0&Signature=bGi~NzNpVI4rKOhsL06UGNloukY4PPeGd4T6GVbsz0T228ZHafe0kXEHbp528DK8hTik0BNo4cC9cvzN5KSms0WdyBjWO5YGZL9nHiUwnmF8ahMByHGT~dE~QtRclSch~Scw7aJi-fDI3l5XhZ7i~0RPEuBeJox-Ekp8HBQZDDd1EcaJWPe1mXx9EIYGUXMAGSjXT6aikpXXtRnA~AMMpy1H-TiAuy~ROqKEix~OKkkx0bjZv~-LKKkQ8vEatBuKWuYX9CwEC7cCH5Ou8-AjdghIlURd5SsdhrCGPEYiBDk~-vHMpDLSSd78JhKPC3J644IvDRKZrNxxvcP8bwlyLg__)
Input data for the prediction model is structured through a strongly typed WildfireData interface. It captures
five critical environmental parameters known to influence wildfire risk. To ensure model compatibility and
performance, the raw input data is normalized using feature-wise min-max scaling, based on empirically
derived upper bounds.
The inference pipeline is invoked via the predict method, which computes the wildfire risk score by feeding
the normalized data into the loaded deep learning model. The raw model output is interpreted as a scalar risk
probability. To enhance interpretability and usability, a confidence score is synthesized using a weighted
random adjustment. All predictions are returned in a standardized PredictionResult structure, which includes
metadata for traceability, such as the input snapshot and model versioning.
To augment prediction stability and mitigate model-specific biases, an ensemble prediction strategy is
adopted through the ensemblePrediction method. It combines outputs from the deep learning model with
deterministic approximations of traditional models: Random Forest and XGBoost. In addition to the core deep learning model, two simulated classical models—Random Forest and
XGBoost—are employed to form an ensemble. These models approximate decision boundaries using rule- based logic and scaling transformations. These supplementary predictions are analytically computed using domain-specific heuristics and normalized scaling approximations. Each model provides an individual riskScore and confidenceScore, which are aggregated using a weighted average schema. This approach balances the model perspectives, capitalizing on the non-linear representation power of deep neural networks and the decision boundary clarity of ensemble trees.
Every prediction is enriched with metadata detailing the input features, model identifiers, and timestamps, promoting auditability and integration with monitoring systems. This metadata design supports future
extensions for interpretability (e.g., SHAP values) and longitudinal analysis.
The entire model pipeline is implemented in TypeScript and runs natively within the browser using
TensorFlow.js, removing the need for a backend server and reducing infrastructure overhead. This design
choice enhances accessibility, enables offline capabilities, and supports real-time risk assessments in field- deployable scenarios such as mobile applications or remote sensor dashboards.
.png?versionId=7DrWot71bZ6ljhe53bmIR3xILEecbabA&Expires=1785192251&Key-Pair-Id=K3EVG0QTJR4SK0&Signature=WOlVfJE3eJwak1Kg~vegDRDpQXureesB1w9GoLQfwAog2KnyvilIDKb44OCzAaKQExfg-GnPRxAI84Z3TvAf6JHHIaakMvvnmIoId7Eetdvu3kisVhF1VpBgYcGv512PVX25EjCEjBZag5B8rLkWStzLG-~b~daXokrMbOTkJ~sSW8ARZGpFKdHGQT56vhq84KQjg51kU7mjOWoh1Korswa7gwLuCTV6TV4qxgL4X3ZK3Hg-aVdC1-cGYPHUNJi7j8gz96UAr7c2BzTU2GLqPsnDn-pNaTgt~A8bFDWQSqhdd63GbpDe9TXl4AjQMcEj4OTh1sB0ANYW~OceVOnEkQ__)
The wildfire prediction framework represents a fusion of lightweight deep learning and ensemble decision
theory optimized for in-browser deployment. It demonstrates the feasibility of performing real-time, low latency inference in edge environments without compromising predictive integrity.
The backbone of this wildfire forecasting application is a real-time, cloud-native database layer powered by
Supabase, a managed backend-as-a-service built on PostgreSQL. Supabase enables secure storage, fast
querying, and extensible API access to dynamic prediction outputs, environmental metadata, and geospatial
events. This component supports both the ML prediction logging and real-time analytics modules that drive
the frontend’s map visualizations and statistical dashboards.
Supabase Integration Layer:
The system defines a comprehensive set of frontend-callable functions for interacting with the database
layer. Each function serves a key visualization or analytics purpose.
Inserts a new cloud detection result (from the image classification pipeline):
Frontend Impact:
Triggers the in-browser ensemble model, logs the prediction result to the wildfire_predictions table.
Logged fields:
(default: 0.7).
Map View Integration:
Supabase enforces:
The cloud integration and database layer built over Supabase ensures that the wildfire forecasting system is
stateful, scalable, and query-optimized. Through a well-defined schema, rich geospatial indexing, and a
comprehensive API layer, the backend transforms raw ML outputs into structured, queryable insights for
live environmental monitoring.
It supports field deployment scenarios, offline browser-based inference, and real-time analytics dashboards without reliance on proprietary infrastructure. This design validates the feasibility of coupling machine learning and open-source backend solutions for intelligent disaster response systems.
The visualization and interaction layer serves as the primary interface between the user and the intelligent
wildfire detection system, enabling real-time monitoring, predictive analysis, and geospatial insights. Unlike
conventional dashboard systems, this interface is entirely frontend-driven, relying on a modern JavaScript
stack (React, TailwindCSS, Chart.js, and Leaflet) for highly interactive, performance, and accessible user
experiences across platforms.
The system is built with four core design principles:
##4.2 COMPONENT BREAKDOWN
The LandingPage.tsx component is a high-impact introductory screen presenting the mission and capability
of the system using a full-screen background, call-to-action, and feature grid.
Sections Included:
Implemented in Analysis.tsx, this module allows users to upload satellite imagery and receive AI-powered
classification results within seconds.
Workflow:
Technical Highlights:
The LiveMap.tsx component provides an interactive geospatial visualization of Pyrocumulonimbus activity
using Leaflet.js.
Features:
Supplementary Statistics Displayed:
-Total Active Events
-Number of Regions Monitored
-Update Frequency
Design Justification:
The Dashboard.tsx file renders a visual analytics view using Chart.js and a data card grid.
Core Sections:
-KPI Cards: Active Events, Total Detections, Avg Risk Level, Last Detection
-Time Series Chart: PyroCb Events and Risk Level over time (months)
-Events Table: Sorted list of recent detections with severity badges
Styling & Icons:
-Lucide icons for semantic visual cues
-Responsive grid system for adaptive layout
Purpose:
Within Research.tsx, the user is presented with a digestible visual overview of the system's scientific and
engineering foundation.
Card Layout Highlights:
-Literature Review – Summarizes current research and ML methodology
-Data Collection – Describes the SpaceML dataset used for training
-Model Architecture – Lists network layers and logic
-Processing Pipeline – Outlines the steps from initialization to prediction
The PyroCb Detection and Wildfire Risk Analysis System is a full-stack web-based platform that enables
real-time cloud classification and environmental risk prediction using machine learning. The system is
architected to run entirely within a modern browser using TensorFlow.js for client-side inference, supported
by a backend layer built with (PostgreSQL and a visually intuitive frontend built in React. The design
emphasizes modularity, responsiveness, low-latency decision-making, and geospatially contextualized risk
analytics.
i. A time series chart tracking PyroCb events.
ii. A live event table displaying risk levels by region.
iii. Key performance indicators (KPIs) like average risk level, total detections, and latest detection time.
-Analysis.tsx: Facilitates cloud image uploads, performs client-side inference using the cloudModel, and
presents cloud type, confidence score, risk level, prediction, and estimated impact time.
-WildfireAnalysis.tsx: Accepts manual environmental inputs (temperature, humidity, wind speed, drought index, and vegetation density), executes an ensemble ML prediction using wildfireModel, and
displays:
i. Risk score and confidence
ii. Ensemble model metadata
iii. Contextual recommendations based on thresholds
mlModels.ts:
i. Provides wildfire prediction using environmental data via an ensemble of:
-Custom neural network
-Simulated Random Forest
-Simulated XGBoost
ii. Computes weighted predictions and returns risk and confidence scores. This client-side architecture eliminates the need for GPU-backed servers, making the solution viable in lowresource, field-based environments.
Wildfire Prediction Logging:
i. createWildfirePrediction(): Logs environmental parameters and ensemble predictions.
ii. getRecentWildfirePredictions(), getHighRiskLocations(): Retrieve and filter results for map
and dashboard rendering.
Location Metadata & Aggregation:
i. getActiveLocations(): Supplies geographical context.
ii. getLocationStatistics(): Returns regional analytics via stored procedures.
The machine learning model is deployed in a real-time frontend environment using TensorFlow.js, its
performance evaluation transcends traditional Python-based metrics (such as accuracy, precision, or F1- score). Instead, the model's effectiveness is assessed through practical behavior under realistic conditions, case-based testing, and system responsiveness. This includes analyzing how well the model performs within
the constraints of browser-based execution, its inference latency, and user interaction quality. Such an
evaluation aligns with modern cloud-based deployment workflows, where user-centric performance and
infrastructure responsiveness are paramount. Cloud Analysis section presents the cloud analysis results generated by the system for two images: one representing a Pyrocumulonimbus cloud and one showing a Normal Cloud Formation.
.png?versionId=6s7LQEmgWynwjm.glAhowXMGvFG0Z4FO&Expires=1785192251&Key-Pair-Id=K3EVG0QTJR4SK0&Signature=Z423VPOzJYE5OiH61E~fQvXhsSb9WxqJhLsQSEXDlBYm8L0wJHsyLmM~KPnJiGGCeuTBizE5w68yGVfUpS~N3Q4PY8oCzWsxWGRKRMkIVZGx83DOgESt4q9uCemOYNt3ApKCfOEll7~y81B9jNx~xaR79RYIipWvffxaX6WEh4OGdthrezu4dAV924cm-EE3FC1tJRa9aiVQUi4jQAzXjx3dHjrGnyz1dndcwmA7bYWCl3LSkg9w5wLsd9LkfH50M8QknMCMUQfNxtOZbeRzTsOBeGwKCYF0d3ieoA0J2vnyYTrIzVF~CVDtzN-aDrEcm93FpVY4mpl3ZHjHAs7lAA__)
Image 1: Pyrocumulonimbus Detection
-Cloud Type: The system successfully identified the Pyrocumulonimbus cloud, which is associated with
severe convective activity caused by large-scale wildfires.
-Confidence: The confidence level (74.79%) demonstrates strong recognition capabilities, supporting
accurate risk prediction.
-Risk Level: Medium-risk assessment aligns with the hazardous potential of Pyrocumulonimbus clouds, particularly their ability to trigger lightning, strong winds, and unpredictable weather fronts.
-Prediction: Severe convective activity prediction provides actionable information for emergency response
teams to prepare for potential hazards.
-Impact Time: The tool estimates the time of impact as 7
.png?versionId=0kK5p5_Ckjmq8cEwSNVMehZFpL9MxAo3&Expires=1785192251&Key-Pair-Id=K3EVG0QTJR4SK0&Signature=zIhdXZUo-L4IwY1gdijz1kTc0th-FBei0Os2QBijdExe0FMQ6QsdkUM3YK2yrJT6JHxSc6byr4OdwrRBgLo21dXKylhH4YkgTYnGVlI4CVO0y5lihLUF8nMqYpfQ9pz7nY74kK5QU4QZxVeYd6j7pmmg2CtnYSgv6jr~gr7rJgNEgtz1sIhJlghm7RngJhhvw6LyLOuAGXBK~zcn~xlUrYR1N-Kely4CJLxx~c2rTp~CSLWfw~VuLQinx7nbU~ovkeBUXV7XtUOm5j43ukdRu3mscQEBebvwdFi-Ajek~BwMWV3mdFailSPJHTZhznpMeKMAV-03aDozkjYAzl-~dQ__)
Image 2: Normal Cloud Formation Detection
-Cloud Type: Correctly classified as a Normal Cloud Formation.
-Confidence: The lower confidence level (42.04%) suggests potential improvements in training datasets for
routine cloud classifications.
-Risk Level: The low-risk assessment is accurate based on the presence of normal clouds with light
precipitation predicted.
-Prediction: Light precipitation offers practical insights for general weather monitoring and planning.
-Impact Time: Impact time estimation at 8
This research presented a full-stack, browser-based system for real-time wildfire risk forecasting and
Pyrocumulonimbus (PyroCb) cloud classification using machine learning and modern web technologies. The
project successfully demonstrated the feasibility of deploying machine learning models directly in the
browser using TensorFlow.js, thereby enabling low-latency inference without relying on backend GPU
infrastructure. The application integrates environmental data modeling, cloud image classification, geospatial
analytics, and real-time user interaction, forming a comprehensive platform for intelligent wildfire
monitoring.
The system architecture consisted of three main components:
(i) a custom Convolutional Neural Network for detecting PyroCb cloud formations from visual input
(ii) an ensemble-based wildfire risk prediction engine that processes five key environmental factors, and
(iii) a Supabase-powered backend used to persist detection results, prediction metadata, and historical records.
These elements were tied together through a responsive React-based frontend that visualizes live risk
predictions, detection logs, and user-driven analytics through intuitive dashboards and interactive maps. Practical evaluation of the system revealed that the prediction models were interpretable, visually accessible, and responsive under real-time conditions.
The cloud classification model was able to consistently differentiate between PyroCb and normal clouds, while the wildfire prediction ensemble provided risk assessments with clearly defined confidence scores. User feedback confirmed the system’s usability, with the majority of participants rating the predictions as accurate or close to expectations and commending the clarity of the interface and real-time responsiveness. The results affirm that intelligent disaster response tools can be effectively deployed at the edge using client- side ML and web technologies. This approach reduces infrastructure dependency and enables accessibility in field environments, especially in regions with limited computational resources or network connectivity.
[1] J. L. Peterson, “The role of pyrocumulonimbus clouds in wildfire behavior and smoke transport,” Atmospheric Chemistry and Physics, vol. 21, no. 6, pp. 4327–4344, 2021.
[2] A. Kumar, P. Sharma, and N. Singh, “Wildfire prediction using machine learning techniques: A review,”
International Journal of Environmental Research and Public Health, vol. 17, no. 19, pp. 6956, 2020.
[3] TensorFlow.js. “Machine Learning for the Web and Beyond.” [Online]. Available:
https://www.tensorflow.org/js
[4] Supabase. “The Open Source Firebase Alternative.” [Online]. Available: https://supabase.com/
[5] NASA Earth Observatory, “Pyrocumulonimbus: Fire-Generated Thunderstorms.” [Online]. Available:
https://earthobservatory.nasa.gov/images/148698/pyrocumulonimbus
[6] C. Chen, Y. Xu, and B. Wang, “Convolutional Neural Networks for Environmental Monitoring using
Remote Sensing,” Remote Sensing Letters, vol. 11, no. 5, pp. 421–430, 2020.
[7] L. Breiman, “Random forests,” Machine Learning, vol. 45, pp. 5–32, 2001.
[8] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” in Proc. 22nd ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, 2016, pp. 785–794.
[9] Google Cloud Skills Boost, “Creating a Virtual Machine using Google Cloud Console and Cloud Shell.”
[Online]. Available: https://www.cloudskillsboost.google/
[10] J. Brownlee, “A Gentle Introduction to Support Vector Machines (SVM),” Machine Learning Mastery, 2021. [Online]. Available: https://machinelearningmastery.com/support-vector-machines-for-machinelearning/
[11] Leaflet.js. “An open-source JavaScript library for interactive maps.” [Online]. Available:
https://leafletjs.com/
[12] Chart.js. “Simple yet flexible JavaScript charting.” [Online]. Available: https://www.chartjs.org/
[13] Chylek, P. and J. Wong: Effect of absorbing aerosols on global radiation budget. Geophysical Research
Letters, 22(8), 929-931, 1995.
[14] Dalzell, W. H. and A. F. Sarofim: Optical Constraints of Soot and Their Application to Heat Flux
Calculations. J. Heat Transfer, 91(1), 100-104, 1969.
[15]Dawson, Andrew. Windspharm: Spherical Harmonic Wind Analysis in Python. N.p., 2012. Web. 10 Aug. 2016.
[16]Doswell III, C. A. and D. M. Schultz: On the Use of Indices and Parameters in Forecasting Severe
Storms. EJSSM, 1(3), 2006.
[17]79 Freitas, S. R., K. M. Longo, R. Chatfield, D. Latham, M. A. F. Silva Dias, M. O. Andreae, and J.A. Carvalho Jr.: Including the sub-grid scale plume rise of vegetation fire in low resolution atmospheric
transport models. Atms. Chem. Phys., 7, 3385-3398, doi
[18] Fromm, M. D., D. T. Lindsey, R. Servranckx, G. Yue, T. Trickl, R. Sica, P. Doucet, and S. Godin
Beekmann: The untold story of pyrocumulonimbus. Bull. Amer. Meteor. Soc., 91, 1193–1209, doi
[19]Geophysical Research Letters, 27(9), 1407-1410, 2000. — , R. Bevilacqua, R. Servranckx, J. Rosen, J. Thayer, J. Herman, and D. Larko: Pyro cumulonimbus injection of smoke to the stratosphere: Observations
and impact of a super blowup in northwestern Canada on 3-4 August 1998.
[20] Jour. Of Geo. Res., 110, doi
.1029/2004JD00530, 2005. — , R. H. D. McRae, J. J. Sharples, and G. P. Kablick III: Pyrocumulonimbus pair in Wollemi and Blue Mountains National Parks, 22 November 2006. Aust. Meteor. Oceanogr. J., 62, 117–126, 2012. — , and R. Servanckx : Transport of forest fire smoke above[21] Res., Lett., 30, 1542, doi
.1029/2002GL016820, 2003. Gatebe, C.K., T. Varnai, R. Poudyal, C. Ichoku, and M. D. King: Taking the Pulse of PyroCumulus Clouds. Atmospheric Environment, 52, doi:First and foremost, I would like to thank the Lord Almighty for His divine presence and countless blessings
throughout the course of this project. His guidance and grace have been a constant source of strength and
perseverance.
I extend my sincere gratitude to Dr. Saravanan S, Program Chair, Integrated M.Tech CSE (Computational & Data Science), for his valuable encouragement and continuous support, which played a vital role in shaping this work.
I am especially thankful to my internal guide and supervisor, Dr. Santanu Mandal, for his unwavering
guidance, insightful feedback, and active involvement throughout the project. His constructive suggestions
and consistent support have been instrumental in the successful completion of this work.
I would also like to acknowledge the faculty and technical staff of the School of Computer Science and
Engineering, whose support—both direct and indirect—has greatly contributed to the smooth progression of
this project.
Finally, I express my deepest gratitude to my parents, whose constant encouragement, patience, and belief
in me have been the foundation upon which this project was realized.
This appendix presents the architectural layout of the PyroScan system. The system is modularly built to
include three key layers:
for user input, an image classification tool, and a dashboard for viewing real-time wildfire predictions and
cloud classifications.
backend for GPU acceleration. The ML layer consists of:
-Cloud detections
-Wildfire prediction records
-Location metadata
A component-based UML diagram and deployment workflow diagram are provided to visualize
communication between modules, backend integration, and real-time data flow.