Wildfire CNN: An Enhanced Wildfire Detection Model Leveraging CNN and VIIRS in Indian Context

2.1. Contextual Method based Wildfire Detection System

The contextual method using VIIRS data comprises an initial threshold for identifying the fire pixel, a circumstantial test for confirming fire in likely fire pixel and threshold to turn down false alarms. VIIRS I4 band is the essential driver of the wildfire detection which provides higher performance for the VIIRS 375 m current fire data when compared to the VIIRS 750 m standard fire product utilizing day and night time data. The VIIRS 375 m fire data achieved greater potential with improved consistency than MODIS fire detection data [5]. This work has limitations in showing the difficulty of low confidence with a high rate of wildfire detections. Contextual wildfire detection algorithm has some weaknesses such as determining static threshold for likely fire pixels, disregarding the perceptiveness of test conditions and ignoring small and cold fires. Table 2 shows the comparative analysis of wildfire detection approaches in existing work with the results and limitations.

Waigl et al. [6] developed the modified contextual wildfire detection algorithm called as VIIRS I-band Fire Detection Algorithm for High Latitudes (VIFDAHL) using the data obtained during Alaska's boreal wildfires on 2016 fire period. The proposed VIFDAHL algorithm successfully classifies the fire pixels as high or low intensity and also removes the duplicate detections due to the bowtie effect in VIIRS data. This method accurately determined 30–90% more fire pixels in VIIRS global fire products than MODIS.

Zhang et al. [7] presented a VIIRS-IM (combination of 750 m M-Band and 375 m I-Band data) for wildfire detection. Eastern China area was considered for this study and showed that the I-Band VIIRS data facilitates smaller active wildfire detection. The study results are compared with Aqua-MODIS and the existent VIIRS I-Band wildfire detections. VIFDAHL captures the fire distri- bution, accurately classifies high and low intensity fires, and correctly detects 30–90% higher fire pixels in comparison with MODIS data. Proposed VIIRS-IM considered the FRP metrics and experimental results show that the lowest FRP pixels in the agricultural burning area of eastern China.

Fire Radiative Power (FRP) is the release rate of radiative fire energy and has been applied for estimation of fire intensity, calculation of smoke injection height, severity of fire and post-fire forest productivity. FRP can be calculated by using the Eq. (1).

(1)

where L_4f and L_4b are radiances of fire and background pixels at 4 µm, A_fp is the fire pixel area, σ is the Stefan–Boltzmann constant (5.67×10-8 Wm⁻² K⁻⁴), τ₄ is the transmittance of atmospheric at 4 µm, and a is an empirical constant (3.0 ∗ 10⁻⁹Wm⁻²sr⁻¹µm⁻¹k⁻⁴).

VIIRS is keen to visible light from wildfires during nighttime. Resampling by pixel radiances of the Day-Night Band (DNB) from M-band footprint pixels to increase nighttime wildfire detection has been presented [8]. A study of the 2018 California Camp Fire has been taken and shown that the method provides higher accuracy during the nighttime fire.

The wildfire detection and performance based on FRP estimation of VIIRS and MODIS fire regimes are analyzed [9]. This study surveyed VIIRS data in Northeastern Asia between 2012–2017 where wildfires and burning of agricultural land are prevalent. Wildfire detection and FRP in MODIS and VIIRS data were compared using the simultaneous observations in the overlapping orbits. The experimental result shows that across land cover categories, the VIIRS is stronger in wildfire detection and high fire omission of MODIS in the area's lands of low-biomass. The detection of false wildfire pixels was lower in VIIRS than in MODIS, but the commission errors were similar. Finally, the low-biomass areas of VIIRS show that is effective for low-intensity wildfire areas, while for high-intensity wildfires MODIS accurately retrieves FRP with dense smoke and solid atmospheric absorption.

In order to overcome the issue of ignoring small fires by contextual method, a new weighted wildfire detection algorithm has been proposed to accurately detect small fires on VIIRS [10]. The mid-infrared channel has been used for wildfire identification on the East Asia region. The crop burning wildfires that occur in East Asia have been detected by adaptive threshold using bubble sort method based on the small fires radiation characteristics. The results show that the accuracy value of 53.85% in summer and 73.53% in winter. However, the missing VIIRS wildfire points have to be analyzed.

Spatio-temporal contextual brightness temperature prediction model (STCM) has also been proposed for wildfire detection. San Diego in Southern California, America has been surveyed for this research study and experimental results show that RMSE is 12.54% lower than that of contextual method and 9.12% lower than that of Temporal-Contextual method. This method uses MODIS products to detect wildfire detection, but it fails to discuss small fires in the wildfire [11].

The performance of MODIS and VIIRS wildfire detection in various land covers has been evaluated in Turkey on the fire data from 2015 and 2019 [12]. The VIIRS achieves accuracy from 1.3% to 25.6% for five various land cover types like herbaceous vegetation, croplands, closed and open forests and shrublands. However, MODIS and VIIRS have limitations of wildfire detections especially less than 10 ha.

The wildfire detection in the Kayamaki Wildlife study area has also been sureveyed using contextual based algorithms. This method analyzed various contextual based algorithms like IGBP, Giglio, Extended, Dynamic, and Giglio and Extended on MODIS data and IGBP algorithm detects 86% wildfire with a 14% error rate. The accuracy can be improved further for the better performance.

The statistical data of the surrounding areas has mostly influenced the use of contextual approaches in wildfire detection. Nevertheless, the threshold may vary depending on the type of land cover, the seasons, and the climate. The performance of the wildfire detection system is impacted by clouds and thick smoke, making it susceptible to omissions and false positives.

2.2. Machine Learning based Wildfire Detection System

Nowadays, Machine Learning (ML) techniques are often useful for finding relationships, detecting and predicting wildfire growth. Using ML algorithms for detecting, rapid monitoring of active fires, and early prediction of active fires could help the current wildfire detection system for further improvement [14].

Muhammad et al. [15] proposed a CNN to produce appropriate images of the fire using surveillance videos. Still, these methods give less performance for huge area wildfire detection. Ba et al. [16] proposed a SmokeNet to improve the scene classification by combining the space and channel attention direction into CNN. The experi- mental results indicate the proposed method classifies smoke scenes based on MODIS data. Gargiulo et al. [17] designed the data fusion method on CNN for wildfire detection of Sentinel-2 images. The precision and recall rate of the proposed method is 84.14% and 56.42% respectively. In this method, the weights of the CNN have fine-tuned based on the geographic study area (Vesuvius, Italy).

Pinto et al. [18] utilized the U-net architecture with Long Short-Term Memory (LSTMs) and CNN in VIIRS 750 m bands data. This model uses VIIRS 750 m Red, middle-infrared (MIR), near-infrared (NIR) bands with VIIRS 375 m active fires for training the model. The authors studied regions of Brazil, California, Portugal, Australia and Mozambique with accuracy of 90.6%. Additionally, Bushnaq et al. [19] used Unmanned Aerial Vehicle (UAV) images or videos to detect wildfires. Rashkovetsky et al. [20] investigated fire-affected areas from satellite image of the infrared, visible and microwave domains. CNN based U Net architecture has been considered with multisensory data (Sentinel-1, Sentinel-2, Sentinel-3 and MODIS) for detecting wildfires. Combination of Sentinel-2 and Sentinel-3 data achieved highest accuracy of 96% in cloudy condition.

Khan, Somaiya, and Ali Khan [21] demonstrated the FFireNet for discriminating fire and non-fire images using MobileNetV2 model. The experimental results show that, 99.47% of recall, 98.42% of accuracy, 1.58% of error rate and 97.42% of precision for the dataset generated by intelligence surveillance system. Ahmed S. Almasoud [22]. combined DL with Computer Vision technology for detecting and monitoring wildfires using IWFFDA-DL model (Intelligent Wild Forest Fire Detection and Alarming system DL). CNN with BLSTM (Bidirectional Long Short Term Memory) for detecting fires and bacterial foraging optimization (BFO) algorithm has been used for enhancing the detection performance. The result analysis showed that the precision of 97.88%, recall of 99.46%, accuracy of 99.56%, F-score of 98.65% and MAE of 1.36% with globally collected fire dataset with condition of smoke less than 50 dB/m, normal fire status, heat value is less than 55°C and flame less than 180 μm.

Abdusalomov et al. [23] proposed an Detectron2 Model with precision of 99.3% for the large custom dataset to address the challenges of day and night fire, light and shadows in the images. Thangavel et al. [24] studied one dimensional CNN using PRISMA hyperspectral imagery to detect wildfires in Australian region and accuracy of the system achieved 97.83% for the validation dataset.

The determination of fire incidents and hotspot actions within the four national parks in Thailand has also been studied using MODIS data. This paper suggests the early detection of forest and land fires through the implementation of Artificial Intelligence (AI) [25]. However, this study aimed to analyze and monitor the effective wildfire detection system.

Priya et al. [26] demonstrated the use of ML to detect vegetation changes and evaluate recovery using VIIRS data based on post-fire satellite imagery with a training error of 0.075.

There are still certain unresolved research issues from the talks above. The literature review revealed a research gap in contextual techniques, where the threshold may change depending on the kind of land cover and the various climate and seasonal factors. As a result, this approach is easily affected by clouds and dense smoke and is susceptible to omissions and false positives. Since ML based techniques have significant adaptability, learning capacity, and superior portability, they offer high accuracy and minimal false alarms [27]. It is still difficult to accurately detect wildfires, and current techniques for doing so with remote sensing sensor data need to be improved [28].

3.1. Proposed Wildfiredetect CNN Model

The steps for detecting wildfire pixels using VIIRS data are shown in Fig. (1).

Wildfire detection algorithm has the following major steps:

3.1.4. Fire Pixel Detection

Potential fire pixels are detected using specialized wildfire detection algorithms.

Wildfiredetect CNN model is proposed to improve the accuracy of wildfire detection using VIIRS data. When compared to contextual methods, VIIRS data can easily eliminate false positives and omission errors and give high resolution for the easy detection of fire points.

Fig. (2) shows the workflow of the Wildfiredetect CNN model. From fire spot location data from VIIRS fire product data, the position and time of the data is taken and a grid of 5 × 5 sizes is formed at the centre of each pixel. Surrounding information of the pixels is calculated by the average and standard deviation of each grid band. In the training data set, the 1:2 ratio of fire and non-fire training points were considered. The characteristics of the fires were learned by the network fully and classified correctly according to the data. In this study, data of fire and non-fire spots are randomly added to the training set and 15% is considered as the validation set for testing the accuracy. The algorithm has a feature extraction method and a classification method. The feature extraction method extracts features from the input data and fully connected layer output the spot is a fire or non-fire.

Fig. (3) shows the proposed Wildfiredetect CNN model. The proposed model consists of 5 convolutional layers, 5 Maxpool layers and 2 fully connected layers to extract features. First convolution layer has 96 kernels of size (11 × 11 × 3) and a stride of 4 pixels followed by Maxpool layer to reduce data’s complexity, second convolution layer has the 256 kernels of size (5 × 5 × 64) with a stride of 2 followed by second Maxpool layer, third convolution layer has 384 kernels of size (3 × 3 × 256) without a pooling procedure and the rest of the convolution layers use filters with kernels of 384 and 256, respectively with a stride of 1. The pooling layer uses 3×3 filters. Finally, 2 fully connected layers and the output layer has Softmax function to distinguish fire/non-fire. In addition to achieving non-linearity in neural networks, the variation of ReLu is the leaky ReLu activation function used, which introduces a small positive slope for negative inputs for addressing the problem of “dead neurons” in ReLu activation function.

5.1. Evaluation Metrics

The performance metrics like Recall Rate (R_r), Precision Rate (P_r), Commission Error (C_e), F-measure (F_m) Accuracy (AC) and Omission Error (O_e) were considered to test the proposed Wildfiredetect CNN model performance. Precision Rate (P_r) declares that the truly predicted fire spots of this model and is calculated by using Eq. (2).

(2)

TP_F denotes correctly classified fire spots, TN_F denotes false fire spots, FP_F denotes pixels misclassified as fire spots, FN_F represents inaccurately classified as non-fire spots. If the reliability of the predicted model is higher, then the PR value is higher.

The C_e denotes the number of the predicted fire spots is wrong. The higher C_e value shows the wrong fire spots recognized by the model and is calculated by using Eq. (3).

(3)

The R_r defines the fire spots correctly identified in the model. If the R_r value is greater, the lesser the fire spots skipped and is calculated by using Eq. (4).

(4)

The O_e shows that the original data fire spot and the extent to it is missed out. The higher the O_e value model, the lower the comprehensiveness. The O_e value is calculated by using Eq. (5).

(5)

The AC is calculated by the ratio of correct predictions to the model predictions and is calculated by using Eq. (6).

(6)

F_m value evaluates the performance of the model and is calculated by using Eq. (7).

(7)

5.2. Experiment Result and Analysis

Fig. (5) shows the result of the proposed Wildfiredetect CNN model in the Tamil Nadu state region. The proposed Wildfiredetect CNN model is compared with two types of algorithms such as contextual algorithms and ML based algorithms. The existing method based on these algorithms has been used for testing the proposed method efficiency. Table 4 shows the comparison of an existing method such as contextual and ML based algorithms with the proposed Wildfiredetect CNN model.

From Table 4, it is observed that the contextual algorithms and an improved Contextual algorithm achieve low accuracies compared to the ML based algorithms. These methods have shown lower performance for detecting active fires when compared to other methods for the use of threshold methods to find the active fires in the dataset. The proposed Wildfiredetect CNN model achieves 26.17% higher accuracy rate than the improved contextual algorithm. Also compared to other ML algorithms, proposed Wildfiredetect CNN model gives higher accuracy in wildfire detection system and it is shown in Fig. (6).

The performance metrics like P_r, R_r, C_e, O_e, AC and F_m were used as the indicators of the proposed Wildfiredetect CNN model performance and the results are shown in Table 5. The proposed Wildfiredetect CNN model achieves the highest accuracy score of 99.7% with a lower omission error of 0.3%.

Confusion matrix gives the predictive analysis of the wildfire dataset classification for the study area. The proposed model shows good results for the wildfire classification problem. The proposed model classifies the fire spot with 99.79% P_r, 99.69% R_r, 0.2% C_e, 0.3% O_e, 99.7% AC and 99.73% F_m values, respectively for training data. Confusion matrix of the proposed model for the testing data is shown in Fig. (7). The proposed results in 227 TP_F and 18 TN_F while the wrong classified active fire spots are 4 FP_F and 1 FN_F, respectively for testing data as shown in Table 6.

Fig. (8) depicts the training and validation accuracy analysis of proposed model. The model runs for 100 epochs and the accuracy values of each epoch are analyzed for the training and validating data on the dataset. The experimental results show that the highest accuracy of 98.3% is achieved at 100 epochs.

5.3. Experiment Result Analysis on Alaska Forests Dataset

In this section, the proposed model is tested with Alaska forest datasets to prove its effectiveness in detecting small fires.

Coffield et al. [32] used machine learning models to predict the small fires in substantial occurrence of wildfires in Alaska forests. Decision trees achieve higher accuracy in detecting small fires in the study area when compared to other machine learning models such as Random Forests (RF), K-Nearest Neighbours (KNN), Gradient Boosting (GB) and Multi-Layer Perceptrons (MLP). The active fire data was collected from MODIS sensor over the year 2001–2017 and it contains fire spots from upto 1168. The Alaska dataset was considered to test the performance of the Wildfiredetect CNN model. The proposed model achieves 99.4% accuracy which is higher than the machine learning models such as RF, DT, KNN, GB and MLP. The proposed model accuracy for small fires in the Alaska forests are 45.3%, 43.6%, 48.7%, 45.4% and 47.6% higher than RF, KNN, GB and MLP models respectively. Fig. (9) depicts the overall performance evaluation metrics comparison of various classification models with the proposed model for the Alaska dataset [33-40].

In this research, 582 training samples (50%), 291 testing samples (25%) and 291(25%) validating samples (10%) were considered to prove the proposed Wildfire CNN model performance on Alaska dataset by considering the six quality indices such as P_r, R_r, C_e, O_e, AC and F_m. Table 7 shows the performance comparison of the proposed model with various existing ML algorithms. The proposed model higher performance in detecting active fire spot and achieves lower error rate for effective fire spot detection. Table 8 shows that the performance metrics result of the proposed model for testing data in Alaska dataset. The proposed model achieves 99.2% P_r, 100% R_r, 0.7% C_e, 0% O_e, 99.3% AC and 99.69% F_m values, respectively.

Fig. (10) depicts the Confusion matrix of the proposed model for the testing data in Alaska dataset. The proposed model results in 279 TP_F and 10 TN_F while the wrong classified active fire spots are 2 FP_F and 0 FN_F, respectively.

The experimental analysis provides the effectiveness of the proposed model for detecting wildfires in the dataset. The proposed model also demonstrates that accurately detecting small fires in Alaska forests dataset when comparing other ML algorithms. Fig. (11) depicts the training and validation accuracy analysis of the proposed model. The model runs for 200 epochs and the accuracy values of each epoch are analyzed for the both training and validating data on the Alaska dataset. The analysis on epochs shows that, at 200 epochs the highest accuracy value of 99.3% achieved.