Res-GeoAI: Satellite-to-Drone Dual-Modality System

Abstract

Geospatial AI UAV Semantic Segmentation Disaster Response Thermal Imaging Flood Detection Victim Localization

Rapid and effective disaster response remains a critical challenge, particularly during floods, where timely identification of affected areas and victims can save lives. Existing systems often rely solely on either satellite imagery or ground-based sensors, limiting their ability to quickly and accurately pinpoint flooded zones and locate stranded individuals.

To address this gap, we propose Res-GeoAI, a geospatial AI system that integrates satellite-based flood detection with drone-assisted victim identification. Our system operates in three stages: (1) flood-zone mapping using satellite images; (2) thermal imaging for detecting victims under low-visibility conditions; and (3) GPS-based coordination for rescue operations.

By combining these capabilities, Res-GeoAI enhances the speed and precision of humanitarian interventions, providing a scalable solution for disaster management.

Method

Res-GeoAI operates through a hierarchical multimodal pipeline that bridges macro-level satellite analysis with micro-level UAV victim localization.

Satellite Flood Mapping

Geographic regions are tessellated into tiles. A modified SegFormer with multiscale attention performs semantic segmentation, computing per-pixel flood probability and overall coverage.

Flood Validation

Candidate flood regions are validated via a lightweight Xception-based classifier. UAV deployment is triggered when confidence > 0.7 and flood coverage > 10%.

UAV Victim Detection

UAVs follow a spiral search pattern and apply night-vision transformation. VarifocalNet with IoU-aware loss detects victims; GPS coordinates are computed from bounding boxes.

System Workflow — **Fig. 1.** End-to-end system workflow for flood victim detection and localization. Level 1 (top): satellite segmentation and flood coverage. Level 2 (bottom): UAV victim detection and GPS localization.

Results

66.30%

mIoU on LoveDA (Segmentation)

76.47%

Water IoU (flood detection)

0.865

VFNet mF1 (night vision)

±3.2m

GPS localization error

Semantic Segmentation — LoveDA

Model	Background	Building	Road	Water	Barren	Forest	Agriculture	mIoU
UNet	43.46	60.59	52.35	69.22	46.60	24.23	41.88	48.95
Semantic FPN	60.70	50.92	51.36	73.93	52.05	56.24	76.94	57.98
FuseNet	47.19	50.77	61.36	68.21	31.71	47.71	57.45	52.62
HRNet	54.61	55.34	57.42	73.36	46.78	45.87	69.88	59.40
Ours (CNN+SegFormer)	60.68	63.25	62.08	76.47	47.93	47.82	72.75	66.30

Confusion matrices — **Fig. 2.** Normalized confusion matrices for the top-3 semantic segmentation models.

Flood Validation — FloodNet

Model	Training Accuracy	Test Accuracy
InceptionNetV3	0.990	0.844
ResNet50	0.974	0.937
Xception (Ours)	0.998	0.946

Victim Detection — HIT-MOB (Night Vision vs RGB)

Model	Normal RGB		Thermal		Infrared		Night Vision
Model	mF1	MPR	mF1	MPR	mF1	MPR	mF1	MPR
YOLOv8	0.841	0.136	0.846	0.136	0.850	0.133	0.857	0.131
VFNet (Ours)	0.858	0.117	0.863	0.114	0.860	0.115	0.865	0.111

System Performance (1000 km², Zoom 11)

Module	Input Size	Time	Memory	Accuracy
Tile Generation	1024×1024	0.80s	17.5 MB	1.00
Flood Detection	1024×1024	0.05s	4.5 MB	0.66
Flood Validation	1024×1024	0.02s	1.5 MB	0.95
Victim Detection	512×512	0.04s	2.0 MB	0.86
GPS Localization	—	0.01s	0.1 MB	±3.2m

Dataset

We introduce HIT-MOB, a harmonized dual-modality UAV dataset constructed by merging the HIT-UAV thermal dataset and the MOBDrone maritime dataset into a unified 9-class schema.

129,068 frames · 25,524 annotations · 9 unified classes
Split: 70% train / 10% val / 20% test
Modalities: Thermal (HIT-UAV) + RGB (MOBDrone)
Images resized to 512×512, normalized with ImageNet statistics

HIT-MOB Dataset on Kaggle

kaggle.com/dsv/12184019 · DOI: 10.34740/KAGGLE/DSV/12184019

BibTeX

If you find this work useful, please cite:

@article{pal2025resgeoai,
  title     = {Res-GeoAI: A Satellite-to-Drone Dual-Modality System
               for Flood Detection and Victim Localization
               in Disaster Response},
  author    = {Pal, Anidipta},
  year      = {2025},
  url       = {https://doi.org/10.13140/RG.2.2.32814.88642}
}