Durga Prasad Dhulipudi supervised by Prof. K. S. Rajan received his Doctorate of Philosophy (Ph.D) in Spatial Informatics (Spat). Here’s a summary of his research work on Automated Detection and Mapping of Airport Surface Features for Enhanced Navigation

Over the past few decades, the field of computer vision, inspired by the intricacies of the human visual system, has undergone a transformative evolution. This revolution has increased its impact across many domains, including generic object recognition, face identification, fashion, traffic sign detection, Advanced Driver Assistance Systems (ADAS), medical image analysis, drug discovery, and the interpretation of satellite images. While satellite imagery is pivotal for various sectors, including agriculture, forestry, disaster management, and navigation, its manual translation into maps remains arduous and time-consuming. A promising solution to this problem is the Deep Convolutional Neural Network (CNN), which has shown remarkable results in remote sensing applications. A major hurdle in advancing AI/ML research in this domain has been the need for high-resolution imagery data. However, recent years have witnessed increased accessibility to such data, thanks to initiatives like the United States Geological Survey(USGS) and European Space Agency(ESA), and platforms like Google Earth Engine. Harnessing these available datasets and advancements in deep learning offers a promising avenue for automating satellite image translation. By exploiting such datasets, research has extensively explored the extraction of roads and buildings from remote sensing images due to their significance in traffic management, city planning, GPS navigation, and map updating. Various methods have been proposed for road extraction, including classification-based, knowledge-based, mathematical morphology, active contour model, deep learning, and dynamic programming. Deep learning approaches are particularly effective for road semantic segmentation. Similar state-of-the-art research has been conducted in the field of aviation for the extraction of geospatial information from remote sensing. However, most studies have only extracted relatively coarse features useful during the landing phase. Fine-grained features suitable for aircraft ground movement are scarce, making it challenging to automatically extract the finer airport surface features required to create a digital map for ground navigation. This thesis aims to develop a method for automatically extracting finer airport surface features from satellite images within the 28 days stipulated by the Aeronautical Information Regulation and Control(AIRAC) cycle, as defined in the Radio Technical Commission for Aeronautics(RTCA) DO-272 standard. The major challenge in this project is the efficient, optimized, and timely extraction of airport features, which differs from the extraction of roads or buildings from aerial or satellite images. Object-level detection is sufficient in road or building extraction, whereas foreground and background are important in airport feature extraction. The unique problem of treating a group of entities as a single object also exists in airport surface feature extraction. This is because airport surface features are sparse and have a semantic meaning, which is crucial for aircraft navigation. Therefore, the proposed method must extract repeated patterns of spatial objects, such as runway threshold markings or chevron markings, which form an object of interest together. These patterns must be spatially contiguous and labeled as one entity or spatial category. Additionally, the proposed method must address the use cases and significance of airport surface feature extraction in the aviation domain, where any changes have to be recorded immediately for operational purposes. Our research journey is divided into several segments. We first look at the literature that focused on applications of GIS data to aviation surface operations. We examine the SOTA methods utilized for extracting airport features, evaluating their accuracy, challenges, and appropriateness for creating aviation surface maps. We then address the need for granularity in existing datasets, particularly concerning aviation ground navigation. In response, we introduce the Airport Ground Navigation Object Detection Dataset (AGNODD). This dataset narrows the granularity gap, enabling digital moving maps for precise ground movements, whereas previously available datasets focused predominantly on coarser landing-centric features. In the third part, we fine-tune SOTA ML algorithms on our dataset, focusing on image patches. While extensive research has been done on extracting roads and buildings, aviation-centric geospatial information demands a distinct approach. Current techniques, harnessing deep learning, image processing, and object-based analysis, show promise in airport and runway detection. However, traditional CNN methods primarily target broader airport features. We adapted these models to detect fine-grained airport features crucial for aircraft safety. In the fourth part, we address challenges encountered during the automated detection on entire satellite images. While our dataset yields satisfactory results on smaller patches using advanced algorithms, performance diminishes when scaled to full images, attributed to their intricate contexts. Traditional sliding window techniques are riddled with false negatives, especially when differentiating visually similar features such as runway thresholds and building rooftops. Despite its widespread use, the conventional sliding window approach must improve in geospatial data applications. Its uniform application often misidentifies or entirely overlooks features that don’t align with the window’s size or orientation. Our research suggests a more effective alternative: a multi-level of detail (LOD) tiling scheme. This method adjusts the scale and granularity based on the image’s context, aiming to boost detection accuracy and reduce false positives and negatives. The fifth segment introduces an solution that combines Generative AI (GenAI) and Prompt Engineering . This approach is designed to refine and validate the outputs of conventional segmentation models. Our methodology employs a two-stage analysis process. In the first stage, initial model detections are made. These are then reviewed in the second stage with the assistance of GenAI, significantly reducing false alarms and improving the accuracy of runway marking data for AMDB applications. To overcome full-scene brittleness, we propose **AERO-DETR**, a dual-perspective extraction pipeline. A top-down stage uses **YOLOv8 oriented bounding boxes** to hypothesize **runway corridors**, constraining search to plausible regions. A bottom-up stage then applies **DETR-based set prediction** to **orientation-standardized runway markings**, leveraging global attention and sequence/layout semantics to assemble precise runway polygons. Post-processing enforces shape and adjacency priors derived from **ICAO Annex 14 0.89** against ground-truth polygons, substantially reducing false positives compared with tiling-only baselines.

July 2026