GeoVisual-Detect Project Summary Inspiration Landslides and avalanches pose significant, often immediate, threats to infrastructure, transportation corridors, and human life. The core challenge in preventing these disasters is the timely detection of slope instability. We found that traditional monitoring methods fall short. In-ground sensors are expensive and difficult to scale. Satellite monitoring is periodic, meaning rapid-onset events can be missed entirely. Even existing camera monitoring often relies on manual, subjective human inspection, which isn't scalable and is prone to error. The most critical, pre-failure warning signs—such as new cracks, minor rockfalls, or slow soil creep—are often too subtle for a human to notice until it's too late. This "data gap" inspired us to create an automated, cost-effective, and continuous monitoring system that can see and quantify these subtle visual pre-cursors. What it does GeoVisual-Detect is an intelligent, end-to-end software engine that ingests visual data from cameras and produces actionable geohazard alerts. It works by establishing a high-fidelity visual baseline of a slope and then continuously watching for change. The system's methodology is based on a "dual-analysis" approach: 2D Semantic Change Analysis (High-Frequency): Every 1-5 minutes, the system analyzes new images. It uses a deep learning model (semantic segmentation) to classify every pixel (e.g., stable rock, loose rock, soil, snow). It then compares this to the baseline to detect rapid changes, such as new cracks, boulder movement, or small rockfalls. 3D Shape/Volume Analysis (Low-Frequency): Every 6-24 hours, the system uses photogrammetry to build a new 3D model of the slope. It compares this to the 3D baseline to perform a "cloud-to-cloud" comparison. This method is extremely sensitive and can quantify slow, cumulative deformation like soil creep, or calculate the exact volume of material lost (erosion) or gained (deposition) in an area. Based on this analysis, an AI-based decision module classifies the severity of the change and issues tiered alerts: Level 1 (Watch): Minor, slow change detected (e.g., 2 cm of soil creep). Level 2 (Alert): Moderate, localized change (e.g., new rockfall cluster). Level 3 (Critical): Large-scale, rapid change. An immediate alert is sent to stakeholders. Users interact with the system through a web-based dashboard that shows camera feeds, change "heatmaps," and volumetric reports. It also has an API to integrate with existing emergency alert systems (SMS, sirens, etc.). How we are going to build it The system will be designed with three main components: Input Module (On-Site): We will use a network of fixed, high-resolution (20MP+) weather-proof cameras, including IR-capable cameras for 24/7 operation. These will be installed on stable mounts overlooking a target slope. An on-site edge computing device (e.g., NVIDIA Jetson) will handle initial pre-processing and the high-frequency 2D analysis to minimize data transmission. Processing Module (Cloud/Server): This will be the core "brain" of the system. Data Ingestion: Will receive data from the edge devices. Image Registration & Normalization: A critical step that will precisely align all incoming images to the pixel level and correct for illumination changes (shadows, sun angle). Semantic Segmentation Model: A deep learning model (e.g., U-Net) will be trained on a custom dataset of geological features. 3D Reconstruction Engine: Will use Structure from Motion (SfM) and Multi-View Stereo (MVS) algorithms to build the 3D models. Change Analysis & Decision Core: Will fuse the 2D and 3D data to classify changes and manage alerts. Implementation Plan: The project will be planned in four phases: Phase 1: Data Collection: Identify pilot sites and install hardware to collect a diverse baseline dataset (all weather/lighting). Phase 2: Model Development: Manually annotate thousands of images to train the segmentation model and benchmark the 3D differencing algorithms. Phase 3: System Integration: Build the end-to-end software pipeline, from ingestion to the user dashboard. Phase 4: Field Deployment & Validation: Deploy the prototype, monitor its performance, and validate its alerts against ground-truth (e.g., drone surveys). Challenges we might run into We anticipate several key environmental and technical challenges: Illumination & Shadows: Variable lighting can create "false" changes. Mitigation: Use IR cameras for nighttime consistency and apply illumination-invariant algorithms and shadow-removal filters. Weather (Rain, Snow, Fog):Obscured visibility is a major issue. During rain or fog, the system can struggle to differentiate between actual terrain changes and transient effects like raindrops or water streaks, which can obscure true surface deformation. Mitigation: Use weather-proof hardware and an algorithm to detect and discard "low-visibility" frames. Train models on weather-affected images to improve resilience. False Positives (Animals, Vegetation): Non-geological movement (birds, wind-blown trees) could trigger false alarms. Mitigation: Our semantic segmentation model is trained to automatically classify and ignore these non-relevant changes. We hope to accomplish:
What we hope to accomplish We aim to develop a system that transforms geohazard monitoring from a reactive approach to a truly proactive one. Our goals focus on innovation, accessibility, and precision: Dual-Analysis (2D/3D) Engine: We hope to perfect a hybrid detection framework that combines the strengths of both 2D and 3D analysis—high-frequency 2D monitoring for rapid events, and precise 3D modeling for gradual, cumulative changes that often precede slope failure.
Cost-Effective & Scalable Deployment: By leveraging commercial-off-the-shelf (COTS) camera systems instead of expensive specialized sensors, we aim to make large-scale geohazard monitoring both affordable and easily deployable across diverse terrains.
High-Fidelity Detection: Our objective is to enable the system to not just detect motion, but to quantify it accurately—for example, identifying a “2 cm soil creep” or a “1.5 cubic meter rockfall”—providing meaningful, actionable data rather than simple alerts.
Intelligent False Positive Reduction: We plan to refine our semantic segmentation model to better understand what is changing—distinguishing between geological movement and irrelevant factors like animals or vegetation—to ensure alerts are both intelligent and reliable.
What we learned This project reinforced several key engineering lessons: Data Pre-processing is Everything: We learned that for visual change detection, the image registration and illumination normalization steps are just as important, if not more so, than the deep learning model itself. Garbage in, garbage out. No Single "Silver Bullet": Relying on only 2D analysis would miss slow creep. Relying on only 3D analysis (which is computationally expensive) would miss rapid-onset events. The fusion of both data types is essential. Understand the "Noise": We learned to characterize all the "normal" changes in an environment (lighting, animals, vegetation) and treat them as part of the baseline, which was critical for reducing false alarms. Edge Computing is a Necessity: For real-world, remote deployments, processing data at the source (edge) is not just a "nice to have"—it's a core requirement for managing data costs and power consumption.
Log in or sign up for Devpost to join the conversation.