Artificial Intelligence vs. landslide

In this article you’ll read,

• how AI can help battling environmental disasters,
• how an AI architect developed an AI in only three months for the World Summit AI
• and which resources from Open Telekom Cloud have been used.

Within just three months, Ferenc Kukucska, AI architect at the Open Telekom Cloud, has developed a landslide detector to identify environmental disasters based on image classifications. The service was launched at the World Summit AI and is available online free of charge.

Global warming and climate change are becoming increasingly noticeable: 2023 was the hottest year since global temperature records began – as confirmed by data from NASA, the World Meterological Organization, the US National Oceanic and Atmospheric Administration, the Climate Research Unit at the University of East Anglia and the British Meteorological Service. The unpleasant side of climate change has been making negative headlines in recent years: Extreme weather events such as prolonged drought and heavy rainfall lead to flooding and landslides that endanger human lives.  

In 2024 alone, there were reports of landslides in Papua New Guinea, India (Kerala), Indonesia and Nepal. But such extreme events are also on the rise in Europe: Switzerland was hit in June, and at the end of October a devastating flood inundated Southern Spain . “For our demo at the World Summit AI 2024, we decided to show how the cloud and artificial intelligence (AI) can help in the fight against such disasters,” explains Ferenc Kukucska, AI architect in the Open Telekom Cloud team. The team wanted to develop a “landslide detector” that analyzes satellite data using artificial intelligence (image classification). The service can be used by public institutions or disaster control to quickly and automatically detect landslides and coordinate rescue operations in a targeted manner.

Popular LLMs could be used to identify landslides, when trained on labeled landslide/non-landslide images and optimal training parameters. But if they are conversational bots, they are only good for chitchat and small talk, not for object/image recognition. “We decided to go for a separate, specific AI model for image classification. We were able to draw on the resources and services of Open Telekom Clous to develop the AI model,” says Kukucska. ModelArts is an AI PaaS that offers development environments, neural networks, frameworks and infrastructure resources from a single source. With ExeML, it is even possible to develop AI services without explicit AI expertise (citizen developer approach). 

But an AI PaaS alone does not make AI. Above all, suitable data in sufficient quantity and quality is important. “Fortunately, satellite data from earth observation is often freely available these days. We decided to use the Bijie dataset for training and testing,” says Kukucska. The Bijie dataset is named after a Chinese city with almost 7 million inhabitants that is regularly exposed to landslides. The dataset includes 2,000 images without landslides and 770 images with landslides. “This set provides a good basis for training. However, this 3:1 ratio can lead to an imbalanced dataset which then leads to inaccurate results in the case of binary questions (landslide: yes or no?),” explains the AI architect. The modeled AI tends to find results that correspond to the majority of the training images more often. In our case, the AI would have a preference for “no landslide".

So, what to do? “We had to shift the balance towards landslide images so that there were roughly similar data sets for both variants,” explains Kukucska. This way, an “unbiased” AI can be created. To do this, the team went “photoshopping”: the 770 landslide images were modified by mirroring, zooming and rotating and added to the data set. All the data was stored in an object storage bucket in the Open Telekom Cloud. This bucket was also used for data engineering, e.g., preprocessing the images for AI training. 

“We first tried the ExeML-based citizen developer approach,” says Kukucska. For the training, he used about three quarters of the available data, but did not use any data augmentation techniques or GPU. With this the first training took about four to five hours. The remaining quarter of the data was used to test the model. “With ExeML, we achieved an analysis accuracy of around 60 percent,” he summarizes the first results. “However, we were aiming for a significantly higher accuracy for our demo. We therefore had no choice but to go deeper and use the advanced AI developer tools from ModelArts. These gave us the opportunity to adjust the hyperparameters for the model and modify the layers of the neural network”.  

Kukucska also tried ExeML at a later stage of the application development, when a lot of additional images were used, selected compute intensive GPU instance as the instance flavor and the Inference Hardware NV_P4. The training then took a few hours.

Convolutional neural networks (CNN) with so-called convolutional layers are often used for image analysis. CNNs can recognize images and patterns. In the convolutional layer, filters (also known as kernels) are moved (folded) over the input image or the input data to extract features or properties from the data. Each filter is specialized for a certain feature such as edges, textures or certain patterns. Running the image through the filter creates a new representation of the image containing only the recognized features. These feature maps are then processed through further layers to identify more complex patterns. In modern models, the CNNs are enriched with other layer types to achieve the best results for image classification, such as Pooling Layers, Fully Connected Layers, ReLU Layers, Batch Normalization Layers, Dropout Layers or Transposed Convolutional Layers.

The model went through a total of 15 - 20 development cycles in just under three months until Kukucska was satisfied with the result. The team completed the work in August with a recognition rate of just under 90 percent – after starting at the end of May. This proves that the on-board resources of the Open Telekom Cloud can generate powerful, specific AI with manageable effort and costs. To reduce development times, the team used GPU resources (8 vCPUs, 64 GB, p2v.2xlarge.8.) for the final training cycles. This resulted in an acceleration by a factor of 3. The new model was available in three to four hours. 

In addition to ModelArts, the Elastic Cloud Servers and Object Storage, container services were used to enable automated deployment on various platforms: Software Repository for Container and the Kubernetes-compatible Cloud Container Engine. The final service was presented at the World Summit AI in October 2024 and is available as a web service. Interested parties can upload satellite images there. The analysis takes place in fractions of a second.