10:45 AM - 12:15 PM
[STT39-P15] Researching on the usefulness of SAR coherence for deep learning-based flood area detection
Keywords:Synthetic Aperture Radar, deep learning, PALSAR-2, flood, coherence
Floods are one of the largest natural disasters that cause extensive damage to people's lives and property. Satellite observation acquires information on the ground surface over a wide area at once, and is an important source of information for quickly assessing the damage and evaluating the impact of floods. Synthetic aperture radar (SAR) observes the reflection of microwave that is not affected by nighttime or weather conditions; the information obtained from SAR includes intensity and coherence images, which are considered effective for flood monitoring in the following ways Since water areas do not reflect microwaves back to satellites, images appear darker. That is used to compare intensity images of flooded and non-flooded areas, in order to identify areas with a large decrease in brightness as flooded areas. SAR coherence images can also be used to identify floods that occur over different types of land cover, such as urban areas and vegetated areas.
Various methods for estimating flooded areas from SAR have been investigated, and methods based on deep learning have gained attention in recent years. Deep learning methods have been widely used to estimate flooding areas by semantic segmentation using pairs of intensity images of flooded and pre-flood areas. On the other hand, there are few cases in which coherence images, have been utilized in deep learning. while it is considered to be useful in flood detection,
In this paper, we compare the performance of a deep learning method trained by adding coherence images to the model input in addition to pairs of SAR flood and pre-flood intensity images, and that only with pairs of flood and pre-flood ones. We show the coherence images are useful for deep learning in determining flooded areas, comparing the performance of the two models. ALOS-2/PALSAR-2 intensity images of flood and pre-flood area are used, and coherence images are generated from single look complex (SLC).
Accuracy evaluation experiments were conducted by comparing conventional methods such as k-mean clustering, deep learning method with flood and pre-flood intensity images, and deep learning method with flood and pre-flood intensity images plus coherence images. The validation target is a flood that occurred in Joso, Ibaraki, Japan, in September 2015. The accuracy was evaluated in terms of true positives (TP), true negatives (TN), false positives (FP), false negatives (FN), and Intersection over Union (IOU). The results showed that in the estimation of inundation areas using deep learning, the use of coherence images as additional information significantly increased IOUs and decreased FPs due to the detection of existing water areas, thereby improving accuracy. This suggests that coherence images contain features that are useful for deep learning to determine inundation areas.
Various methods for estimating flooded areas from SAR have been investigated, and methods based on deep learning have gained attention in recent years. Deep learning methods have been widely used to estimate flooding areas by semantic segmentation using pairs of intensity images of flooded and pre-flood areas. On the other hand, there are few cases in which coherence images, have been utilized in deep learning. while it is considered to be useful in flood detection,
In this paper, we compare the performance of a deep learning method trained by adding coherence images to the model input in addition to pairs of SAR flood and pre-flood intensity images, and that only with pairs of flood and pre-flood ones. We show the coherence images are useful for deep learning in determining flooded areas, comparing the performance of the two models. ALOS-2/PALSAR-2 intensity images of flood and pre-flood area are used, and coherence images are generated from single look complex (SLC).
Accuracy evaluation experiments were conducted by comparing conventional methods such as k-mean clustering, deep learning method with flood and pre-flood intensity images, and deep learning method with flood and pre-flood intensity images plus coherence images. The validation target is a flood that occurred in Joso, Ibaraki, Japan, in September 2015. The accuracy was evaluated in terms of true positives (TP), true negatives (TN), false positives (FP), false negatives (FN), and Intersection over Union (IOU). The results showed that in the estimation of inundation areas using deep learning, the use of coherence images as additional information significantly increased IOUs and decreased FPs due to the detection of existing water areas, thereby improving accuracy. This suggests that coherence images contain features that are useful for deep learning to determine inundation areas.