Advancing space-borne computational vision for rare object mapping
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
This work investigates and proposes solutions to common challenges in remote sensing for mapping rare targets. The first part focuses on the classification and precise localization of rare targets like Maasai Boma homesteads across broad geographic areas. A mapping workflow is developed using a Proxyless-NAS deep convolutional neural network (DCNN) for broad-area scanning and a two-phase weighted mean-shift algorithm for efficient clustering. This workflow achieves remarkably high precision in pinpointing the precise locations of Boma targets and identifies numerous human false-negatives, substantiating DCNN's feasibility for large-scale scanning. The second part explores using multi-temporal medium-resolution Sentinel-2 data instead of costly very-high-resolution imagery. A Deep Seasonal Network with an EfficientNet backbone is designed to effectively leverage multi-temporal data, significantly improving the F1 score for Boma classification compared to the plain EfficientNet architecture, demonstrating the viability of leveraging medium-resolution multi-temporal data. The third part tackles few-shot scene classification, critical when data annotation is limited. A Multi-head Hierarchical ProtoNet architecture is proposed, utilizing the hierarchical relationships among labels for few-shot classification. Multiple output integration methods are evaluated, with a hybrid approach combining weighted summation and joint probability leading to the best performance, significantly boosting few-shot classification across backbone architectures. The research addresses key rare target mapping challenges through novel architectures validated on real-world case studies and benchmark datasets, providing valuable insights for precise mapping, spatial-temporal data fusion, few-shot learning with label ontology, and analogous problems for rare objects.
Table of Contents
DOI
PubMed ID
Degree
Ph. D.