Human-wildlife conflicts in residential/agricultural settings rely on ineffective deterrents like rodenticides or fences. We introduce InvisibleFence, a modular 3D-printed Vision Pod system with off-the-shelf deterrents. The Vision Pod fuses a 2K camera and 240° motion sensing with an edge-optimized pipeline trained on 44,000 wildlife images of eleven classes—achieving 86.7% mAP. Benchmarking YOLO variants (416p–2K) ensures performance. Upon detection, it sends MQTT commands to drive deterrent units—ultrasonic speakers or lighting/spray modules—that emit tones without affecting humans or pets. InvisibleFence creates adaptive zones that reduce false triggers and limit habituation.

Flood is one of the common natural disasters that can severely affect human life and properties. Early detection, therefore, is of paramount importance to provide help through an emergency response team. Robust flood detection techniques so far have been based on computer vision using images either from cameras, satellite imagery, remote sensing, or radar-based images. However, sound signal-based flood event detection has not been widely explored. In this work, we design an end-to-end architecture for a deep learning-based flood-related sound event detection model. We employ Mel-Spectrogram-based auditory signal analysis and deep learning models for sound event detection (SED). We evaluated four deep learning models under the following two categories: (i) Binary classification Flood/No Flood, vs. Windy vs. Non-Windy, and (ii) Multi-classification for more granular flood and wind events. The experimental results performed in these settings on the datasets collected from real deployment showed an accuracy of around 78%.

Monitoring flash floods and providing just-in-time notification to city officials for taking appropriate action and prompt intervention is crucial for any smart city located in flood-prone areas around the world. Flood monitoring systems that exploit image analysis via Machine Learning (ML) techniques have been already proposed in literature. Such systems, however, adopt a cloud-based approach that generates significant data traffic and could be susceptible to failures due to network outages. In such a framework, images are continuously offloaded from cameras deployed in flood-prone areas of the city towards a cloud infrastructure where a service is deployed to analyze the images and detect the rise of water in rivers or city canals in a timely way. In this paper, we present the activities of the project EdgeFlooding, which aims at investigating the opportunity of adopting a distributed approach based on edge computing for the implementation of more resilient and reliable flash flood monitoring systems, that helps mitigate the limitations of the cloud-based systems. We have developed a prototype of an edge computing flood monitoring system based on micro-services, and we run an extensive set of experiments exploiting one European Fed4Fire+ testbed, i.e., the Grid'5000 testbed. The aim of those experiments is to assess whether a distributed edge/cloud computing approach is feasible for the implementation of future flood or environmental monitoring systems.

In this paper, we propose an end to end goal-oriented conversational AI agent that can provide contextual information from a potential hazard site. We posit the conversational agent as a FloodBot capable of seeing, sensing, assessing hazard condition, and ultimately conversing about them. We present our domain-specific FloodBot design-solution and learning-experience from the real-time deployment in a flash flood devastated city that uses state-of-the-art deep learning models. We specifically used computer vision and pertinent natural language processing technologies to empower the conversation power of the FloodBot. To deliver such practical and usable AI, we chain multiple deep learning frameworks and create a human-friendly question-answer based dialogue system. We present our deployment details from the last five months and validate the results using ongoing COVID19's impact on the area as well.

We investigate the practical challenge of localized flood detection in real smart city environment using the fusion of physical sensor and social sensing models to depict a reliable and accurate flood monitoring and detection framework. Our proposed framework efficiently utilize the physical and social sensing models to provide the flood-related updates to the city officials. We deployed our flood monitoring system in Ellicott City, Maryland, USA and connect it to the social sensing module to perform the flood-related sensor and social data integration and analysis. Our ground-based sensor network model record and performs the predictive data analytic by forecasting the rise in water level (RMSE=0.2) that demonstrates the severity of upcoming flash floods whereas, our social sensing model helps collect and track the flood-related feeds from Twitter. We employ a pre-trained model and inductive transfer learning based approach to classify the flood-related tweets with 90% accuracy in the use of unseen target flood events. Finally our flood detection framework categorizes the flood relevant localized contextual details into more meaningful classes in order to help the emergency services and local authorities for effective decision making.

Successful implementation of the Internet of Thing (IoT) is precursory to a thriving smart city. However, the technical, physical, and environmental conditions can often pose challenges in their successful deployments. The deployment is further complicated if the time and location of implementation are amidst a natural disaster. In this work, we use flash flood detection as a natural hazard testbed and describe various IoT deployment, our progression, and first-hand experience from those implementations. We compare and contrast three IoTs and their performance in real-time execution. Next, we discuss systems architecture and their end-to-end design and present lessons learned from these heterogeneous deployments. Additionally, we evaluate and outline our observations, challenges, and opportunities for further improvement. We also formulate standard evaluation metrics for their scoring and document our deployment journey.

Talking to the electronic device and getting the required information at a minimal time has become today's norm. Although AI-powered conversational agents have percolated the commercial market, their use in a communal setting is still evolving. We postulate that the deployments of chatbots in disaster-prone areas can be beneficial to watch, monitor, and warn people during the crisis. Furthermore, the successful implementation of such technology can be life-saving. In this work, we discuss our deployment of a real-time flood monitoring chatbot called FloodBot. We collect, annotate and visually parse images from potentially hazardous areas. We detect the flood conditions and identify objects in harm's way by stacking deep learning models such as a convolutional neural network (CNN), single-shot multi-box object detection (SSD). We then feed the image contents to a knowledge base of our artificially intelligent FloodBot and explore its AI-Conversing power using end to end memory network. We also showcase the power of cross-domain transfer learning and model fusion techniques. In this work, we discuss our deployment of a real-time flood monitoring chatbot called FloodBot. We collect, annotate and visually parse images from potentially hazardous areas. We detect the flood conditions and identify objects in harm's way by stacking deep learning models such as a convolutional neural network (CNN), single-shot multi-box object detection (SSD). We then feed the image contents to a knowledge base of our artificially intelligent FloodBot and explore its AI-Conversing power using end to end memory network. We also showcase the power of cross-domain transfer learning and model fusion techniques.

Overground Wireless Sensor Network (WSN) is reaching its final stage of being able to pass the test of resiliency and reliability. However, the network protocol for underground communication and sensor applications are still evolving. Thus, in this work, we propose to study the sustainability of a wireless underground sensor network (WUSNs) built by fine-grained spatiotemporal sensing grids. The metrics for our system's sustainability measure are reliability, resiliency and the longevity of the device. In this work, we propose to build a low-cost automated WUSN system that can collect and map out the subsurface groundwater movement. Our system senses spatiotemporal soil moisture levels and generates a predictive model to quantify the groundwater movement in soil strata. We propose an end to end system that can monitor soil moisture level during and after the rainfall. The prototype consists of a rain gauge, soil moisture sensors mesh and data storage/transfer unit. The results from this work will document the feasibility and challenges associated with the deployment of the underground sensor network.

Detecting flash floods in real-time and taking rapid actions are of utmost importance to save human lives, loss of infrastructures, and personal properties in a smart city. In this paper, we develop a low-cost low-power cyber-physical System prototype using a Raspberry Pi camera to detect the rising water level. We deployed the system in the real word and collected data in different environmental conditions (early morning in the presence of fog, sunny afternoon, late afternoon with sunsetting). We employ image processing and text recognition techniques to detect the rising water level and articulate several challenges in deploying such a system in the real environment. We envision this prototype design will pave the way for mass deployment of the flash flood detection system with minimal human intervention.

Computer Vision and Image Processing are emerging research paradigms. The increasing popularity of social media, micro- blogging services and ubiquitous availability of high-resolution smartphone cameras with pervasive connectivity are propelling our digital footprints and cyber activities. Such online human footprints related with an event-of-interest, if mined appropriately, can provide meaningful information to analyze the current course and pre- and post- impact leading to the organizational planning of various real-time smart city applications. In this paper, we investigate the narrative (texts) and visual (images) components of Twitter feeds to improve the results of queries by exploiting the deep contexts of each data modality. We employ Latent Semantic Analysis (LSA)-based techniques to analyze the texts and Discrete Cosine Transformation (DCT) to analyze the images which help establish the cross-correlations between the textual and image dimensions of a query. While each of the data dimensions helps improve the results of a specific query on its own, the contributions from the dual modalities can potentially provide insights that are greater than what can be obtained from the individual modalities. We validate our proposed approach using real Twitter feeds from a recent devastating flash flood in Ellicott City near the University of Maryland campus. Our results show that the images and texts can be classified with 67\% and 94\% accuracies respectively