Real-Time Forest Fire and Human Anomaly Detection using IoT and Machine Learning

Real-Time Forest Fire and Human Anomaly Detection using IoT and Machine Learning

Arya Pradeep Chavhan

Electronics and Telecommunications Sardar Patel Institute of Technology, Mumbai, India

Dikshita Naresh Hate

Electronics and Telecommunications Sardar Patel Institute of Technology, Mumbai, India

Akshada Dilip Butte

Electronics and Telecommunications Sardar Patel Institute of Technology, Mumbai, India

Dr. Sukanya Kulkarni

Assistant Professor Sardar Patel Institute of Technology, Mumbai, India

Abstract – Early detection of forest threats like illegal human activity, wildlife movements, and re outbreaks is essential for protecting the environment and preventing disasters. This research introduces a monitoring framework that combines acoustic anomaly detection with forest re risk prediction. The system uses Mel-Frequency Cepstral Coefcients (MFCC) from real environmental audio data stored in an Amazon S3 bucket. It detects anomalies with a trained autoencoder and classies these anomalies as either human or animal presence based on MFCC centroid similarity. For assessing re risk, an XGBoost classier, trained on historical data from forest sensors, predicts risk levels and spots potential re hazards. A real-time dashboard based on Flask shows both systems at the same time, allowing for ongoing monitoring of acoustic anomalies and re risk trends. Experimental results show that the system can effectively detect environmental anomalies and estimate re risks in near real-time.

1. Introduction

   Forests are complex ecosystems that are vital for biodi-versity, carbon storage, climate control, and human health. However, various human-made and natural threats weaken their stability. Forest res alone cause billions of dollars in damage each year while increasing greenhouse gas emissions and reducing biodiversity. Additionally, illegal logging, poach-ing, and human encroachment disrupt wildlife habitats and speed up deforestation. Traditional monitoring methods, such as satellite imaging, manual patrolling, and isolated sensor networks, have drawbacks like slow response times, limited coverage, or high deployment costs.

   Recent developments in the Internet of Things (IoT), ma-chine learning, and cloud computing allow for real-time forest monitoring. Acoustic analysis has shown promise as a strong indicator of environmental activity because many forest dis-turbances make distinct sounds. At the same time, data from environmental sensors, such as temperature, humidity, and gas levels, can help predict re risk. However, monitoring sound or re risk separately does not provide a complete view.

   This research proposes a system that detects sound anoma-lies, classies the type of disturbance, and predicts forest re risk within a single integrated framework. The uniqueness of this system is in combining acoustic anomaly detection with machine learning-based re risk prediction and dashboard monitoring for visual analytics.

2. Literature Review

   A literature review examined recent developments in forest re detection, monitoring human and wildlife anomalies, and using IoT and machine learning for environmental monitoring. The studies show a shift from traditional sensor-based methods to AI-powered frameworks. This reveals both successes and ongoing challenges. The main references in this review include works on audio-based anomaly detection [1], [2], IoT-driven re detection [3][5], and machine learning methods for environmental monitoring [6], [7].

   1. Audio-Based Anomaly Detection

      Acoustic monitoring has become an effective way to iden-tify unusual environmental events, like human intrusion, an-imal activity, and early signs of re. Kumari and Saini [1] proposed a method using adaptive Huffman coding to process audio data in real time. This approach reduced false positives caused by background noise or changes over time. Similarly, Nunes [2] conducted a systematic review of detecting unusual sounds with machine learning. He highlighted challenges such as small labeled datasets, environmental noise, and differences in audio quality. These studies show the promise of using audio for detecting anomalies, but they also point out the need for strong preprocessing and feature extraction methods. One example is Mel Frequency Cepstral Coefcients (MFCCs), which help capture important sound features effectively.

   2. IoT-Based Forest Fire Detection

      IoT-based solutions for monitoring forest res use sensors to detect changes in the environment like temperature, gas concentration, and smoke levels. Avazov et al. [3] introduced a hybrid system that combines IoT modules with AI algorithms. They showed that analyzing gas concentration, temperature changes, and visual signs at the same time can improve early detection. Miriyala et al. [4] created an affordable IoT detection system using NodeMCU and environmental sensors to send alerts in real time. Bharadwaj et al. [5] highlighted the need for low-power sensor networks and timely alerts to respond to forest res quickly. While IoT solutions are cost-effective and provide real-time monitoring, challenges like reliance on networks, limited range, and sensitivity to environmental conditions still exist.

   3. Machine Learning and Computer Vision Approaches

      Machine learning and computer vision techniques have been used to nd anomalies and predict environmental hazards. Mehta et al. [6] showed how supervised learning models can detect problems in IoT sensor data. They pointed out the advantages of automated data analysis. Han et al. [7] looked at IoT anomaly detection when data quality is low. They stressed the importance of preprocessing and feature engineering to keep the model effective. These studies suggest that machine learning can improve prediction accuracy and reliability, but there are still issues with computational cost, real-time use, and generalizing across different environmental conditions. Combining these methods with acoustic and IoT-based sensing can create a more complete monitoring system.

   4. Summary of Gaps and Motivation for Current Study

      The literature shows signicant progress in forest monitor-ing using IoT, acoustic sensing, and machine learning. How-ever, existing studies often view audio anomaly detection, re risk prediction, and visual monitoring as separate tasks. There is a clear gap in integrated approaches that combine different types of sensing, real-time processing, and easy-to-understand visualization for monitoring both environmental and human activities at the same time. This study aims to ll those gaps by merging acoustic anomaly detection and XGBoost-based re risk prediction into one real-time dashboard system. This will provide better situational awareness and useful insights for managing forests.

3. Project Objectives

   The proposed system aims to develop a real-time, integrated framework for detecting forest res and human anomalies with the following objectives:

   1. Environmental Sensing and Fire Prediction: Deploy IoT sensors to monitor temperature, smoke, CO, and gas concentrations. Use an XGBoost classier to process sensor data for precise re prediction.

   2. Human Anomaly Detection: Record ambient audio with MEMS microphones. Examine audio signals with

      an autoencoder neural network to spot unusual human activity.

   3. Data Visualization and Alerts: Create a web-based dashboard for real-time monitoring. Send instant alerts for both re risks and human presence.

   4. Scalability and Energy Efciency: Ensure the system works in remote forest areas. Improve sensor and mi-crocontroller power use for long-term deployment.

   5. System Integration: Merge sensor acquisition, ML processing, and web visualization into a smooth, real-time monitoring solution.

4. System Architecture

   The proposed system uses a multilayered architecture that combines sensing, cloud storage, machine learning inference, and real-time visualization into a single monitoring frame-work. Fig. 1 shows the architectural ow, starting at the data acquisition layer and ending at an interactive dashboard. This design focuses on scalability, modularity, and asynchronous operation. It ensures that each subsystem can work indepen-dently while still contributing to an overall situational picture.

   1. Data Acquisition Layer

      The lowest tier comprises heterogeneous environmental data sources:

      1. Audio Sensors: Forest-mounted microphones capture ambient acoustic signals like human speech, animal sounds, and general forest noise. We choose a sampling frequency that maintains the clarity of speech and ani-mal calls while reducing bandwidth and storage needs. Later, we convert the recorded audio signals into Mel Frequency Cepstral Coefcients (MFCCs). This process allows for a compact representation of features suitable for detecting anomalies and classifying tasks.

      2. Environmental Sensors: Temperature and humidity sensors measure atmospheric conditions that closely relate to the chance of forest res starting. A multi-gas oxygen sensor checks the oxygen level and its changes due to combustion or smoke. Lower oxygen levels and shifts in gas concentration are early signs of re starting and spreading. The combined dataset, which includes temperature, humidity, and oxygen concentration, serves as the input for the XGBoost-based re risk prediction model. This setup allows for real-time evaluations of re risk and potential threats to people or animals nearby.

   2. Cloud Storage and Communication Layer

      In the proposed system, the cloud layer is crucial for storing and retrieving data used for anomaly detection and estimating forest-re risk. The Raspberry Pi sensing units continuously capture short audio segments and convert them into Mel Frequency Cepstral Coefcient (MFCC) feature vectors. These MFCC les are directly uploaded to an Amazon S3 bucket. The choice of S3 is due to its ability to reliably store large amounts of small les generated at regular intervals, while

      providing durability and simple URL-based access for later retrieval.

      S3 allows the system to securely store all incoming MFCC vectors without needing manual storage management. This ensures that historical and real-time audio samples remain ac-cessible for inference. The bucket acts as the central repository for incoming eld data, with each MFCC le timestamped and stored for later processing.

      To perform machine learning inference, an Amazon EC2 instance retrieves MFCC les from S3 using the boto3 SDK. EC2 serves as the computation layer where it loads the autoen-coder model, calculates reconstruction errors, and performs classication. The communication ow is clear: Raspberry Pi, S3, EC2. This method reduces complexity and keeps sensing, storage, and inference separate, while still allowing efcient communication through AWS services.

   3. Processing and Machine Learning Layer

   The processing and machine learning tasks run entirely on an AWS EC2 instance. This instance hosts all the models and algorithms used for real-time decision-making. After downloading MFCC feature vectors from S3, the EC2 server processes them through three main components.

   First, the EC2 server feeds the MFCC vectors into a trained autoencoder model (.p) to compute the reconstruction error. Since the autoencoder has trained only on normal forest acoustic patterns, abnormal events, such as human voices or animal sounds, produce signicantly higher error values. This helps the system detect anomalies without needing to label every possible sound ahead of time.

   Second, a centroid-based approach categorizes detected anomalies as either human or animal activity. This classier compares the MFCC vector with stored centroid values that represent typical human and animal feature patterns; it assigns the closest match.

   Lastly, the EC2 instance also hosts the XGBoost re-risk prediction model (.pkl). This model uses environmental sensor data, primarily temperature and humidity, to decide whether the current forest conditions indicate a safe state or a higher risk of re. Running both models on the same EC2 instance keeps latency low and ensures synchronized processing for audio anomaly detection and re-risk estimation.

   Together, S3 and EC2 form a compact and efcient cloud-based processing pipeline where S3 serves solely as the storage point for MFCC data, and EC2 handles all machine learning inference operations.

   VI. METHODOLOGY

   The methodology adopted in this work follows a structured ve-phase pipeline encompassing data collection, preprocess-ing, model development, deployment, and visualization. Each stage is designed to ensure seamless integration between the sensor units deployed in forest environments and the cloud-based inference mechanisms responsible for both anomaly detection and re-risk assessment.

   1. Data Collection

      The data collection phase includes environmental mea-surements and acoustic recordings. Low-power environmental sensors on Raspberry Pi edge devices capture temperature, humidity, and oxygen levels. These parameters serve as input for re-risk classication and are sampled regularly to show changes in atmospheric conditions over time.

      At the same time, microphones mounted in the forest monitor environmental audio signals. These recordings pick up background forest sounds and specic events like human speech or animal calls. To support the anomaly detection system and offer enough variability for centroid comparison, open-source datasets of human and animal vocalizations are combined with real recordings.

      All acquired data, specically MFCC feature vectors for audio and scalar sensor readings for environmental parameters, are uploaded to an Amazon S3 bucket. S3 serves as the remote storage option because it is scalable, reliable, and works well with EC2-based inference systems. The structured storage format allows for timestamp-based retrieval and makes model deployment more efcient.

   2. Preprocessing

      Raw sensor and audio data go through several preprocessing steps to change and standardize the input before it is fed into machine learning models.

      For acoustic data, noise reduction uses spectral gating to lower background noise and improve relevant frequency com-ponents. Next, the audio stream is divided into short frames, allowing for focused analysis of brief events. Each frame is then changed into Mel Frequency Cepstral Coefcients (MFCCs). These represent the spectral envelope and capture important sound features. A xed number of MFCC coef-cients, specically 13 in this case, are extracted from each frame to create feature vectors that are suitable for detecting anomalies and classication.

      Environmental sensor data undergoes normalization to lessen the effects of different scales and measurement units. Min-max or z-score normalization is used for temperature, humidity, and oxygen concentration values to improve model stability and reduce bias.

      Label encoding is used during re-risk model training. In this process, environmetal readings are matched to binary output labels: safe or risk. This structured representation simplies supervised learning.

   3. Autoencoder Training

      The autoencoder acts as the main tool for detecting anoma-lies and trains using an unsupervised learning approach. It only uses audio recordings that represent normal forest conditions, which are free from human intervention or unusual animal activity, during training.

      The encoder compresses MFCC vectors into a smaller latent representation, and the decoder reconstructs the original signal. The training process reduces reconstruction loss, usually mean squared error. This helps the autoencoder learn the typical

      Fig. 1. Autencoder-Based Anomaly Detection Workow

      patterns of normal forest sounds. During inference, unusual acoustic events lead to higher reconstruction errors because of unfamiliar feature distributions.

      An anomaly threshold T is dened statistically:

      T = mean(error) + k · std(error)

      where k controls sensitivity. Frames that exceed this thresh-old are marked as anomalies. The next step is centroid-based classication, which identies whether the detected anomalies relate to human or animal activity. This process allows for contextual interpretation.

   4. XGBoost Training

      Forest-re risk prediction uses the XGBoost algorithm be-cause it can capture nonlinear interactions among environ-mental variables. The model takes normalized temperature, humidity, and oxygen readings as input and provides a binary risk score.

      We optimize the model with hyperparameters like learning rate, tree depth, gamma, and the number of estimators. We tune these hyperparameters through k-fold cross-validation to prevent overtting and ensure it generalizes well across seasonal changes.

      The output logits are assigned to discrete classes:

      0 Safe, 1 Fire Risk

      The trained XGBoost model is saved as a .pkl le, allowing integration with the cloud-based inference system.

      Fig. 2. XGBoost Fire Risk Prediction Workow

   5. Real-Time Deployment

   Real-time operation is achieved by hosting both models on an Amazon EC2 instance that runs a Flask backend. The backend periodically fetches MFCC feature les from S3 using boto3 and applies the autoencoder to calculate reconstruction error. If the anomaly threshold is exceeded, centroid classi-cation determines if the event involves a human or an animal. At the same time, environmental sensor data uploaded to S3 is retrieved and sent to the XGBoost model to evaluate re probability. The results from both inference pipelines are

   logged, timestamped, and sent to a web-based dashboard. The dashboard shows real-time anomaly scores, human/ani-

   mal classication results, and forest-re risk graphs over time. By combining anomaly detection with re-risk prediction, the system identies threats and their possible impact on

   humans or wildlife at the same time.

   Fig. 3. End-to-End Unied Forest Monitoring System Workow

5. Simulation and Experimental Results

   The proposed system was tested using Raspberry Pi edge devices, MEMS microphones, and environmental sensors. It relied on cloud processing through AWS S3 and EC2. The main goal was to assess how well acoustic anomaly detection worked alongside forest re risk prediction.

   1. Audio Anomaly Detection

      Ambient audio signals were captured in a forest-like setting. The system pulled MFCC features from these recordings and used a trained autoencoder model to spot anomalies. When unusual events, like human voices or animal calls, took place, the reconstruction error went beyond a set threshold. This allowed the system to recognize abnormal sounds. A centroid-based classier then separated human activity from animal activity, providing context for the detection.

   2. Forest Fire Risk Prediction

      Environmental sensors tracked temperature, humidity, and oxygen levels. This data was analyzed with an XGBoost model to predict re risk. The system detected changes in conditions that suggested potential re hazards and classied them in real-time.

   3. Dashboard Visualization

   The web-based dashboard displayed results for both audio anomaly detection and re risk predictions. It showed real-time alerts for anomalies and trends in re risk, allowing for ongoing monitoring of the forest environment.

   Overall, the simulation showed that the integrated system could function in real-time, enabling simultaneous tracking of human or animal intrusion and possible forest re dangers.

   Fig. 4. Real-time monitoring dashboard (a)

   Fig. 5. Real-time monitoring dashboard (b)

6. Conclusion

   In this work, we developed an integrated framework for real-time forest monitoring that combines acoustic anomaly detec-tion with machine learning-based forest re risk prediction. The system leverages MEMS microphones and environmental sensors to continuously capture forest sounds and atmospheric parameters, which are processed using an autoencoder and an XGBoost model, respectively. By analyzing MFCC features, the autoencoder effectively identies unusual acoustic events, distinguishing between human intrusion and animal activity through centroid-based classication. Simultaneously, the XG-Boost model assesses re risk using temperature, humidity, and oxygen level data, allowing for early warnings of potential forest res.

   The proposed framework was deployed in a cloud-based en-vironment using Amazon S3 for storage and EC2 for process-ing, ensuring scalability, reliability, and near real-time opera-tion. The results, visualized through a web-based dashboard,

   demonstrate the systems ability to provide continuous situ-ational awareness, simultaneously highlighting environmental anomalies and re hazards. This dual monitoring approach is especially valuable for forest conservation, enabling authorities to respond quickly to illegal human activity, wildlife distur-bances, and early signs of re, thereby mitigating environ-mental and economic damage. Overall, the research shows that combining IoT, machine learning, and cloud computing offers a practical, effective solution for modern forest monitoring challenges.

7. Future Scope

In the future, we can improve the proposed system for broader coverage and better intelligence. Adding more sensors over larger forest areas would boost spatial monitoring and detection accuracy. Integrating the system with UAV-mounted cameras or thermal imaging can offer visual proof of unusual events and re outbreaks. Using Edge AI on Raspberry Pi devices could lower delays and reliance on the cloud, allowing for quicker alerts in critical situations. Additionally, predictive models can use weather forecasts, types of vegetation, and terrain data to better anticipate the likelihood and spread of forest res. Energy-harvesting solutions, like solar-powered sensor nodes, could make the system more sustainable for long-term use in remote areas. Expanding the audio detection capabilities to identify a broader range of events, including different types of animals or specic human activities, would also provide more useful insights to forest authorities.

References

1. Kumari, P., & Saini, M. (2021). Anomaly Detection in Au-dio with Concept Drift using Adaptive Huffman Coding. arXiv. https://doi.org/10.48550/arXiv.2102.10515

2. Nunes, E. (2021). Anomalous sound detection with machine learning:

   A systematic review. arXiv. https://doi.org/10.48550/arXiv.2102.07820

3. Avazov, K., Hyun, A., Sami, S., Khaitov, A., Abdusalomov, A., & Cho, Y. (2023). Forest Fire Detection and Notication Method Based on AI and IoT Approaches. Future Internet 15, 61. https://doi.org/10.3390/15020061

4. Miriyala, T., Ragipati, K., Mahitha, J., & Reddy, V. (2018). IoT-based forest re detection system. International Journal of Engineering & Technology, 7, 124. https://doi.org/10.14419/ijet.v7i2.7.10277

5. Bharadwaj, B. S., Bharath, A., Kiran, M., Pooja, M., & Sahana,

   S. (2023). Earlier detection of forest re using IoT. International Journal of Emerging Trends in Engineering Research, 11, 125135. https://doi.org/example-doi

6. Mehta, V., Sharma, S., & Abirami, G. (2024). Detecting anomalies in IoT sensor data using machine learning algorithms. Proceedings of Machine Learning Applications, 9. https://doi.org/10.1201/9781003559085-9

7. Han, S., Wu, Q., & Yang, Y. (2022). Machine learning for Internet of Things anomaly detection under low-quality data. International Journal of Distributed Sensor Networks, 18, Article 155013292211337. https://doi.org/10.1177/15501329221133765

8. Thapa, S., Chitale, V., Pradhan, S., Shakya, B., Sharma, S., Regmi, S., Bajracharya, S., Adhikari, S., & Dangol, G. (2021). Forest Fire Detection and Monitoring. Springer. https://doi.org/10.1007/978-3-030-73569-28

9. Atmosphere. (2024). 15(11), 1284.

https://doi.org/10.3390/atmos15111284