SmokeNet: Satellite Smoke Scene Detection Using Convolutional Neural Network with Spatial and Channel-Wise Attention

A variety of environmental analysis applications have been advanced by the use of satellite remote sensing. Smoke detection based on satellite imagery is imperative for wildfire detection and monitoring. In order to detect the fire smoke in satellite imagery, we propose the image-level CNN-based method (SmokeNet) which utilizes spatial and channel-wise attention.

We also construct a new satellite imagery dataset (USTC_SmokeRS) based on Moderate Resolution Imaging Spectroradiometer (MODIS) data for smoke scene detection. The USTC_SmokeRS dataset consists of 6225 RGB images from six classes of cloud, dust, haze, land, seaside, and smoke. This dataset has been released as the benchmark dataset for smoke scene detection with satellite remote sensing.

Problem & Motivation

One important challenge for detecting fire smoke in satellite imagery is the similar disasters and multiple land covers. The commonly used smoke detection methods mainly focus on smoke discrimination from a few specific classes, which reduces their applicability in different regions of various classes. In addition, there is no satellite remote sensing smoke detection dataset so far.

To this end, we construct the USTC_SmokeRS dataset and integrate more smoke-like aerosol classes and land covers in the dataset, for example, cloud, dust, haze, bright surfaces, lakes, seaside, vegetation, and so on. To improve the effectiveness and applicability of the method, we propose a new convolution neural network (CNN) model, dubbed SmokeNet, which incorporates the spatial and channel-wise attention in CNN to enhance feature representation for scene classification.

USTC_SmokeRS Dataset

The USTC_SmokeRS dataset contains a total of 6225 RGB images from six classes: cloud, dust, haze, land, seaside, and smoke. Each image was saved as the ".tif" format with the size of 256 × 256. The spatial resolution is 1 km. The number of images in each class is presented in the following table, and some example images of these six classes are shown in the figure.

SmokeNet Model

The proposed SmokeNet model incorporates the spatial and channel-wise attention in CNN to improve classification capacity. The model is combined with the spatial, channel-wise attention block, and the residual attention module. This makes the SmokeNet model can not only dynamically recalibrate the channel-wise features and generate the attention-aware features, but also optimize the representation of spatial information using the proposed spatial attention mechanism.

Related Publication

Ba, R., Chen, C., Yuan, J., Song, W., & Lo, S. (2019). SmokeNet: Satellite Smoke Scene Detection Using Convolutional Neural Network with Spatial and Channel-Wise Attention. Remote Sensing, 11(14), 1702.

Download: [Paper] [Dataset: Link/链接]