Perception & AI |

UAV Multispectral Time-Series and Weather-Based Hurdle Modeling for Onion Diseases and Pest Estimation

Mon, 03 Apr 2023 00:00:00 +0000

Figure: UAV multispectral and weather-based framework for onion disease and pest estimation

Onion crops are highly vulnerable to a range of diseases and pests, which can lead to significant yield losses if timely detection and intervention are not achieved. This study presents a multimodal framework that integrates spectral indices derived from UAV-based multispectral imagery with meteorological data to estimate the incidence of major onion stressors under controlled field conditions. Data were collected from experimental plots of the Bhima onion variety across multiple growth stages, where stemphylium blight, anthracnose, purple blotch, and thrips co-occurred at naturally varying intensities. Fourteen vegetation indices (VI) and four weather variables were evaluated under untreated (no-spray) conditions to determine which features most reliably predict crop health and stress intensity. Beyond conventional indices, we incorporated derived features capturing temporal changes, lag effects, and biomass normalization, which significantly enhanced sensitivity to stress dynamics. Among all evaluated features, a newly introduced vegetation index, referred to as the Green Stress Ratio, consistently demonstrated the strongest associations across all disease and pest categories. A two-stage classification–regression framework was employed to first detect the presence of a stressor and subsequently quantify its severity, with models evaluated independently for each disease and pest. The results highlight the effectiveness of combining UAV-derived multispectral data with weather information for non-invasive, field-scale monitoring of onion crop health. This integrated approach provides a foundation for advanced decision-support systems aimed at targeted and sustainable pest and disease management in onion cultivation.

UAV Multispectral Imaging for Oil and Water Leak Detection

Mon, 03 Apr 2023 00:00:00 +0000

Pipeline leakages involving oil and water pose significant environmental and economic risks, motivating the need for timely and reliable surface leak detection techniques. This paper presents a proof-of-concept study that investigates the use of UAV-based multispectral imagery and deep learning for detecting and distinguishing surface oil and water leaks over soil backgrounds. Multispectral data were acquired using a UAV-mounted sensor under controlled experimental conditions and processed through a radiometric and geometric correction pipeline. Spectral analysis was conducted to assess band-wise separability, highlighting the effectiveness of Red and Red-Edge bands for oil–water discrimination. Multiple object detection models, including YOLOv11, RT-DETR, Faster R-CNN, and a fusion-based YOLOv11-RGBT variant, were evaluated using these inputs. Experimental results demonstrate that single-band models achieve robust detection performance, with YOLOv11 attaining detection accuracies of 92.8% for oil and 98.1% for water, while RT-DETR and Faster R-CNN demonstrated comparable accuracy. The fusion-based YOLOv11-RGBT model achieved lower performance, likely due to limited dataset size for effective cross-band feature learning. Overall, the findings demonstrate that UAV-based multispectral sensing combined with deep learning offers a practical solution for localized surface leakage monitoring.

Object Detection and Tracking

Mon, 03 Apr 2023 00:00:00 +0000

At AitoeLabs, I worked with a team of engineers to develop real-time video analytics solutions for security applications. For ATM surveillance, we developed deep learning models and computer vision algorithms operating on live video streams.

I worked on the object detection and tracking pipeline, developing algorithms to detect and track people in real-time video. The tracking system assigned unique identities to detected individuals, enabling their movement and behavior to be analyzed over time.

Figure: Real-time detection and tracking of people in a live video stream

The system was used to identify potentially suspicious situations, including excessive occupancy and prolonged loitering inside ATM premises. Detected events could trigger alerts for security personnel or the responsible organization.

Jewellery Segmentation

Mon, 03 Apr 2023 00:00:00 +0000

In digital image processing and computer vision, image segmentation is the process of partitioning a digital image into multiple segments. It simplifies image analysis by assigning labels to different regions, which can support subsequent object detection and image-processing tasks.

I worked as a Project Assistant at the Multimodal Perception Laboratory, International Institute of Information Technology, Bangalore, where I analyzed different image segmentation algorithms and developed an algorithm for segmenting the background of shared jewellery images.

The objective was to replace the original background with a transparent background, allowing vendors to further edit the images and prepare them for use on online selling platforms.

Figures 1 and 2 show the performance of the segmentation algorithm on original images provided by the vendors.

Figure 1: Segmentation algorithm — best case

Figure 2: Segmentation algorithm — worst case