Abstract:

Objective: Among the current forest fire detection methods, the research based on deep learning is the most active. However, when dealing with real scenes, the model detection effect is often poor due to insufficient samples of forest fire, imbalanced distribution of categories and weak expression ability of scenes, et al. To alleviate this problems, we presented a newly developed data augmentation method called self-adaptive mix augmentation (SMA). Method: This study takes UAV(unmanned aerial vehicle) forest fire images collected from Chongli district, Zhangjiakou city, Hebei Province as the research object. The work is as follows: 1) Preprocess the UAV video to construct the original data set. 2) Use methods such as category statistics and annotation box centralization to analyze and find out the problems existing in the data, such as: nimiety of small targets, unbalanced target distribution and scattered annotation box size. 3) For the problem of imbalanced categories, we introduced self-adaptive parameters to achieve the dynamic adjustment of samples. 4) In order to ensure the effectiveness of cross-sample information fusion, IOA(intersection over aim) was proposed as a judgment threshold to give a reasonable reference value. 5) According to the principle of control variables, we designed 12 ablation experiments with UAV data as samples, and compared the results of forest fire detection of original sample, ordinary data augmentation, Mosaic and SMA methods in SSD, YOLOv3 and YOLOv4 mainstream algorithms, respectively. 6) MAP(mean average precision) was selected as the index to evaluate the results of different data augmentation methods in the same algorithm. Result: The results of ablation test showed that in SSD, YOLOv3 and YOLOv4 algorithms, the MAP performance of SMA method was 48.16%, 82.02% and 67.79%, compared with the original data, it increased by 12.14%, 11.50%, 36.83%, compared with traditional random augmentation, it increased by 11.95%, 4.86% and 16.33%, compared with the Mosaic method, it increased by 1.06%, 18.24%, and 1.79%. Conclusion: Traditional data augmentation methods did not fully explore the information contained in samples in forest fire data set. The SMA method in this study introduces self-adaptive parameters to alleviate the problem of sample imbalance, and the introduction of IOA achieves cross-sample fusion. The experimental results showed that the SMA method improves the MAP performance of SSD, YOLOv3 and YOLOv4 algorithms compared with the traditional method, which proves the effectiveness of SMA method on forest fire data set.

Key words: forest fire prevention, smoke and fire detection, the sample balance, adaptiveness, data augmentation, information fusion