Hyperspectral satellite imagery detects wildfires

In recent years, climate change and other human-related environmental issues have garnered significant attention in scientific literature. Historically, wildfires were largely detected by monitoring vast areas from fire observation towers and using simple devices such as fire detectors. However, this method was not very accurate, and its effectiveness could be affected by human fatigue accumulated during long observation periods. On the other hand, alternative sensors used to detect gases, flames, smoke, and heat emissions typically require extended measurement times. Furthermore, due to the limited measurement range of these sensors, only a large number of sensors could cover large areas. The rapid development of target recognition, deep learning, and remote sensing technologies has provided us with new methods for finding and tracking wildfires.

1. Main technical classifications

Currently widely used technologies can be broadly categorized into three types, capable of identifying or observing active fires or smoke in real-time or near real-time: ground-based, airborne, and satellite systems. These technologies are typically combined with visible light, infrared, multispectral, or hyperspectral sensors; once data is collected, it can be processed using applicable artificial intelligence (AI) algorithms (usually machine learning (ML) methods). These technologies rely on extracting human-made features or powerful AI methods to detect wildfires in their early stages and simulate the behavior of smoke and fire. Different types of fire detection methods are illustrated in Figure 1.

Figure 1. Fire detection methods

In recent years, extensive research has been conducted attempting to identify wildfires using satellite imagery, primarily due to the large number of satellites launched and the associated decreasing costs. Specifically, a suite of satellites for Earth observation (EO) has been developed (e.g., Planet). Satellites are typically categorized based on their orbits, each with its own advantages and disadvantages. Table 1 lists the most important categories.

Table 1. Satellite Categories

Most geostationary orbit (GEO) satellites in low polar SSO orbits have precise altitude and inclination estimates to ensure that spacecraft observe the same scene at the same angle of sunlight each time, and that the shadows are identical with each pass. Sun-synchronous satellite data has high spatial resolution but low temporal resolution (LandSat-7/8 has an 8-day repeat cycle, while Sentinel 2A/2B has a 2-3 day repeat cycle in mid-latitude regions), while GEO satellites have lower spatial resolution but higher temporal resolution. Therefore, they are not effective at detecting active wildfires in real time; instead, they are better suited for less time-sensitive tasks, such as estimating burned area. Geostationary orbit satellite systems are able to detect wildfires because they can see large areas. Most satellites that photograph Earth use multispectral imaging sensors and are located in geostationary or near-polar sun-synchronous orbits.

Table 2. List of some relevant remote sensing satellite systems and their characteristics

2. Current detection methods

Based on hyperspectral data, preliminary direct information can be obtained by examining individual bands. For example, smoke can be identified by examining the visible near-infrared (VNIR) band, while active wildfires can be retrieved by examining saturated pixels in the short-wave infrared (SWIR) band. In fact, within the 2000nm to 2400nm band, the signal easily saturates when observing active wildfires because the signal captured from Earth is greater than the signal from the sun (since wildfires act as active power emitters). However, analyzing the entire spectral features using convolutional neural networks can avoid errors and improve the reliability of classification.

The multi-class classification CNN model is shown in Figure 2. The input pixel spectrum of the hyperspectral data contains SWIR and VNIR channels. Therefore, it is a hypercube containing C = 2^34 elements (after removing some useless raw data from the input hypercube). A one-dimensional convolutional layer with a kernel of 3, an n1 = 112 filter, the same padding, a ReLU activation function, and an l2 kernel regularizer is the first hidden layer. After the convolutional layer, there is a max-pooling layer with a pooling size of 2 and a stride of 2 (note that n2 = n1 as shown in Figure 2). The result of this max-pooling is then passed through a flattening layer and then connected to a fully connected layer of 128 units (with a ReLU activation function). The last layer is a dense unit layer with a SoftMax activation function for multi-class classification. It is worth noting that the metrics in the C1 and C2 figures are easy to evaluate and depend on the network architecture. We use the Adam optimizer and the classification cross-entropy loss function to train the model.

Figure 2. Multi-class classification CNN model