As the cost of GPUs continues to rise, GPU-sharing solutions have become increasingly important for improving efficiency and maximizing resource utilization. At the same time, large-scale operational deployments of such solutions remain relatively less explored, especially in heterogeneous production environments where workload dynamics and orchestration complexity introduce new practical considerations. In this paper, we introduce eGPU, an elastic, efficient, and scalable GPU-sharing framework tailored for production-scale concurrent machine learning (ML) training and inference. eGPU enables fine-grained, runtime-adjustable sharing of GPUs across multiple jobs, while preserving high resource utilization and fault isolation. To address communication bottlenecks, eGPU supports native NVLink/NCCL-based communication between shared GPU instances, capabilities that are limited or unavailable in many existing designs. Built with production deployment in mind, eGPU integrates with Kubernetes (K8s) to support large-scale orchestration. It has been deployed and running stably in production clusters with over$\text{1 0, 0 0 0 ~ G P U s}$for five years. Our evaluation results show that eGPU achieves elastic and precise control over instance sizes, improves job efficiency by 21 % to 31% than SOTA sharing solutions, saves the number of GPUs required by up to$8 \times$, and improves cluster GPU utilization by more than$\mathrm{3} \times$.

Heavy rainfall and earthquake in mountain areas usually trigger numerous landslides. Fast and accurate mapping of landslides is crucial for risk management and emergency rescue. Deep learning-based landslide detection methods can automate identification, but convolutional neural network (CNN) models focus primarily on local features, often missing crucial global context in landslide images. Conversely, Transformer-based models excel at capturing global features but are hindered by high computational complexity. As a result, existing detection models struggle to strike an effective balance between accuracy and efficiency. To address this issue, this article presents a lightweight landslide detection method based on the newly proposed Mamba network. Specifically, a landslide detection model named SegMamba2D with an encoder–decoder structure is proposed. In the encoder, the Mamba network is used to extract multiscale features. A state-space model (SSM) is employed to reduce computational complexity while maintaining accuracy. In the decoder, a multilayer perceptron is used to build a lightweight decoder, ensuring that the model’s overall complexity remains low. The experimental results on both public and new datasets demonstrate that SegMamba2D achieves a superior landslide detection accuracy, with an approximately 2% improvement in$F1$score across various scenarios over conventional models, while significantly reducing computational costs. Additionally, SegMamba2D demonstrates robust generalization performance across diverse research areas. These advancements highlight the model’s potential to enhance accuracy in creating landslide inventories and expedite emergency response times during landslide disasters. The source code is available athttps://github.com/xiaochuan-tang/SegMamba2D

Landslide hazards pose a great threat to the local residents and infrastructure in mountain areas. Numerous technologies have been invented to monitor landslides, and large amounts of high-resolution spatio-temporal data are consistently emerging. These data are highly related to the national security. The local governments release legislation to regulate the sharing of these data. However, the existing landslide detection models explicitly or implicitly assume that landslide monitoring and mapping data are directly shared in a centralized server. There is a gap between landslide detection models and landslide data sharing. To bridge this gap, this letter proposes a privacy-preserving machine learning method named federated learning-based landslide detection (FedLD) for landslide detection. First, horizontal federated learning (HFL) is introduced to protect the data privacy of the modeling process of landslide detection, enabling the development of landslide detection models without direct sharing of the original landslide monitoring data. Second, a new marginal contribution (MC) metric is proposed to measure the contribution of the participants of federated landslide detection models and is used to develop a model aggregation algorithm for federated landslide detection. Experimental results demonstrated that FedLD is able to protect the privacy of popular deep learning-based landslide detection models and achieves competitive landslide classification performance. Therefore, federated learning (FL) provides an effective solution for promoting data sharing in landslide detection.

Landslide recognition is widely used in natural disaster risk management. Traditional landslide recognition is mainly conducted by geologists, which is accurate but inefficient. This article introduces multiple instance learning (MIL) to perform automatic landslide recognition. An end-to-end deep convolutional neural network is proposed, referred to as Multiple Instance Learning–based Landslide classification (MILL). First, MILL uses a large-scale remote sensing image classification dataset to build pre-train networks for landslide feature extraction. Second, MILL extracts instances and assign instance labels without pixel-level annotations. Third, MILL uses a new channel attention–based MIL pooling function to map instance-level labels to bag-level label. We apply MIL to detect landslides in a loess area. Experimental results demonstrate that MILL is effective in identifying landslides in remote sensing images.