Knowledge-Guided Floodwater Level Reasoning for Cloud-Edge Collaborative Monitoring

Zhengshu Zhou; Shaoyu Wang; Ziyi Geng; Lu Tao; Qian Long

cs.CVNov 1, 2026

Knowledge-Guided Floodwater Level Reasoning for Cloud-Edge Collaborative Monitoring

Authors: Zhengshu Zhou, Shaoyu Wang, Ziyi Geng, Lu Tao, Qian Long

Organizations: College of Artificial Intelligence, Tianjin University of Science and Technology, Tianjin 300457, China · School of Resource and Environmental Sciences, Wuhan University, Wuhan 430079, China

Abstract

Frequent urban floodwater poses a serious threat to the safety and efficiency of intelligent transportation systems. Existing vision-based methods for floodwater level detection face two main limitations. First, the lack of explicit reasoning leads to poor detection performance and little interpretability. Second, these methods struggle to generalize to the complex and varied real-world road scenes seen from different viewpoints. To address this, we reformulate the task as an explicit visual reasoning problem and propose HydroReasoner, a floodwater level reasoning framework designed for cloud-edge collaboration. HydroReasoner builds on Qwen2.5-VL and is trained with Group Relative Policy Optimization. To address the scarcity of reasoning process annotations, we build a floodwater domain knowledge graph HydroKG and design a novel KG path reward. Through deterministic graph traversal over HydroKG, this reward supervises the model's reasoning process at a fine granularity, which improves both reasoning and generalization. For edge deployment, we apply 4-bit quantization to the model, cutting GPU memory usage by 57% and inference latency by 43% while out-of-domain accuracy drops by only 0.2 percentage points. We also propose a confidence-based cloud-edge routing mechanism, which improves system safety and eases the computational load on the cloud. In addition, we build HydroLevel, a floodwater benchmark dataset of 3,000 images, for training and evaluating the model. Experimental results show that HydroReasoner achieves detection accuracy 4.0 and 13.6 percentage points higher than conventional vision baselines in the in-domain and out-of-domain evaluations respectively, with performance dropping by only 4.4 percentage points out of domain.

Figures & tables

Fig. 1: Overview of the proposed collaborative edge-cloud traffic monitoring network. The system leverages 5G technology to aggregate multi-source floodwater imagery from Unmanned Aerial Vehicles (UAVs), vehicles (dashcams), and Roadside Units (RSUs). These diverse data streams, captured from heterogeneous viewpoints, are transmitted to the cloud server or processed locally, where the proposed HydroReasoner framework performs efficient and interpretable floodwater level reasoning.

Fig. 2: The overall framework of HydroReasoner. The pipeline comprises three stages: (I) Reasoning Path Generation: the policy model performs explicit reasoning on the input floodwater image and outputs a structured reasoning path. (II) HydroKG Path Verification: the reasoning path is parsed into candidate triples and verified against HydroKG for structural completeness and consistency with domain grading rules. (III) Reward Shaping and GRPO Update: the verification result, together with format and accuracy signals, forms a composite reward that drives the policy update.

Fig. 3: Sample images from the HydroLevel dataset. The dataset comprises floodwater images across multiple views, diverse scenes, and varying severity levels. The viewing perspective is annotated in the bottom-right corner of each image: CCTV (Closed-Circuit Television) denotes a surveillance view, while UAV denotes an Unmanned Aerial Vehicle view. In selected images, red bounding boxes highlight key reference objects used for assessing the floodwater level, accompanied by the classification criteria.

Dataset Subset	Domain	Content	Class Distribution				Total
Dataset Subset	Domain	Content	L1	L2	L3	L4	Total
HydroLevel-Train	Source	$\langle I,y\rangle$	400	400	400	400	1,600
HydroLevel-Val	Source	$\langle I,y\rangle$	50	50	50	50	200
HydroLevel-VQA	Source	$\langle I,\text{CoT},y\rangle$	450	450	450	450	1,800
HydroLevel-InD	Source	$\langle I,y\rangle$	50	50	50	50	200
HydroLevel-OutD	Target	$\langle I,y\rangle$	250	250	250	250	1,000

TABLE I: Detailed specifications of the HydroLevel Benchmark.

Fig. 4: Statistical distributions of visual attributes in the HydroLevel dataset. The subfigures illustrate the dataset’s diversity in (a) road scenes, (b) reasoning referents (e.g., vehicles and pedestrians), (c) capture viewpoints, and (d) illumination conditions.

Model Name	4-class Metrics				2-class Metrics		Param.	Modality
Model Name	Acc	Prec.	Rec.	F1	Acc	F1	Param.	Modality
HydroReasoner (Ours)	0.920	0.9319	0.920	0.9216	0.975	0.975	7B	VL
DenseNet	0.875	0.8763	0.875	0.8751	0.945	0.945	20M	Vision
EfficientNet	0.880	0.8892	0.880	0.8788	0.955	0.956	12M	Vision
ViT-base	0.880	0.8898	0.880	0.8818	0.935	0.936	86M	Vision
Swin Transformer	0.870	0.8757	0.870	0.8709	0.935	0.9341	88M	Vision

TABLE II: Comparison of Model Performance on In-domain Data.

Fig. 5: Comparison of fine-grained confusion matrices on the in-domain and out-of-domain test sets. We select representative models from each category of baselines, including HydroReasoner, SFT variant, Gemini 3, and ViT-Base, to compare their performance differences on the fine-grained four-class task.

Model Name	4-class Metrics				2-class Metrics		Param.	Modality
Model Name	Acc	Prec.	Rec.	F1	Acc	F1	Param.	Modality
HydroReasoner (Ours)	0.876	0.8811	0.876	0.8775	0.955	0.9549	7B	VL
DenseNet	0.735	0.7444	0.735	0.7319	0.858	0.8570	20M	Vision
EfficientNet	0.726	0.7292	0.726	0.7254	0.860	0.8596	12M	Vision
ViT-base	0.740	0.7527	0.740	0.7381	0.823	0.8223	86M	Vision
Swin Transformer	0.704	0.7075	0.704	0.7012	0.805	0.8042	88M	Vision

TABLE III: Comparison of Model Performance on Out-of-domain Data.

Fig. 6: Generalization gap analysis between in-domain and out-of-domain settings. The figure shows the performance differences between HydroReasoner and models trained with supervised learning strategies, including conventional image classification models and SFT variants, on in-domain and out-of-domain evaluations. Hollow and solid markers indicate performance on the in-domain and out-of-domain test sets, respectively, and the length of the connecting line represents the performance gap between the two evaluations.

Fig. 7: Comprehensive performance analysis of HydroReasoner versus baseline models. The subfigures illustrate: (a) the relationship between in-domain/out-of-domain four-class accuracy and safety; (b) the relationship among parameter scale, generalization gap, and accuracy; (c) the trade-offs among coarse-/fine-grained task performance, safety, and model size; and (d) an overall assessment that integrates out-of-domain accuracy, safety, and robustness (note that all metrics in the figure are normalized).

Configuration	Reasoning Quality		Accuracy
Configuration	SC	LC	OutD 4-class
w/o KG path reward	0.782	0.923	0.843
w/o $s_{\mathrm{struct}}$	0.828	0.956	0.851
w/o $s_{\mathrm{pred}}$ , $s_{\mathrm{gt}}$	0.979	0.943	0.863
w/o $s_{\mathrm{branch}}$ , $s_{\mathrm{amb}}$	0.978	0.980	0.870
Full (Ours)	0.983	0.991	0.876

TABLE IV: Component-Level Ablation of the KG Path Reward.

Configuration	Inference Time	Memory	Avg. Power	4-class Accuracy
Configuration	Inference Time	Memory	Avg. Power	InD	OutD
BF16, $K{=}1$	7.6 s	16.7 GB	37 W	0.920	0.876
INT4, $K{=}1$	4.3 s	7.1 GB	26 W	0.920	0.874
INT4, $K{=}3$	6.8 s	7.3 GB	30 W	0.920	0.874

TABLE V: Ablation Study on Quantization Strategy

Model Precision	Format Validity		KG Path Validity		Sampling Consistency		Path-Alignment
Model Precision	InD	OutD	InD	OutD	InD	OutD	InD	OutD
BF16	1.000	0.995	0.995	0.978	1.000	0.980	0.900	0.855
INT4	1.000	0.995	0.990	0.970	1.000	0.980	0.900	0.850

TABLE VI: Effect of Quantization on Structured Output Quality.

Fig. 8: Reasoning processes and final conclusions of HydroReasoner, Qwen2.5-VL-Instruct-72B, Gemini 3 and the SFT variant of HydroReasoner on two typical floodwater scenarios (both Level 4). The content contained within the $<$ think $>$ $<$ $/$ think $>$ tags represents the reasoning process, in which the red-colored text indicates the key judgment basis. The content contained within $<$ answer $>$ $<$ $/$ answer $>$ represents the final conclusion.

Prompt Strategy	In-domain		Out-of-domain
Prompt Strategy	4-class Acc	2-class Acc	4-class Acc	2-class Acc
w/o Reasoning (Direct Output)	0.890	0.975	0.845	0.933
w/ Reasoning (Ours)	0.920	0.975	0.876	0.955

TABLE VII: Ablation Study on Reasoning Strategy.

Metric	In-domain	Out-of-domain
Edge Accuracy (4-class)	0.920	0.876
Cloud-Routing Rate	0.150	0.253
Edge-Accepted Accuracy	0.976	0.969
System-Level High-Risk Recall	1.000	1.000
High-Risk Miss Interception Rate	–	1.000
Error Recall	0.750	0.815

TABLE VIII: Evaluation of the Cloud–Edge Collaborative Inference.

Metric	In-domain	Out-of-domain
Reference-Object Accuracy	0.980	0.930
Waterline-Evidence Accuracy	0.900	0.855
Path-Alignment Accuracy	0.900	0.855
Classification Accuracy (ref.)	0.920	0.885

TABLE IX: Reasoning Grounding Evaluation.

Layer	Node types	Relation	Target layer
Anchor	Physical reference (6)	has_wl_rel.	Evidence
	No reliable reference (1)	has_env_cue	Evidence
	Surface-state start (2)	has_surf_st.	Evidence
Evidence	Waterline relation (18)	indicates	Depth
	Environmental cue (5)	indicates	Depth
	Surface state (3)	supports	Level

TABLE X: The Schema of HydroKG.

Head	Relation	Tail	Head	Relation	Tail
Level 1 (short chain): road-surface state			Reference object: e-bike
road_surface	has_surf_st.	dry_road_surface	ebike	has_wl_rel.	below_footboard
road_surface	has_surf_st.	slight_wetness	ebike	has_wl_rel.	footboard~half_wheel
natural_water_region	has_surf_st.	natural_water_body	ebike	has_wl_rel.	above_half_wheel
dry_road_surface	supports	level_1	below_footboard	indicates	depth_lt_15cm
slight_wetness	supports	level_1	footboard~half_wheel	indicates	depth_15_30cm

TABLE XI: The complete edge set of HydroKG

Zhengshu Zhou received the B.E. degree in software engineering from Tianjin University, Tianjin China, in 2013. He received the Master and Ph.D. degree in informatics from Nagoya University, Nagoya Japan, in 2020 and 2022, respectively. He is currently an Associate Professor with the College of Artificial Intelligence, Tianjin University of Science & Technology (TUST). Previously, he was a system engineer with the Japanese automotive industry. And he was a researcher at Hitachi Yokohama Research Laboratory, Yokohama, Japan. He has published over 30 academic papers, including in top journals in the fields of Intelligent transportation and reliability engineering, such as IEEE TITS, IEEE TVT, IEEE TNSM and RESS. He has also led several research projects related to intelligent driving and computer vision. His research interests mainly include software engineering, autonomous driving, and high-precision dynamic map.

Table 20

Shaoyu Wang received the B.E. degree in computer science and technology from North China Institute of Aerospace Engineering, Langfang, China, in 2024. He is currently pursuing the M.S. degree with Tianjin University of Science and Technology, Tianjin, China. His current research interests include deep learning and autonomous driving.

Table 21

Ziyi Geng received the B.E. degree in software engineering from Hebei University of Economics and Business, Shijiazhuang, China, in 2024. He is currently pursuing the M.S. degree with Tianjin University of Science and Technology, Tianjin, China. His current research interests include autonomous driving and software engineering.

Table 22

Lu Tao is currently a tenure-track associate professor at Wuhan University. He received his B.S. and M.E. degrees in Geographic Information Systems and Navigation, Guidance, and Control from Wuhan University in 2014 and 2017, respectively. He obtained his Ph.D. degree in Informatics from Nagoya University in 2023, and was then a researcher at the Institute of Innovation for Future Society from 2023 to 2024. His research interests include vehicle motion prediction, sensor fusion, high-definition maps, dynamic maps, edge and cloud computing, and connected autonomous vehicles.

Table 23

Qian Long received the B.S. and Ph.D. degrees from the University of Science and Technology of China, Hefei, China, in 2000 and 2007 respectively. He was a faculty member at the Department of Precision Machinery and Precision Instrumentation of the University of Science and Technology of China until 2015. He has worked as a visiting scientist at Brookhaven National Laboratory, Upton, USA (2009-2010), as a research associate at the University of Texas at Brownsville, Brownsville, USA (2010-2011), as a postdoctoral fellow at the Research Center for Smart Vehicles, Toyota Technological Institute, Nagoya, Japan (2012-2015), as a researcher at Nippon Soken, Inc., Nishio, Japan (2015-2016), and as a professor at Yunnan Observatories, Chinese Academy of Sciences, Kunming, China (2016-2023). Currently, he is a professor at the College of Artificial Intelligence, Tianjin University of Science & Technology, Tianjin, China. His research interests include computer vision, machine learning, autonomous driving, and artificial intelligence.

Table 24

Reasoning for VLMs

See all 78 papers

Aug 15, 2026

FloodReasonBench: Benchmarking VLM Reasoning Segmentation for Embodied Flood Response at the Edge

Reasoning segmentation enables vision-language models (VLMs) to translate mission-relevant language requests into pixel-level visual grounding, offering a natural perception interface for embodied agents. However, existing benchmarks largely focus on generic visual scenes and overlook the domain and resource constraints encountered in flood-response platforms. We present FloodReasonBench, a benchmark for VLM reasoning segmentation for embodied flood response at the edge. At its core, FloodReasonBench introduces FloodResponseSeg, a flood-specific reasoning-segmentation dataset constructed from real-world scenes and response-relevant targets. Beyond task accuracy, the benchmark characterizes reasoning-segmentation pipelines under lightweight visual encoding, hierarchical split inference, and compressed intermediate representations. We observe strong partition-dependent accuracy variation in the generic pre-adaptation setting, while the flood-adapted target-workload design space exhibits a substantially more compact accuracy range across partitions. Evaluation on an NVIDIA Jetson AGX Xavier further exposes the tradeoffs among reasoning-segmentation accuracy, edge-side latency, energy, and communication footprint, enabling quality-constrained selection of edge operating points. Together, these results provide a task- and system-level characterization of reasoning segmentation for resource-constrained embodied flood response at the edge.

Rajat Bhattacharjya, Yoomee Jung, Minwoo Kim +4

University of California, Irvine · Kookmin University, South Korea

Reasoning Segmentation VLM Evaluation Efficient VLM Inference Disaster Response

Ask PDF

Mar 17, 2026

LLM-Powered Flood Depth Estimation from Social Media Imagery: A Vision-Language Model Framework with Mechanistic Interpretability for Transportation Resilience

Urban flooding poses an escalating threat to transportation network continuity, yet no operational system currently provides real-time, street-level flood depth information at the centimeter resolution required for dynamic routing, electric vehicle (EV) safety, and autonomous vehicle (AV) operations. This study presents FloodLlama, a fine-tuned open-source vision-language model (VLM) for continuous flood depth estimation from single street-level images, supported by a multimodal sensing pipeline using TikTok data. A synthetic dataset of approximately 190000 images was generated, covering seven vehicle types, four weather conditions, and 41 depth levels (0-40 cm at 1 cm resolution). Progressive curriculum training enabled coarse-to-fine learning, while LLaMA 3.2-11B Vision was fine-tuned using QLoRA. Evaluation across 34797 trials reveals a depth-dependent prompt effect: simple prompts perform better for shallow flooding, whereas chain-of-thought (CoT) reasoning improves performance at greater depths. FloodLlama achieves a mean absolute error (MAE) below 0.97 cm and Acc@5cm above 93.7% for deep flooding, exceeding 96.8% for shallow depths. A five-phase mechanistic interpretability framework identifies layer L23 as the critical depth-encoding transition and enables selective fine-tuning that reduces trainable parameters by 76-80% while maintaining accuracy. The Tier 3 configuration achieves 98.62% accuracy on real-world data and shows strong robustness under visual occlusion. A TikTok-based data pipeline, validated on 676 annotated flood frames from Detroit, demonstrates the feasibility of real-time, crowd-sourced flood sensing. The proposed framework provides a scalable, infrastructure-free solution with direct implications for EV safety, AV deployment, and resilient transportation management.

Nafis Fuad, Xiaodong Qian

Civil and Environmental Engineering Wayne State University 5050 Anthony Wayne Drive, Detroit, MI, USA, 48202 · Civil and Environmental Engineering Institute for AI and Data Science (AIDaS) Wayne State University 5050 Anthony Wayne Drive, Detroit, MI, USA, 48202

Depth Estimation Intelligent Transportation Systems VLM Interpretability Flood Inundation Mapping

Ask PDF

Aug 2, 2026

Mechanistic Interpretability-Guided Selective Fine-Tuning of Vision-Language Models for Centimeter-Level Flood Depth Estimation

Urban flooding poses an escalating threat to transportation infrastructure, yet no operational system provides real-time, street-level flood-depth estimates at centimeter resolution. This paper presents three vision-language models fine-tuned for continuous flood-depth estimation from street-level imagery: FloodLlama-Dense, a fully fine-tuned QLoRA baseline, and FloodLlama-MI5 and FloodLlama-MI6, interpretability-guided sparse variants that fine-tune only the top five and six causally relevant cross-attention layers identified through mechanistic interpretability analysis, respectively. Training uses an approximately 610,000-image subset of a 2.81-million-image synthetic corpus generated in Unreal Engine 5. The dataset combines single-vehicle subsets with 5 cm depth increments and mixed-vehicle subsets with 1 cm depth increments, spanning seven vehicle types, four weather conditions, and flood depths from 0 to 40 cm. FloodLlama-Dense achieves an MAE of 0.40 cm, an RMSE of 1.97 cm, and an Acc@5cm of 97.59%. Mechanistic interpretability analysis combining linear probing, logit lens, centered kernel alignment (CKA), and cross-attention entropy reveals a two-stage adaptation pattern: layers L13-L22 restructure visual representations, while depth first becomes linearly decodable at layer L23. FloodLlama-MI5 and FloodLlama-MI6 leverage this insight by fine-tuning only five or six of the eight cross-attention layers, achieving an 86-88% reduction in trainable parameters (6.55-7.86 million versus 54.4 million) with minimal accuracy loss. On a real-world benchmark, FloodLlama-MI6 achieves 98.62% accuracy, compared with 86.61% for the published STURM-FloodDepth baseline.

Nafis Fuad, Xiaodong Qian, Dongxiao Zhu

Department of Civil & Environmental Engineering Wayne State University Detroit, MI, USA · Department of Civil & Environmental Engineering Institute for AI and Data Science (AIDaS) Wayne State University Detroit, MI, USA · Department of Computer Science, Institute for AI and Data Science (AIDaS) Wayne State University Detroit, MI, USA

VLM Adaptation Depth Estimation VLM Interpretability Mechanistic Interpretability Parameter-Efficient Fine-Tuning

Ask PDF

Abstract

Reasoning for VLMs

Explore similar work

FloodReasonBench: Benchmarking VLM Reasoning Segmentation for Embodied Flood Response at the Edge

LLM-Powered Flood Depth Estimation from Social Media Imagery: A Vision-Language Model Framework with Mechanistic Interpretability for Transportation Resilience

Mechanistic Interpretability-Guided Selective Fine-Tuning of Vision-Language Models for Centimeter-Level Flood Depth Estimation