Source-linked AI summary
ISW-Agent: Integrating LLM-Driven Dialogue with n8n Workflow for Internal Solitary Wave Remote Sensing Image Processing
Hao Zhang, Hengrui Liu, Lina Sun, Junmin Meng
TL;DR
ISW remote-sensing workflows must transfer task parameters and spatial products across acquisition, detection, extraction, and prediction, while maintaining usable intermediate outputs. ISW-Agent combines LLM task interpretation, n8n orchestration, and scientific models through explicit constraints and validation. In the reported tests, it achieved 94.2% strict task matching for GPT-5.6 Sol and DeepSeek, and completed the full chain in 39 of 50 timed scenes.
Problem
ISW processing requires transferring task parameters and spatial information across stages, while fragmented tools and outputs complicate those handoffs and validation.
Source: §1 · pp. 1–2Method
ISW-Agent connects LLM-based task interpretation, n8n orchestration, and independent scientific models through structured task constraints, geospatial transformations, and stage-specific output checks.
Source: §2.1 · p. 3, §2.2 · p. 3, §4 · pp. 10–11Results
GPT-5.6 Sol and DeepSeek each achieved a 94.2% strict task-match rate, and 39 of 50 timed scenes completed detection, extraction, and prediction with a 48.04 s median turnaround time.
Source: §4 · p. 11Takeaways & Limitations
ISW-Agent provides a traceable workflow foundation for combining task continuity, valid spatial structures, and consistent provenance in multimodel ISW analysis.
Source: §4 · p. 11Detection and stripe extraction reuse an existing YOLO11 model, whose accuracy and regional transferability require evaluation with complete training records and independent regional annotations.
Source: §3.6 · p. 10
Abstract
from arXiv · showhide
Internal solitary wave (ISW) remote sensing requires task parameters and spatial information to be transferred across image acquisition, detection, stripe extraction, and propagation prediction. Here, we present ISW-Agent, which integrates large language model (LLM) interpretation with n8n workflows and independent scientific models through structured task constraints, geospatial transformations, and stage-specific output validation. In tests of 120 Chinese instructions per backend, GPT-5.6 Sol and DeepSeek both achieved a strict task-match rate of 94.2%, with all intents and execution decisions matching the reference. Sentinel-1 cases demonstrated continuous processing from dialogue requests to wavefront prediction for the specified scene. Of 50 timed scenes, 39 completed detection, extraction, and prediction with a median turnaround time of 48.04 s; 11 stopped after empty extraction. Across 3,408 controlled fault-injection runs, retries recovered 234 transient-failure conditions and output validation rejected 234 structural defects, while 61 provenance or completeness errors indicated the need for content-consistency checks. Overall, ISW-Agent thus connects natural-language requests to traceable spatial processing chains for multimodel ISW analysis.
1 Introduction
ISW remote-sensing research spans detection, stripe extraction, and propagation prediction, but its tools and outputs remain fragmented. ISW-Agent addresses the resulting handoff and validation challenges by separating natural-language interpretation from deterministic execution.
1 Introduction
- Connecting detection, stripe extraction, and propagation prediction turns scene-level wave-packet detections into spatially and temporally referenced analysis products.SAR records ISWs as bright and dark bands caused by short-wave roughness modulation.Source: §1 · p. 1
- Earlier ISW methods advanced across detection, extraction, and prediction, but were typically tailored to specific inputs and evaluation objectives.Moving images, boxes, masks, and crestline coordinates between methods still requires data preparation and parameter configuration.Source: §1 · p. 1
- Users must transfer intermediate products and interpret service-specific fields, coordinates, and formats; a successful response may still produce an output unsuitable for the next stage.A truncated contour can remain writable while changing the spatial extent used to construct the initial wavefront.Source: §1 · p. 1
- ISW-Agent separates natural-language interpretation from deterministic execution to support task interpretation, spatial-product transfer, and explicit service-failure handling.The architecture is evaluated through dialogue contracts, spatial-representation preservation, execution time, and terminal outcomes under controlled faults.Source: §1 · p. 2
2 System Design
ISW-Agent combines dialogue understanding, n8n orchestration, and scientific model services to coordinate ISW processing. Explicit task constraints, data dependencies, and stage-specific checks govern transfers from requests through model outputs.
2 System Design
- Three layers divide ISW-Agent into dialogue understanding, domain-aware orchestration, and scientific model services.Dialogue specifies tasks, orchestration applies execution constraints, and model services perform scientific computation.Source: §2 · p. 2, Figure 1 · p. 2
2.1 Dialogue Understanding Layer
- Each request becomes a task contract that combines ordered intents, spatiotemporal slots, sensor and data specifications, model parameters, and output and provenance requirements.The LLM proposes values; the controller supplies data specifications from the asset registry and fixed service configurations for parameters and outputs.Source: §2.1 · p. 3
- Dialogue rules retain known information and request clarification for missing or ambiguous fields, while rejecting out-of-domain or unsupported requests before spatial processing.Prediction requires a crestline extracted in the same run, but does not automatically add detection.Source: §2.1 · p. 3
2.2 Domain-Aware Orchestration Layer
- Orchestration links independent models through data dependencies and allows each stage to proceed only when required inputs and validated preceding outputs are available.Checks cover readability, coordinates, geometry, and count consistency; empty extraction is valid but supplies no crestline for prediction.Source: §2.2 · p. 3
- The shared controller centralizes parameter validation, retries, and output checks, while bounded retries address transient interruptions and invalid spatial structures stop further processing.Each scientific service call has a 300 s deadline; completed outputs are preserved when processing terminates.Source: §2.2 · p. 3
2.3 Model Service Encapsulation
- The model-service layer links detection, stripe extraction, and prediction, passing the extracted vector file to prediction within the same run.The selected stripe’s georeferenced polygon is reduced to a centerline resampled to 200 points for pointwise propagation calculations.Source: §2.3 · p. 4
- Predicted speed and azimuth determine spherical displacement over the specified forecast interval, after which finite-valued points are filtered and fitted with a weighted spline.Postprocessing applies a 10 km coastal exclusion and a 40 km distance constraint relative to the reference point set.Source: §2.3 · p. 4
3 Results and Discussion
The results evaluate ISW image data, dialogue decisions, end-to-end spatial processing, execution time, and fault handling, with limitations in instruction coverage and operational scope. Tests show strong task matching and spatial processing, while content-consistency errors remain undetected by structural validation.
3.1 Image Acquisition and Model Data
- The 50 Sentinel-1 scenes were from the Maluku Sea, acquired from 19 February 2022 to 17 May 2026, and processed as 100 m SAR quicklooks.Propagation inputs used GEBCO 2024 bathymetry and WOA2023 monthly temperature and salinity.Source: §3.1 · p. 4
3.2 Dialogue Understanding and Execution Decisions
- GPT-5.6 Sol and DeepSeek each achieved a 94.2% strict task-match rate, with all 120 operation intents and execution decisions correct.Each backend also achieved slot F1 of 0.978; seven forecast-interval representation mismatches still triggered clarification.Source: §3.2 · pp. 5–6, Figure 3 · p. 6
3.3 End-to-End ISW Processing and Propagation Prediction
- Sentinel-1 case studies linked detections and stripe geometries to forecast fronts, including a 4 h forecast for Maluku Sea ISW 5 and 4 h and 8 h forecasts in Andaman and Celebes Sea cases.The selected Maluku Sea crestline was approximately 65.7 km long; the Celebes Sea case had a 92.3 km initial crestline.Source: §3.3 · pp. 6–7, Figure 4 · p. 7, Figure 5 · p. 8
- Full-pixel vectorization produced 21 valid geometries with IoU = 1, whereas the original export produced 14 invalid polygons and median IoU of 0.682.The tested conversion preserved mask extents, and the maximum pixel-to-geographic round-trip residual was well below one pixel.Source: §3.3 · pp. 7–8
3.4 Workflow Execution Time
- Across 50 Sentinel-1 scenes, 39 completed detection, extraction, and 3 h prediction with a median turnaround of 48.04 s; 11 stopped after empty extraction.The completed scenes had an interquartile range of 47.08–49.92 s; model services accounted for most time cost.Source: §3.4 · p. 9, Figure 6 · p. 9
3.5 Fault Handling and Strategy Ablation
- All 234 transient-failure conditions recovered within three attempts, and validation rejected all 234 structural defects; all configurations missed 61 content errors.The missed errors comprised 39 wrong-scene products and 22 feature-omission cases with consistent declared counts; these counts describe controlled faults, not routine incidence.Source: §3.5 · p. 10, Figure 7 · p. 10
3.6 Limitations
- Evaluation used rule-generated instructions and mainly serial timing and controlled fault tests on Maluku Sea imagery; regional transfer, concurrent users, and long-term operation remain unquantified.Detection and stripe extraction reuse an existing YOLO11 model, whose accuracy and regional transferability require further evaluation.Source: §3.6 · p. 10
4 Conclusion
ISW-Agent links natural-language understanding, n8n orchestration, and scientific models through explicit task constraints and geospatial transformations. Tests show strong task matching and end-to-end processing, while fault tests reveal that provenance and completeness still need content-consistency checks.
4 Conclusion
- ISW-Agent connects LLM-based natural-language understanding, n8n orchestration, and ISW models through explicit input constraints and geospatial transformations.Stage-specific checks govern data transfer, linking independently configured algorithms into coordinated processing with traceable intermediate products.Source: §4 · pp. 10–11
- 39 of 50 timed scenes completed detection, extraction, and prediction, with a median turnaround time of 48.04 s; 11 stopped after empty extraction.Across 120 controlled instructions, GPT-5.6 Sol and DeepSeek achieved a strict task-match rate of 94.2%, with all operation intents and execution decisions matching the reference.Source: §4 · p. 11
- Retries recovered 234 transient interruptions and structural checks rejected 234 invalid products, but 61 undetected content errors point to a need for provenance and completeness checks.The fault ablation distinguishes service recovery from scientific-output validation.Source: §4 · p. 11