Wheat Field Fire Smoke Detection from UAV Images Using CNN-CABM

Abstract

Wheatfield fires all seriously threaten farmers income, agricultural resources, and wildlife habitats. These fires have become much more frequent nowadays, due to climate change and activities of humans in farming areas. Since smoke detection is the initial step in early fire detection, action must be taken right away. We proposed an improved CNN model that can detect smoke in wheat field fires to address this problem. Features are extracted from UAV images using CNN. CNN is a deep learning model that automatically adjusts to spatial information in input images. As part of this method, 3900 UAV images were collected to illustrate smoke from wheat field fires. The suggested method divides Wheatfield fire images into four categories: high, medium, low, and no fire. The dataset displays smoke from burning wheat. Included are the following categories of smoke photos:(a) a lot of smoke and fire; (b) a lot of smoke and fire (c) Mild smoke and fire. There are three types of fire and smoke: (d) medium; (e) small; and (f) small.The CNN-CBAM and the model that combines CBAM and CNN were trained and validated on this dataset. One strategy for accomplishing this is to apply the channel along with the pores to the important aspects of the images, which is known as the attention mask method. The proposed task is to extract features for smoke patch characteristics of wheat field fires. Also, the CNN design incorporates disconnected heads that extract relevant data from a variety of sources. Image categorization and feature extraction using convolutional neural networks comprise the two components of the suggested CNN-CABM approach. The usefulness of this suggested method was validated through testing on our wheat field fire smoke dataset, outperforming the prior model and achieving an accuracy of up to 0.92. The study found that the combination of CBAM and CNN is particularly effective for the task of detecting fires in wheat fields, indicating a new bright side to the area of early warning systems. To cross-validate the findings of the same dataset using basic SVM and CNN, we determined that the suggested model CNN-CABM outperforms simple SVM and CNN to produce better results. The next steps will be to improve this versions stability in the face of shifting multiplicative nuisances, as well as to increase speed and precision, according to those who contributed.