The following work was done by me and Dr. Shay Strong, while I was a data engineer consultant under the supervision of Dr. Strong at OmniEarth Inc. All the work IP rights belong to OmniEarth. Dr Strong is the Chief Data Scientist at OmniEarth Inc.
以下要介绍的工作是我在OmniEarth公司做数据工程师的时候和Shay Strong博士共同完成的工作。工作的知识产权归OmniEarth公司所有，我的老板Shay Strong博士是OmniEarth公司的数据科学家团队的领头人。
A wildfire had been burning in the Great Smoky Mountains of Tennessee and raced rapidly northward toward Gatlinburg and Pigeon Forge between late Nov. and Dec. 2nd, 2016. At least 2000 buildings were damaged or destroyed across 14,000 acres of residential and recreational land, while the wildfire also claimed 14 lives and injured 134. It was the largest natural disaster in the history of Tennessee.
After obtaining 0.4 m resolution satellite imagery of the wildfire damage in Gatlinburg and Pigeon Forge from Digital Global, OmniEarth Inc created an artificial intelligence (AI) model that was able to assess and identify the property damage due to the wildfire. This AI model will also be able to more rapidly evaluate and identify areas of damage from natural disasters from similar issues in the future.
从Digital Global获得大约为0.4米分辨率的高分辨率遥感图像（覆盖了火灾发生的Gatlinburg 和Pigeon之后）我们建立了人工智能模型。该人工智能模型可以帮助我们快速定位和评估火宅受灾面积和损毁程度。我们希望该模型未来可帮助消防人员快速定位火灾险情和火灾受损面积。
Fig 1. The final result of fire damage range in TN from our AI model. 该图是通过人工智能模型生成的火灾受灾范围图。
1. Artificial intelligence model behind the wildfire damage火灾模型背后的人工智能
With assistance from increasing cloud computing power and a better understanding of computer vision, more and more AI technology is helping us detect information from trillions of photos we produce daily.计算机图像识别和云计算能力的提升，使得我们能够借助人工智能模型获取数以万计甚至亿计的照片地图等图片中获取有用的信息。
Before diving into the AI model behind the wildfire damage, in this case, we only want to identify the differences between fire-damaged buildings and intact buildings. We have two options: (1), we could spend hours and hours browsing through the satellite images and manually separate the damaged and intact buildings or (2) develop an AI model to automatically identify the damaged area with a tolerable error. For the first option, it would easily take a geospatial technician more than 20 hours to identify the damaged area among the 50,000 acres of satellite imagery. The second option poses a more viable and sustainable solution in that the AI model could automatically identify the damaged area/buildings less than 1 hour over the same area. This is accomplished by image classification in AI, using convolutional neural networks (CNN) specifically, because CNN works better than other neural network algorithms for object detection and recognition from images.
Fig 2. Our AI model workflow. 我们的人工智能模型框架。
Artificial intelligence/neural networks are a family of machine learning models that are inspired by biological neurons of our human brain. First conceived in the 1960s, but the first breakthrough was Geoffrey Hinton’s work published in the mid-2000s. While our human eyes work like a camera seeing the ‘picture,’ our brain will process it and be able to construct the objects we see through the shape, color, and texture of the objects. The information of “seeing” and “recognition” is passing through our biological neurons from our eyes to our brain. The AI model we created works in a similar way. The imagery is passed through the artificial neural network, and objects that have been taught to the neural network are identified with certain accuracy. In this case, we taught the network to learn the difference between burnt and not-burnt structures in Gatlinburg and Pigeon Forge, TN.
2. How did we build the AI model
We broke down the wildfire damage mapping process into four parts (Fig 1). First, we obtained the 0.4m resolution satellite images from Digital Globe (https://www.digitalglobe.com/). We created a training and a testing dataset of 300 small images chips (as shown in Fig 3, A and B) that contained both burnt and intact buildings, 2/3 of which go to train the AI model, CNN model in this case, and 1/3 of them are for test the model. Ideally, the more training data used to represent the burnt and non-burnt structures are ideal for training the network to understand all the variations and orientations of a burnt building. The sample set of 300 is on the statistically small side, but useful for testing capability and evaluating preliminary performance.
|Fig 3(A). A burnt building||Fig3(B). Intact buildings|
Our AI model was a CNN model that built upon Theano (GPU backend) (http://deeplearning.net/software/theano/). Theano was created by the Machine Learning group at the University of Montreal, led by Yoshua Bengio, who is one of the pioneers behind artificial neural networks. Theano is a Python library that lets you define and evaluate mathematical expressions with vectors and matrices. As a human, you can imagine our daily decision-making is based on the matrices of perceived information as well, e.g. which car you want to buy. The AI model helps us to identify which image pixels and patterns are fundamentally different between burnt and intact buildings, similar to how people give a different weight or score to the car brand, model, and color they want to buy. Computers are great at calculating matrices, and Theano brings it to next level because it calculates multiple matrices in parallel, and so speeds up the whole calculation tremendously. Theano has no particular neural network built-in, so we use Keras on top of Theano. Keras allows us to build an AI model with a minimalist design on training layers of a neural network and run it more efficiently.
Our AI model was run on AWS EC2 with a g2.2xlarge instance type. We set the learning rate (lr) to 0.01.. A smaller learning rate will force the network to learn more slowly but may also lead to optimal classification convergence, especially in cluttered scenes where a large amount of object confusion can occur. In the end, our AI model with came out with 97% of accuracy, less than 0.3 loss over three runs within a minute, and it took less than 20 minutes to run on our 3.2G satellite images.
The model result was exported and visualized using QGIS (http://www.qgis.org/en/site/). QGIS is an open source geographic information system that allows you to create, edit, visualize, analyze and publish geospatial information and maps. The map inspection was also done through comparing our fire damage results to the briefing map produced by Wildfire Today (https://inciweb.nwcg.gov/incident/article/5113/34885/) and Incident Information System (https://inciweb.nwcg.gov/incident/article/5113/34885/).
Fig 4. (A). using OmniEarth parcel level burnt and intact buildings layout on top of the imagery.
Fig 4 (B). The burnt impact (red color) on top of the Great Smoky Mountains from late Nov. to early Dec 2016.
Satellite image classification is a challenging problem that lies at the crossroads of remote sensing, computer vision, and machine learning. A lot of currently available classification approaches are not suitable to handle high-resolution imagery data with inherent high variability in geometry and collection times. However, OmniEarth is a startup company that is passionate about the business of science and scaling quantifiable solutions to meet the world’s growing need for actionable information.
Contact OmniEarth for more information:
For more detailed information, please contact Dr. Zhuangfang Yi, email: email@example.com; twitter: geonanayi.
Dr. Shay Strong, email: firstname.lastname@example.org; twitter: shaybstrong.