Methods and equipment for early warning of flash floods and landslides

By performing event time correction and source tracing on multi-source observation data, allocating rainfall to flash flood and slope response branches, establishing conditional response time windows and directed event interpretation diagrams, the problems of response mismatch and data misjudgment in flash flood and landslide early warning were solved, and integrated early warning of flash flood and landslide disasters was realized.

CN123072175APending Publication Date: 2026-11-10WUYI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611392927.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-09
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing flash flood and landslide early warning methods suffer from problems such as delays in multi-source observation systems, mismatches in fixed time windows, duplicate weighting of derived indicators from the same source, and misjudgments and response mismatches caused by missing data. These issues make it difficult to achieve asynchronous escalation and escalation early warnings for flash floods, landslides, and cascading consequences.

Method used

Event records are generated by performing event time correction, source tracing, and data lineage marking on multi-source observation data. Based on rainfall input and slope status, the data is assigned to flash flood and slope response branches. Conditional multi-layer response time windows and directed event interpretation graphs are established. Candidate event verification and alternative interpretation structure search are performed to ensure asynchronous and graded early warning of flash floods, landslides, and cascading events.

Benefits of technology

It reduces pseudo-time delays caused by delays in multi-source observation systems, avoids response mismatches, suppresses redundant weighting of homogeneous derived indicators, prevents misjudgments due to missing data, and enables asynchronous escalation and escalation early warnings for flash floods, landslides, and blockage-break cascade consequences.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123072175A_ABST
    Figure CN123072175A_ABST
Patent Text Reader

Abstract

This invention provides a method and device for early warning of flash floods and landslides. The method includes: performing event time correction, source tracing, and quality labeling on rainfall, river hydrology, and slope hydro-mechanical data; allocating rainfall to flash flood response branches and slope response branches based on infiltration capacity, pre-wetting conditions, and fissure status, and establishing conditional multi-layer response time windows and directed event graphs; removing candidate nodes and their dependent nodes, and re-searching alternative interpretation structures for subsequent observation sets of the same target, confirming or rejecting candidates based on interpretation cost increment, uncertainty increment, coverage loss, independent evidence, and data integrity; and updating the cascade states of flash floods, landslides, and dammed-out floods asynchronously, thereby achieving integrated early warning of flash floods and landslides. This flash flood and landslide early warning method is applicable to monitoring and early warning of flash floods and landslides induced by heavy rainfall in small mountain watersheds.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and early warning technology, and in particular to a method and device for early warning of flash floods and landslides. Background Technology

[0002] Short-duration heavy rainfall can generate surface runoff, gully confluence, and river flood peaks within a small watershed. Flash flood responses are typically characterized by rapid, sudden, and network-wide propagation. Rainfall can also alter slope moisture content, groundwater level, pore water pressure, and matrix suction through matrix infiltration, fissure preferential flow, and riverbank exchange, leading to reduced shear strength, localized damage, enhanced microvibration, and accelerated displacement. While flash flood processes and slope instability processes share the same rainfall input, their response rates, propagation media, and recovery times differ.

[0003] Existing early warning methods often use cumulative rainfall over a fixed period, uniform empirical thresholds, single water level rises, or single slope parameters as triggering conditions. If flash floods and landslides share the same fixed time window, a longer window will smooth out minute-level abrupt changes in flash floods, while a shorter window will miss delayed responses such as the advance of wetting fronts, peak pore pressure, and post-rain displacement acceleration. If the cessation of rainfall is used as a common release condition, the slope risk may be prematurely released after the flood peak recedes. Furthermore, existing methods usually focus on judging the intensity of disaster response based on observed changes, making it difficult to determine whether a candidate disaster event has an irreplaceable explanatory contribution to subsequent observations of the target when there are time biases in multi-source observations, duplication of evidence derived from the same source, and competing explanations from multiple causes.

[0004] Sites, radar, satellites, forecasts, river monitoring, slope sensing, and remote sensing imagery have different source times, reception times, spatial scales, sampling frequencies, sensor responses, and communication delays. If ordered directly based on platform reception time, false chronological relationships may emerge between multi-source observations. Furthermore, high-frequency data such as micro-vibrations and slow variables such as rainfall and pore pressure are difficult to compare traceably within the same event chain without multi-rate aggregation.

[0005] Water level, rise, slope, and abrupt change index calculated from the same water level sensor, or multiple segmentation indicators derived from the same remote sensing image, share a common source of error. Repeatedly weighting these as multiple independent pieces of evidence may lead to false high-confidence candidates when a single device drifts or a single image is confused. Static susceptibility, model probability, and algorithm anomaly scores can be used for candidate recall and path ranking, but cannot independently prove that a dynamic event is in a true catastrophic chain.

[0006] The same anomalous observation can be explained by different causes. For example, increased pore pressure may be caused by current rainfall infiltration, previous water storage, river recharge, pipeline leakage, or sensor drift; enhanced microvibrations may be caused by slope damage, construction, earthquakes, rockfalls, or equipment noise; and remotely sensed floods or landslides may be caused by permanent water bodies, shadows, bare land, vegetation changes, or registration errors. Simply selecting the path with the highest score or the strongest correlation does not necessarily indicate whether a candidate makes a substantial contribution to interpreting subsequent observations.

[0007] Existing methods rarely employ the following structured verification: first, obtain a baseline interpretation structure containing the candidate to be verified; then, rigorously remove the candidate and all intermediate states, model outputs, and alarm evidence derived from it; finally, re-search for alternative interpretation structures for subsequent observation sets of the same target, and compare the interpretation cost, prediction uncertainty, and observation coverage before and after deletion. Without this verification, derived evidence generated by the candidate itself may backfire to prove the candidate, forming a circular proof; transitions that accidentally fall within a wide time window may also be mistakenly incorporated into propagation relationships.

[0008] Landslides can also enter gullies, forming blocking dams, leading to backwater, seepage erosion, overtopping, and ultimately, breach flooding. If flash floods, landslides, and cascading consequences are consistently assessed using a single level or simply taking the maximum value, it fails to reflect the situation where a flash flood has receded while the slope continues to deteriorate, or the escalating downstream flood risk after a landslide blocks a river. Therefore, establishing an integrated early warning method that unifies input, branch responses, candidate verification, disaster type confirmation, and asynchronous recovery has become an urgent technical problem to be solved. Summary of the Invention

[0009] This invention provides a method and device for early warning of flash floods and landslides, which can reduce pseudo-time delays caused by delays in multi-source observation systems, avoid response mismatches caused by flash floods and landslides sharing a fixed time window, suppress redundant weighting and candidate self-verification of derived indicators, prevent "confirmation without substitution" caused by missing data or insufficient model applicability domain, and enable the asynchronous escalation and escalation of flash floods, landslides, and blockage-break cascade consequences according to their respective physical recovery processes, thereby achieving integrated early warning of flash floods and landslides.

[0010] In a first aspect, embodiments of the present invention provide a method for early warning of flash floods and landslides, including: Acquire multi-source observation data of the target area, and perform event time correction, source tracing and data lineage marking on the multi-source observation data to generate event records with event time, event time uncertainty interval, original source identifier, dependency group identifier, data lineage and evidence quality; Based on rainfall input, previous wet conditions, state-related infiltration capacity, and fissure conditions, rainfall is allocated to flash flood response branches and slope hydro-mechanical response branches to generate flash flood candidate events and slope response candidate events, respectively. Based on the parent event, location, monitoring parameters, and current state, a conditional multi-layer response time window is established, and based on the conditional multi-layer response time window and spatial topology, a directed event interpretation graph containing candidate event nodes, observation nodes, derived dependency nodes, and alternative cause nodes is constructed. For a candidate event to be verified, fix the subsequent observation set of the target and search the complete directed event interpretation graph for a benchmark interpretation structure that contains the candidate event to be verified and can cover the subsequent observation set of the target. Perform a counterfactual deletion operation to remove the candidate event to be verified and the data lineage closure derived only from it from the directed event interpretation graph, forming a pruned directed event graph, and then search for alternative interpretation structures in the pruned directed event graph for subsequent observation sets of the same target. Based on the existence of alternative interpretation structures, the increment of interpretation cost, the increment of uncertainty, the loss of observation coverage, the independent evidence index, and the data integrity, the non-substitutable contribution state of the candidate event to be verified is determined, and the candidate event to be verified is confirmed or rejected based on the non-substitutable contribution state. Based on confirmed flash flood events, landslide events, and cascading events, the status of flash floods, landslides, and cascading events are updated respectively, and asynchronous graded early warnings are implemented according to disaster type.

[0011] In some embodiments, the event time correction of the multi-source observation data includes: When the source clock quality meets the preset conditions, the trusted source time correction mode is adopted. The event time is obtained by mapping the source device clock to a unified clock and subtracting the sensing delay. When the source clock quality does not meet the preset conditions, the receiving time backtracking mode is adopted. The event time is obtained by subtracting the sensing delay, processing delay and communication delay from the platform receiving time in sequence. The event time uncertainty interval is determined by the upper bound of the non-negative uncertainty of the clock error, sensing error, processing error, and communication error involved in the time estimation.

[0012] In some embodiments, the distribution of rainfall to flash flood response branches and slope hydro-mechanical response branches includes: The distribution ratio of rapid flow in the fractures is determined based on fracture aperture, connectivity, surface cover, and pre-wetting conditions. Based on the fissure rapid flow allocation ratio, the distributable rainfall after deducting prior losses is allocated as fissure rapid flow and matrix infiltration; wherein, the excess runoff is used as the input of the flash flood response branch, and the matrix infiltration and fissure rapid flow are used as the input of the slope hydro-mechanical response branch; when there are independent channel outflow observations, the event-scale water consistency residual is calculated and used as evidence of data conflict.

[0013] In some embodiments, the conditionalized multi-layer response time window includes a rainstorm and pre-rain condition window, a flash flood rapid response window, a slope hydrological response window, a micro-vibration mechanical response window, a deformation instability response window, a cascade consequence window, and a post-rain recovery window; wherein, when the predicted response interval of the parent event and the event time interval of the child observed event intersect, the time matching is determined to be valid; the physical response delay and the event time uncertainty interval of the parent and child events are represented separately, and are updated in groups according to the site's historical confirmed events.

[0014] In some embodiments, constructing a directed event interpretation graph comprising candidate event nodes, observation nodes, derived dependency nodes, and alternative cause nodes includes: Generate candidate event nodes in the spatial topology of slope unit-valley-river channel-disaster-bearing body; Based on the conditions of time matching, spatial connectivity, propagation direction permission, physical consistency satisfaction, evidence quality reaching the edge construction threshold, and being within the model's applicable domain, directed edges are established between candidate event nodes. Derived nodes originating from the same original observation or the same model output are grouped into the same dependency group based on their original source identifier and data lineage.

[0015] In some embodiments, both the baseline interpretation structure and the alternative interpretation structure consist of one or more directed paths capable of covering the subsequent observation set of a fixed target; the interpretation cost of the interpretation structure is a weighted sum of time interval bias cost, spatial propagation bias cost, propagation mechanism prior bias cost, evidence quality cost, model applicability domain distance cost, and normalized uncertainty cost; wherein, topological disconnection, propagation direction disallowance, and unacceptable temporal conflicts or physical conservation conflicts serve as hard constraints, directly rendering the corresponding directed edges unusable.

[0016] In some embodiments, determining the non-substitutable contribution state of the candidate event to be verified includes: Calculate the observation coverage, overall uncertainty, and key data integrity of the baseline and alternative interpretation structures; Calculate the increment of interpretation cost, the increment of uncertainty, and the loss of observation coverage before and after deleting the candidate event to be verified; The independent evidence index is calculated based on the independent evidence groups after deduplication by source and common mode correlation. When the baseline observation coverage and key data integrity reach the corresponding thresholds, and the incremental cost of no feasible alternatives or explanations, the incremental uncertainty and the loss of observation coverage all reach the corresponding thresholds, and the independent evidence index reaches the corresponding threshold, the candidate event to be verified is confirmed. When a low-cost, low-uncertainty, and sufficiently comprehensive alternative explanation structure exists, the state level of the candidate event to be verified is rejected or reduced. When key data is incomplete, time synchronization is unavailable, the model's applicable domain is exceeded, or the evidence window is not closed, the candidate event to be verified will be set to a state of insufficient data or pending review.

[0017] In some embodiments, the execution of asynchronous, graded early warning based on disaster type includes: The status of flash floods is upgraded or downgraded independently based on the propagation and receding of the flood peak; The landslide status is independently downgraded after the slope hydro-mechanical and deformation status enters the preset recovery domain and remains there for a preset time; however, the receding of flash flood peaks does not automatically trigger the landslide status to be lifted, the confirmation of a landslide event does not automatically confirm a blockage event, and the confirmation of a blockage event does not automatically confirm a breach event.

[0018] In some embodiments, it also includes: Only confirmed events with clear causes, traceable time, and verified through candidate event contribution are subject to controlled updates of response time windows, explanatory cost weights, or dynamic thresholds. Before updating, save the current parameter version and rollback point. After updating, use independent time period playback, subsequent observation or subsequent confirmation events for verification. If the verification result does not meet the preset conditions, restore the version before the update. After the action is taken, the execution feedback is recorded and the affected hydrological-slope state, response time window, directed event graph and disaster type level are recalculated. The execution feedback is not separated from independent measured evidence and is used as the sole basis for confirming the authenticity of the candidate event.

[0019] In a second aspect, embodiments of the present invention also provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the flash flood and landslide disaster early warning method as described in the first aspect.

[0020] According to embodiments of the present invention, a method and device for early warning of flash floods and landslides are provided. The method includes: acquiring multi-source observation data of a target area, and performing event time correction, source tracing, and data lineage marking on the multi-source observation data to generate event records with event time, event time uncertainty interval, original source identifier, dependency group identifier, data lineage, and evidence quality; allocating rainfall to flash flood response branches and slope hydro-mechanical response branches based on rainfall input, previous wetting state, state-related infiltration capacity, and fracture state to generate flash flood candidate events and slope response candidate events respectively; establishing a conditional multi-layer response time window based on the parent event, location, monitoring parameters, and current state, and constructing a directed event interpretation graph containing candidate event nodes, observation nodes, derived dependency nodes, and alternative cause nodes based on the conditional multi-layer response time window and spatial topological relationships. For candidate events to be verified, the subsequent observation set of the target is fixed, and a baseline interpretation structure containing the candidate event to be verified and covering the subsequent observation set of the target is searched in the complete directed event interpretation graph. Counterfactual deletion is performed to remove the candidate event to be verified and the data lineage closure derived only from it from the directed event interpretation graph, forming a pruned directed event graph. For the same subsequent observation set of the target, alternative interpretation structures are searched again in the pruned directed event graph. Based on the existence of alternative interpretation structures, the increment of interpretation cost, the increment of uncertainty, the loss of observation coverage, the independent evidence index, and the data integrity, the irreplaceable contribution state of the candidate event to be verified is determined, and the candidate event to be verified is confirmed or rejected based on the irreplaceable contribution state. Based on the confirmed flash flood events, landslide events, and cascading events, the flash flood state, landslide state, and cascading state are updated respectively, and asynchronous graded early warning is performed according to the disaster type. This invention employs event time correction, source tracing, and quality labeling to rainfall, river hydrology, and slope hydro-mechanical data. Based on infiltration capacity, prior wetting, and fissure status, rainfall is allocated to flash flood response branches and slope response branches, establishing conditional multi-layered response time windows and directed event graphs. For candidates to be validated, their dependent nodes and those derived solely from them are removed. For subsequent observation sets of the same target, alternative interpretation structures are re-searched, and candidates are confirmed or rejected based on interpretation cost increment, uncertainty increment, coverage loss, independent evidence, and data integrity. The states of flash floods, landslides, and dammed-out cascades are updated asynchronously. This invention is applicable to monitoring and early warning of flash floods and landslides induced by heavy rainfall in small mountain watersheds. Based on this, the embodiments of the present invention can reduce the pseudo-time delay caused by the delay of multi-source observation systems, avoid response mismatch caused by flash floods and landslides sharing a fixed time window, suppress the repeated weighting and candidate self-verification of derived indicators from the same source, prevent "confirmation without substitution" caused by missing data or insufficient model applicability domain, and enable the consequences of flash floods, landslides and blockage-break cascades to be asynchronously upgraded and downgraded according to their respective physical recovery processes, thereby realizing integrated early warning of flash flood and landslide disasters. Attached Figure Description

[0021] Figure 1A This is a main flowchart of a flash flood and landslide disaster early warning method provided in one embodiment of the present invention; Figure 1B This is an embodiment of the present invention providing a candidate event alternative path verification and an overall closed-loop diagram for asynchronous early warning of flash floods and landslides; Figure 2 An eight-step flowchart of candidate event generation, counterfactual deletion, and alternative interpretation verification is provided for one embodiment of the present invention; Figure 3 This is a schematic diagram of the topology of slope unit-valley-river channel-disaster-bearing body, monitoring deployment, edge evidence package and system connection provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of storm input state-related diversion, independent water volume consistency, and dual-branch system provided in one embodiment of the present invention; Figure 5 A schematic diagram of a conditional multi-layer hydrological-slope response time window provided in one embodiment of the present invention; Figure 6 A candidate event diagram, direct parent node, and alternative cause diagram are provided as an embodiment of the present invention; Figure 7 A flowchart of candidate event counterfactual deletion and alternative path reconstruction verification provided in one embodiment of the present invention; Figure 8 A schematic diagram illustrating disaster type confirmation, cascaded candidate step-by-step verification, joint early warning, and asynchronous upgrade / downgrade according to an embodiment of the present invention; Figure 9 This is a schematic diagram of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the following drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0024] In this embodiment of the invention, the terms "furthermore," "exemplarily," or "optionally" are used as examples, illustrations, or descriptions and should not be construed as being more preferred or advantageous than other embodiments or designs. The use of the terms "furthermore," "exemplarily," or "optionally" is intended to present the relevant concepts in a specific manner.

[0025] To facilitate a more convenient description of the working principle of the embodiments of the present invention, the following introduction of relevant technical scenarios is given first.

[0026] Short-duration heavy rainfall can generate surface runoff, gully confluence, and river flood peaks within a small watershed. Flash flood responses are typically characterized by rapid, sudden, and network-wide propagation. Rainfall can also alter slope moisture content, groundwater level, pore water pressure, and matrix suction through matrix infiltration, fissure preferential flow, and riverbank exchange, leading to reduced shear strength, localized damage, enhanced microvibration, and accelerated displacement. While flash flood processes and slope instability processes share the same rainfall input, their response rates, propagation media, and recovery times differ.

[0027] Existing early warning methods often use cumulative rainfall over a fixed period, uniform empirical thresholds, single water level rises, or single slope parameters as triggering conditions. If flash floods and landslides share the same fixed time window, a longer window will smooth out minute-level abrupt changes in flash floods, while a shorter window will miss delayed responses such as the advance of wetting fronts, peak pore pressure, and post-rain displacement acceleration. If the cessation of rainfall is used as a common release condition, the slope risk may be prematurely released after the flood peak recedes. Furthermore, existing methods usually focus on judging the intensity of disaster response based on observed changes, making it difficult to determine whether a candidate disaster event has an irreplaceable explanatory contribution to subsequent observations of the target when there are time biases in multi-source observations, duplication of evidence derived from the same source, and competing explanations from multiple causes.

[0028] Sites, radar, satellites, forecasts, river monitoring, slope sensing, and remote sensing imagery have different source times, reception times, spatial scales, sampling frequencies, sensor responses, and communication delays. If ordered directly based on platform reception time, false chronological relationships may emerge between multi-source observations. Furthermore, high-frequency data such as micro-vibrations and slow variables such as rainfall and pore pressure are difficult to compare traceably within the same event chain without multi-rate aggregation.

[0029] Water level, rise, slope, and abrupt change index calculated from the same water level sensor, or multiple segmentation indicators derived from the same remote sensing image, share a common source of error. Repeatedly weighting these as multiple independent pieces of evidence may lead to false high-confidence candidates when a single device drifts or a single image is confused. Static susceptibility, model probability, and algorithm anomaly scores can be used for candidate recall and path ranking, but cannot independently prove that a dynamic event is in a true catastrophic chain.

[0030] The same anomalous observation can be explained by different causes. For example, increased pore pressure may be caused by current rainfall infiltration, previous water storage, river recharge, pipeline leakage, or sensor drift; enhanced microvibrations may be caused by slope damage, construction, earthquakes, rockfalls, or equipment noise; and remotely sensed floods or landslides may be caused by permanent water bodies, shadows, bare land, vegetation changes, or registration errors. Simply selecting the path with the highest score or the strongest correlation does not necessarily indicate whether a candidate makes a substantial contribution to interpreting subsequent observations.

[0031] Existing methods rarely employ the following structured verification: first, obtain a baseline interpretation structure containing the candidate to be verified; then, rigorously remove the candidate and all intermediate states, model outputs, and alarm evidence derived from it; finally, re-search for alternative interpretation structures for subsequent observation sets of the same target, and compare the interpretation cost, prediction uncertainty, and observation coverage before and after deletion. Without this verification, derived evidence generated by the candidate itself may backfire to prove the candidate, forming a circular proof; transitions that accidentally fall within a wide time window may also be mistakenly incorporated into propagation relationships.

[0032] Landslides can also enter gullies, forming blocking dams, leading to backwater, seepage erosion, overtopping, and ultimately, breach flooding. If flash floods, landslides, and cascading consequences are consistently assessed using a single level or simply taking the maximum value, it fails to reflect the situation where a flash flood has receded while the slope continues to deteriorate, or the escalating downstream flood risk after a landslide blocks a river. Therefore, establishing an integrated early warning method that unifies input, branch responses, candidate verification, disaster type confirmation, and asynchronous recovery has become an urgent technical problem to be solved.

[0033] Based on this, the present invention provides a method and device for early warning of flash floods and landslides. The method includes: acquiring multi-source observation data of the target area, and performing event time correction, source tracing, and data lineage marking on the multi-source observation data to generate event records with event time, event time uncertainty interval, original source identifier, dependency group identifier, data lineage, and evidence quality; allocating rainfall to flash flood response branches and slope hydro-mechanical response branches based on rainfall input, previous wet conditions, state-related infiltration capacity, and fracture state to generate flash flood candidate events and slope response candidate events respectively; establishing a conditionalized multi-layer response time window based on the parent event, location, monitoring parameters, and current state, and constructing a directed event interpretation graph containing candidate event nodes, observation nodes, derived dependency nodes, and alternative cause nodes based on the conditionalized multi-layer response time window and spatial topological relationships; and fixing the data for the candidate events to be verified. For the target's subsequent observation set, a baseline interpretation structure containing the candidate event to be verified and capable of covering the target's subsequent observation set is searched in the complete directed event interpretation graph. A counterfactual deletion operation is performed to remove the candidate event to be verified and the data lineage closure derived only from it from the directed event interpretation graph, forming a pruned directed event graph. For the same target's subsequent observation set, an alternative interpretation structure is searched again in the pruned directed event graph. Based on the existence of the alternative interpretation structure, the increment of interpretation cost, the increment of uncertainty, the loss of observation coverage, the independent evidence index, and the data integrity, the irreplaceable contribution status of the candidate event to be verified is determined, and the candidate event to be verified is confirmed or rejected based on the irreplaceable contribution status. Based on the confirmed flash flood event, landslide event, and cascading event, the flash flood status, landslide status, and cascading status are updated respectively, and asynchronous graded early warning is performed according to the disaster type. This invention employs event time correction, source tracing, and quality labeling to rainfall, river hydrology, and slope hydro-mechanical data. Based on infiltration capacity, prior wetting, and fissure status, rainfall is allocated to flash flood response branches and slope response branches, establishing conditional multi-layered response time windows and directed event graphs. For candidates to be validated, their dependent nodes and those derived solely from them are removed. For subsequent observation sets of the same target, alternative interpretation structures are re-searched, and candidates are confirmed or rejected based on interpretation cost increment, uncertainty increment, coverage loss, independent evidence, and data integrity. The states of flash floods, landslides, and dammed-out cascades are updated asynchronously. This invention is applicable to monitoring and early warning of flash floods and landslides induced by heavy rainfall in small mountain watersheds. Based on this, the embodiments of the present invention can reduce the pseudo-time delay caused by the delay of multi-source observation systems, avoid response mismatch caused by flash floods and landslides sharing a fixed time window, suppress the repeated weighting and candidate self-verification of derived indicators from the same source, prevent "confirmation without substitution" caused by missing data or insufficient model applicability domain, and enable the consequences of flash floods, landslides and blockage-break cascades to be asynchronously upgraded and downgraded according to their respective physical recovery processes, thereby realizing integrated early warning of flash flood and landslide disasters.

[0034] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0035] like Figure 1A As shown, Figure 1A This is a flowchart of a flash flood and landslide disaster early warning method provided in an embodiment of the present invention. The flash flood and landslide disaster early warning method may include, but is not limited to, steps S101 to S107.

[0036] Step S101: Acquire multi-source observation data of the target area, and perform event time correction, source tracing and data lineage marking on the multi-source observation data to generate event records with event time, event time uncertainty interval, original source identifier, dependency group identifier, data lineage and evidence quality; Step S102: Based on rainfall input, previous wetting status, state-related infiltration capacity and fracture status, the rainfall is allocated to flash flood response branches and slope hydro-mechanical response branches to generate flash flood candidate events and slope response candidate events respectively. Step S103: Based on the parent event, location, monitoring parameters and current state, establish a conditional multi-layer response time window, and based on the conditional multi-layer response time window and spatial topology, construct a directed event interpretation graph containing candidate event nodes, observation nodes, derived dependency nodes and alternative cause nodes; Step S104: For the candidate event to be verified, fix the subsequent observation set of the target, and search in the complete directed event interpretation graph for a benchmark interpretation structure that contains the candidate event to be verified and can cover the subsequent observation set of the target. Step S105: Perform counterfactual deletion operation to remove the candidate events to be verified and the data lineage closures derived only from them from the directed event interpretation graph, forming a pruned directed event graph, and for subsequent observation sets of the same target, search for alternative interpretation structures again in the pruned directed event graph; Step S106: Determine the non-substitutable contribution status of the candidate event to be verified based on the existence of alternative interpretation structures, interpretation cost increment, uncertainty increment, observation coverage loss, independent evidence index, and data integrity, and confirm or reject the candidate event to be verified based on the non-substitutable contribution status. Step S107: Based on the confirmed flash flood events, landslide events, and cascading events, update the flash flood status, landslide status, and cascading status respectively, and execute asynchronous graded early warning for different disaster types.

[0037] Understandably, this method involves performing event time correction, source tracing, and quality labeling on rainfall, river hydrology, and slope hydro-mechanical data; allocating rainfall to flash flood response branches and slope response branches based on infiltration capacity, pre-wetting conditions, and fissure status; establishing conditional multi-layer response time windows and directed event graphs; removing candidate data and their dependent nodes; re-searching alternative interpretation structures for subsequent observation sets of the same target; and confirming or rejecting candidates based on interpretation cost increment, uncertainty increment, coverage loss, independent evidence, and data integrity; and updating the flash flood, landslide, and dam-break cascade states asynchronously. This flash flood and landslide disaster early warning method is applicable to monitoring and early warning of flash floods and landslides induced by heavy rainfall in small mountain watersheds.

[0038] It is understood that the overall framework of the present invention is as follows: Figure 1B As shown, the system first generates multi-source event records with event time, uncertain interval of event time, original source identifier, data lineage and evidence quality; then, the rainstorm input is allocated to the flash flood response branch and the slope hydro-mechanical response branch according to state-related infiltration capacity, previous wetting and fissure state; then, conditional response time windows are established according to parent event, location, depth, parameters and current state, and a directed event graph is generated in the topology of slope unit-gully-river-disaster-bearing body.

[0039] Understandably, for each candidate to be verified, the system fixes the target subsequent observation set that does not belong to the candidate and its derived dependency set before deletion. It then searches the complete directed event graph for a baseline interpretation structure that includes the candidate and covers the observation set, recording its interpretation cost, overall uncertainty, and observation coverage. Based on the original source identifier and data lineage, the system forms a dependency set derived only from the candidate. It removes the candidate and its dependency set from the available node and edge sets in the alternative interpretation search, and re-searches for alternative interpretation structures for the same target subsequent observation set in the pruned directed event graph.

[0040] Understandably, when a feasible alternative explanation structure exists, the system determines the candidate's irreplaceable contribution based on its incremental explanation cost relative to the baseline explanation structure, the incremental comprehensive uncertainty, the observation coverage loss, and the independent evidence index after data lineage deduplication. When a feasible alternative explanation structure does not exist, the optimal solution is not computed on an empty set. Instead, key observations, communication, time synchronization, and model applicability domain are checked first. Only after the data integrity check is passed is "no feasible alternative" considered as evidence of an irreplaceable contribution. Low-cost and adequately covered alternative explanations can explain when a candidate is rejected or downgraded during observation; when key data is not available or the evidence window is not closed, the candidate enters a state of insufficient data or awaiting review. When configuring rolling prediction or digital twin models, the bi-branch results of retaining and deleting candidate effects can be used as auxiliary verification quantities, but they do not replace independent measured evidence and alternative explanation re-search.

[0041] It is understood that this invention maintains separate states for flash floods, landslides, and landslide blockage-breakdown cascade states, and forms a joint early warning state based on confirmed event combinations, river-slope toe coupling, landslide entry into gullies, upstream backwater, disaster stage, estimated remaining treatment time, and exposure relationship of affected objects. Flash flood states are upgraded or downgraded based on the arrival, propagation, and receding of the flood peak; landslide states are downgraded based on whether the slope's hydrological, mechanical, and deformation parameters enter the recovery domain and remain there for a certain period; after confirming that a landslide may enter a gully, candidate landslides are generated for gully entry, blockage, backwater, seepage erosion, overtopping, and breakdown, and each candidate is validated using a similar alternative interpretation.

[0042] Understandably, this invention also sets boundaries for model and parameter update permissions. Static susceptibility, remote sensing probability, machine learning anomaly scores, and physical model predictions are used for candidate generation, path priors, explanation structure ranking, window initials, and uncertainty estimation, but are not used to separately confirm the highest internal risk level. Only confirmed events with clear causes, traceable timelines, and verified through alternative explanation contributions are allowed to update the core response time window, explanation cost weights, or dynamic thresholds. Before updating, parameter versions and rollback points are saved. After updating, independent time-segment replay, subsequent observations, or subsequent confirmed events are used for verification. If the verification does not meet preset conditions, the previous version is restored. After the action is executed, the execution feedback is recorded and the affected state is recalculated. The execution feedback is not separated from independent measured evidence to separately confirm the disaster event.

[0043] It is understood that, compared with fixed threshold or single-path scoring methods, this invention has the following beneficial effects: First, by correcting the time of trusted sources or backtracking the receiving time and explicitly transmitting the time uncertainty intervals of parent and child events, system delay and physical hysteresis are separated; Second, by using state-related rainfall diversion and conditional multi-layer response time windows, both rapid response to flash floods and hydro-mechanical hysteresis response of slopes are taken into account; Third, by using original source identification and data lineage aggregation of homologous derived evidence, circular proofs and duplicate weighting are reduced; Fourth, by using fixed target subsequent observation sets, strict removal of candidates and their derived dependencies, alternative interpretation re-search, and data integrity gating for infeasible alternatives, the verifiability of candidate irreplaceable contributions is enhanced; Fifth, by using disaster type confirmation, cascaded step-by-step verification, and independent recovery gating, the premature removal of landslides after rain and the underreporting of downstream risks after blockage are reduced; Sixth, by using versioned parameters, controlled updates, and post-processing review, the traceability and rollback of the operating results are improved.

[0044] The following describes the implementation process of the present invention with reference to four typical embodiments. Specific implementation methods are as follows: Figure 2The eight core steps are as follows: event recording and unified time, rainstorm parent event and dual-branch diversion, conditional response time window, candidate events and directed event graph, baseline interpretation structure, execution of candidate event counterfactual deletion, removing candidate event h and the data lineage closure D(h) derived only from it from the event interpretation graph to form a pruned event graph G(-h), and re-searching for alternative interpretation structures for the same target observation set Y, contribution gating and candidate lifecycle, asynchronous state and cascading for different disaster types; risk state-driven sampling communication, versioned auditing, and controlled updates are described separately as preferred operating mechanisms. Among them, Example 1 is the main example of the entire chain of rainstorm-flash flood-delayed landslide-blockage-secondary flood; Example 2 illustrates that the flash flood rapid branch is dominant while the slope branch is not confirmed; Example 3 illustrates the delayed deterioration and independent recovery of the slope after the rain stops; Example 4 illustrates that equipment drift is rejected by disaster candidates through low-cost alternative interpretation. The specific values ​​in the following examples are all executable calculation examples and do not constitute a limitation on the scope of protection.

[0045] Step 1: Establish a monitoring system, slope-ditch-river topology, and record and unify the time of multi-source events.

[0046] like Figure 3 As shown, this embodiment establishes a spatial topology of slope unit—gully—river channel—disaster-bearing body in the target small watershed and configures monitoring nodes. The slope can be monitored according to the slope crest or trailing edge, middle or main landslide body, toe, and old landslide zone or deep unit. Monitoring nodes for rainfall, cracks, volumetric water content, pore water pressure, micro-vibration, earth pressure, displacement, or inclination angle are deployed at corresponding locations. Water level, flow rate, water depth, or flow velocity monitoring nodes are deployed in the gully, river channel, and its upstream and downstream areas. Downstream roads, bridges, or residential areas serve as disaster-bearing body nodes. A directed topological association is established between each slope unit and the gully, river segment, and disaster-bearing body based on topographic confluence, potential landslide movement, upstream-downstream propagation, and spatial adjacency, and the propagation distance, cross-section, flow capacity, and confluence time are recorded. The observations formed by each monitoring node or external data source are encapsulated into edge evidence packages by edge nodes and transmitted to the central computing platform for subsequent event time correction, response time window matching, candidate event graph construction and alternative interpretation verification; the central computing platform can also communicate with manual, drone or remote sensing field verification units and early warning release units. Figure 3 The spatial topology arrows and data connection arrows in the text represent physical adjacency or potential propagation direction and data transmission relationship, respectively, and are not directly equivalent to the directed event edges used in the confirmation of the authenticity of candidate events; the candidate event edges are still established according to the time matching, spatial connectivity, propagation direction, physical consistency, evidence quality and model applicability domain conditions specified in step four.

[0047] Table 1 Monitoring Objects, Locations, and Outputs

[0048] Each observation generated by each monitoring node or external data source is encapsulated into an edge evidence package by edge nodes. An edge evidence package includes at least the event identifier, device identifier, original source identifier, dependency group identifier, parent event identifier, source time, reception time, event time, event time uncertainty interval, location, slope unit, depth, sampling rate, observation value, unit, range marker, missing measurement rate, calibration time, communication status, model or rule version, data lineage, and original fragment reference. Means, slopes, increases, mutation indices, and model features calculated from the same original observation carry the same original source identifier and are associated through data lineage to form dependency sets.

[0049] The event time adopts either trusted source time correction mode or received time backtracking mode, and the event time interval is represented by equation (1): (1) In equation (1), Indicates the time of the event; Indicates the source time; This represents a correction function that maps the source device clock to a uniform clock. Indicates the platform's receiving time; , and These represent sensor response and acquisition delay, edge processing delay, and communication transmission delay, respectively. Indicates the time interval of the event; This represents the upper bound of the non-negative uncertainty formed by the clock, sensing, processing, and communication errors that actually participate in the corresponding time estimation. The corrected source time is used when the source clock quality meets the preset conditions; the receiving time is used to backtrack when the source clock is unavailable or does not meet the quality conditions. Equation (1) only corrects the system clock and system delay; the physical hysteresis caused by infiltration, confluence, pore pressure diffusion, and deformation development is represented separately in the response time window.

[0050] The quality of evidence is determined by formula (2): (2) In equation (2), Indicates the quality of evidence; Indicates sensor range, drift, calibration, and missing measurement quality; Indicates the quality of the data source; This indicates time freshness and time synchronization quality; Indicates communication quality in terms of packet loss, retransmission, and integrity. This represents the quality of the domain to which the model or rule applies. All quality quantities are normalized to zero or one. If any necessary quality quantity is below the rejection threshold, the corresponding observation can only form a candidate for anomaly or a state of insufficient data, and cannot support high-level confirmation on its own.

[0051] Rainfall, water content, groundwater, pore pressure, and slow displacement are sampled at the second to minute level; micro-vibrations and acoustic emissions are sampled at higher frequencies, and the root mean square value, peak value, frequency band energy, duration, and key original fragment fingerprints are output at the edge. The sampling and transmission frequency of associated nodes is increased in the candidate state, and key original fragments are retained in the confirmed state; features generated by different sampling frequencies are converted to a unified event clock and then used in mapping.

[0052] Step 2: Identify the parent event of the rainstorm and perform dual-branch diversion and hydrological-slope state calculations.

[0053] The system segments continuous rainfall at configurable silent intervals and, combined with station-radar or station-satellite consistency and spatial movement of the rain field, identifies the start, peak, and end of the parent event of the rainstorm. The attenuation of the initial rainfall is calculated using equation (3): (3) In equation (3), Indicates the first The attenuation of the initial rainfall over each calculation period; Indicates the first Rainfall over a calculation period; Indicates the backtracking sequence number; Indicates the number of backtracking periods; This represents the attenuation coefficient calibrated with season, soil layer, and evapotranspiration conditions. The anterior wet conditions can also be determined by combining the water content, groundwater level, pore pressure, and baseflow conditions before the event begins, rather than replacing the entire anterior conditions with cumulative rainfall over a fixed number of days.

[0054] like Figure 4 As shown, the system inputs distributable rainfall into the rapid flow channels in the fissures and the matrix infiltration channels respectively, based on the previous wetting, fissure state and state-related infiltration capacity, using Equation (4): (4) In equation (4), Indicates the first Distributable rainfall after deducting prior losses in each calculation cycle; Indicates the rapid flow rate of the fracture; The proportion of rapid flow distribution in fractures is determined by fracture aperture, connectivity, surface cover, and pre-wetting conditions. Indicates the amount of matrix infiltration; This represents the state-related matrix infiltration capacity consistent with the current calculation period and controlled area. Excess runoff. The data are obtained from the runoff generation and confluence model, cross-sectional flow inversion or other independent hydrological estimates and input into the flash flood response branch; matrix infiltration and fissure fast flow are input into the slope hydro-mechanical response branch; if the original infiltration capacity is given as a rate, it is first converted into equivalent water volume according to the calculation period before participating in equation (4).

[0055] To avoid using the same balance to infer excess runoff and then using that balance to verify itself, this implementation method uses event-scale independent water volume consistency residuals when there are independent river outflow observations: (5) In equation (5), This represents the event-scale water quantity consistency residual; Indicates rainfall at the event scale; This represents the outflow volume of the event obtained by integrating independent channel flow observations; Indicates the controlled area; and These represent event-scale matrix infiltration and fracture rapid flow, respectively. Indicates the amount of loss, evaporation, or other losses at the event scale; This represents changes in water storage on slopes, in soil layers, and in channels. The system reduces the quality of evidence for the corresponding interpretative structure or generates data conflict candidates only when independent outflow observations are available and the residuals exceed the tolerance determined by measurement error propagation; if independent outflow observations are unavailable, the residuals are used only as model numerical closure quantities.

[0056] In step two, the rapid response to flash floods and the hydro-mechanical state of the slope are further calculated.

[0057] The rapid branch road for flash floods calculates water level, flow rate, water depth, and flood peak based on slope runoff generation, gully confluence, and river propagation. Kinematic waves, shallow water equations, unit hydrographs, data-driven hydrological models, or equivalent propagation models calibrated in the field can be used. The continuity relationship in this implementation method is expressed by equation (6): (6) In equation (6), Indicates the cross-sectional area of ​​the water passage; Indicates physical time; Indicates cross-sectional flow rate; Indicates the spatial coordinates along the ditch or river channel; This indicates lateral inflow. The peak arrival time and prediction interval output by the model are used to determine the flash flood response window and path cost, while measured water level, flow rate, or spatial flood objects are used for independent validation. Water level rise, slope, and abrupt change index derived from the same water level sensor are grouped into the same dependency set.

[0058] The hydrological state of a slope is described by available parameters including water content, groundwater level, pore water pressure, matrix suction, and resistivity. The equivalent diffusion relationship of pore water pressure along the depth direction can be expressed as: (7) In equation (7), Indicates pore water pressure; Indicates the equivalent hydraulic diffusion coefficient; Indicates the depth of the slope; Represents physical time. When fast flow exists in fractures, the system sets fast channel nodes or edges, and does not force the use of a single diffusion coefficient to interpret the deep fast response.

[0059] The shear strength of unsaturated soil can be expressed as: (8) In equation (8), Indicates shear strength; Indicates effective cohesion; Indicates the total normal stress; Indicates pore gas pressure; Indicates pore water pressure; Indicates the effective internal friction angle; The strength parameter represents the increase in suction force. The material parameters are determined by field or laboratory tests. Equation (8) is used to form strength states and stability candidates, and landslides are not confirmed solely by fixed stability thresholds not specified by the field.

[0060] Displacement velocity and displacement acceleration are expressed by equation (9): (9) In equation (9), Indicates the first Displacement over one calculation cycle; Indicates displacement velocity; Indicates displacement acceleration; This represents a time interval greater than zero, determined by the time of adjacent valid events. Single-point abrupt changes in displacement, tilt, or cracks only generate deformation candidates, which need to be verified in conjunction with alternative interpretation contributions from different locations, different measurement mechanisms, or slope hydrological-mechanical evidence.

[0061] Step 3: Establish a multi-layered response time window with parameter-location-state conditions.

[0062] like Figure 5 As shown, the system maintains in parallel the following windows: rainstorm and pre-rainstorm state window, flash flood rapid response window, slope hydrological response window, micro-vibration-mechanical response window, deformation-instability response window, cascade consequence window, and post-rain recovery window. Each type of response time window is set staggered along a unified time axis according to its corresponding physical response stage. The flash flood rapid response window has a shorter response delay than the slope hydrological response window. The slope hydrological response, micro-vibration-mechanical response, and deformation-instability response can be sequentially delayed. The cascade consequence window is determined based on direct parent events such as landslides entering gullies or blockages. The post-rain recovery window remains active after rainfall ends until the slope condition meets the recovery criteria. The physical start delay and end delay of the response are denoted as follows: and The temporal matching of candidate events uses the intersection criterion between the predicted response interval of the parent event and the interval of the observed child event: (10) In equation (10), This indicates the predicted response range given by the parent event; Indicates the interval of the sub-observed event; and Represent the time and uncertainty of the parent event, respectively; and These represent the time and uncertainty of the sub-observation event, respectively. Indicates slope position, depth, channel cross-section, or river section location; Indicates the monitoring parameters; This represents a condition vector consisting of slope position, depth, soil layer or lithology, fissures, season, previous wetting, and current state. This indicates the time matching result. The physical response delay, the time uncertainty interval of parent and child events, and the forecast data validity period are determined separately. The forecast validity period does not replace the physical response delay. Figure 5 Each response time window in the diagram represents the conditional time range of the corresponding physical process. When determining the boundary between specific parent and child events, the time matching is based on whether the predicted response interval of the parent event intersects with the interval of the observed child event. The order and width of the windows in the diagram do not represent a fixed global duration.

[0063] Table 2 Multi-layer response time windows and edge objects

[0064] When there are few historically confirmed events, and Initial values ​​are provided by confluence time, wetting front advance, pore pressure diffusion, sensor resolution, and engineering tests. After confirming event accumulation, the system is grouped by season, soil layer, slope position, depth, fissure, and pre-wetting, and the window is updated using one or more methods among cross-correlation, time-shift optimization, change point, or recovery state convergence. The cross-correlation candidate time delay adopts equation (11): (11) In equation (11), This represents the candidate time delay that maximizes the correlation between the driving sequence and the response sequence; Indicates the parent event or driving state sequence; This represents the time-shifted response sequence; This indicates a correlation calculation. As a window candidate, it must simultaneously pass constraints related to topological connectivity, propagation direction, physical reachability, and quality. If an excessively wide window leads to increased false associations, the system shortens the window while maintaining coverage of historical confirmed events; if pore pressure or deformation continues to deteriorate after the rain stops, Keep it open.

[0065] Step 4: Generate candidate events and construct a directed event graph.

[0066] Figure 6 The system displays candidate event blocks. It generates candidates for heavy rainfall, flash floods, slope hydrology, micro-vibration-mechanical events, deformation, cascade consequences, and anomaly causes. Candidate sources can include measured change points, threshold exceedances, physical model prediction intervals, remote sensing objects, or data-driven anomaly scores. Different sources only change candidate quality, model applicability, and uncertainty; they do not change the requirement that candidates must undergo alternative interpretation contribution verification. Derived nodes originating from the same original observation or the same model output are grouped into the same dependency group based on their original source identifier and data lineage.

[0067] Table 3. Candidate Events, Independent Evidence, and Alternative Causes

[0068] Candidate event directed graph is built on Figure 3 Based on the spatial topology of slope unit—gully—river channel—disaster-bearing body shown, and combined with Figure 6 The candidate events, direct parent nodes, and alternative cause relationships are constructed as shown. The directed event graph is notated as follows: ,in Represents a set of candidate events, states, observations, and alternative cause nodes. This indicates the set of allowed relational edges. Rainfall nodes can serve as ancestors of flash floods or slope hydrology; current aquifer, groundwater, pore water pressure, suction, toe scour, or local mechanics nodes are candidates for direct parent nodes of deformation; confirmed landslide nodes can be connected to gully entry, blockage, and breach nodes. Nodes indicating equipment failure, time synchronization errors, communication anomalies, construction, earthquakes, river recharge, remote sensing cloud shadows, permanent water bodies or other object obfuscation, and nodes outside the model's applicable domain point to observation nodes that they can interpret.

[0069] Candidate edges The compatibility condition adopts equation (12): (12) In equation (12), Indicates that by node Pointing to node The candidate edge compatibility value; Indicates spatial topological connectivity; Indicates the direction of propagation that can be marked; Indicates mass conservation, state transition, or other physical consistency markers; Represents a node The quality of evidence; Indicates the edge construction quality threshold; Indicates the applicable domain label of the model; The indicator function is represented. Candidate edges are established only when all terms in equation (12) are one; source independence is not treated as a unilateral attribute, but is processed in the dependency grouping and independent evidence index calculation stages.

[0070] Step 5: Fix the target's subsequent observation set and search for a benchmark interpretation structure containing candidates to be verified.

[0071] The system first filters infeasible edges according to equation (12). For the candidate edges to be verified... Fixed target subsequent observation set before deletion The set must not contain The model outputs, alarm scores, or interface labels in the derived dependency set are primarily composed of independent measured or independent remote sensing observations. The system searches the complete directed event graph for elements containing these elements. And coverage A low-cost benchmark interpretation structure consisting of one or more directed paths The soft cost of the structure is explained using equation (13): (13) In equation (13), Representing the interpretation structure The cost of explanation; This represents the normalized time deviation between the observed event interval and the predicted response interval; This represents the cost of propagation distance, location residual, or spatial scale inconsistency under the premise of topological connectivity. This represents the prior deviation between multiple allowed propagation mechanisms. If the propagation direction only adopts a binary rule of allowed or disallowed, this term can be omitted. This represents the aggregate cost of node and edge evidence quality; Indicates the distance of the applicable domain of the model; The normalized aggregate represents the prediction interval, parameters, and observation uncertainties; each soft cost is normalized to zero to one. , , , , and For each non-negative weight, the sum of the weights is one. Inadmissible temporal conflicts, physical conservation conflicts, topological disconnections, or propagation direction restrictions are treated as hard constraints and the corresponding edges are directly deleted without compensation by finite soft cost; interpretation structures of different lengths are compared after normalization based on the coverage observation weights and structure length.

[0072] Table 4 Path cost terms and low cost conditions

[0073] Benchmark Explanation Structure Once determined, the system records its interpretation cost. Comprehensive uncertainty Subsequent observation sets for a fixed target benchmark coverage And the flash flood, landslide, or cascading conditions supported by this explanatory structure. Only when Candidate contribution confirmation only begins when the preset positive baseline coverage threshold is reached; the baseline interpretation structure itself does not equate to the candidate being confirmed.

[0074] Step 6: Strictly remove candidate and derived dependency sets, and re-search for alternative interpretation structures.

[0075] like Figure 7 As shown, the system constructs candidate data for verification based on the original source identifier, parent event identifier, derivation relationship, and model call record. Derived dependency set The set is calculated based on data lineage closure, including only those derived from... The generated intermediate hydrological or mechanical states, model rolling outputs, source-derived features, alarm scores, and interface prompts; independent measured observations, other cause nodes, and independent data. The model state does not enter Subsequent observation set of a fixed target Determined before deletion and must not belong to .

[0076] The search rule for deleting directed event graphs and feasible alternative explanation structures adopts equation (14): (14) In equation (14), Represents a complete directed event graph; This indicates strict removal of candidates to be verified. and its derived dependency set The subsequent deletion of the directed event graph; This represents the set of subsequent observations of a fixed target with the same interpretation structure as the baseline. This indicates the removal of overlays in a directed event graph. A set of feasible alternative explanatory structures; only if The lowest-cost alternative interpretation structure is defined only when it is not empty. Core contribution verification employs strict removal or disabling, and does not replace counterfactual deletion with "de-weighting"; de-weighting can only be used for candidate ranking or sensitivity analysis.

[0077] when When the system fails, it does not compute the optimal solution on an empty set. Instead, it first checks whether key observations are missing or not present, whether communication is interrupted, whether time synchronization is available, and whether the model's applicable domain is exceeded. If any of these checks fail, the candidate enters a data insufficiency state. Only when all checks pass, the baseline coverage reaches the threshold, and independent support from different devices or mechanisms is retained, is a "no feasible alternative" flag set. This serves as evidence of the candidate's irreplaceable contribution, without artificially setting alternative interpretation costs or creating increments through uncertainty upper bounds. As an optional implementation, the system can run a two-branch replay with both candidate retention and candidate deletion effects under the same initial state, boundary conditions, and external inputs, using the differences in subsequent predictions of water level, pore water pressure, displacement, blockage status, or flood peak as auxiliary verification quantities; the two-branch replay does not replace independent measured evidence, strict deletion, alternative interpretation re-search, and data integrity checks.

[0078] Step 7: Calculate the contribution gating value and update the candidate lifecycle.

[0079] The observation coverage, overall uncertainty, key data integrity, and contributions before and after deletion for the baseline interpretation structure and feasible alternative interpretation structures are calculated using Equation (15): (15) In equation (15), This represents the versioned non-negative weights of subsequent observations of the target; Representing the interpretation structure For observation The combined uncertainty of normalized time, observations, models, and parameters; Represents the key data set; Indicates the weight of key data; This indicates whether key data arrived effectively within the corresponding time window; This indicates the data quality and takes a value between zero and one. Indicates the integrity of critical data; , and These represent the cost, uncertainty, and coverage of the benchmark interpretation structure, respectively. , and These represent the corresponding quantities of feasible alternative explanation structures; , and These represent the incremental cost of explanation, the incremental uncertainty, and the loss of observation coverage, respectively.

[0080] The Independent Evidence Index is formed by deduplicating evidence from source-dependent groups: (16) In equation (16), Indicates candidate The index of independent evidence support; This indicates that the candidate data is supported after deduplication based on original source, data lineage, and known common-mode correlation. A collection of independent equipment groups or physical mechanism groups; Indicates a separate group of evidence; This indicates the support level of the independent evidence group after combining the evidence quality, and its value ranges from zero to one. Multiple indicators from the same original sensor, the same model output, or known to have common mode error are first merged into the same evidence group, and then Equation (16) is executed; It is a support index and is not interpreted as the probability of an event occurring.

[0081] Candidate confirmation gating adopts formula (17): (17) In equation (17), Indicates the candidate confirmation gating result; , , , , and These represent baseline coverage, data integrity, independent evidence, incremental interpretation costs, incremental uncertainty, and coverage loss threshold, respectively. Indicates that there is no feasible alternative; Indicates an indicator function. When When a candidate is confirmed, it is moved to the confirmed state; when a low-cost, low-uncertainty, and sufficiently comprehensive alternative explanation exists, the candidate is rejected or its status level is downgraded; when key data is unavailable, time synchronization is unavailable, the model's applicable domain is exceeded, or the evidence window is not closed, the candidate is placed in the data insufficiency or pending review state. Equations (14) to (17) are re-executed after new evidence arrives. Each threshold is determined primarily using the separation results of historical confirmed events and confounding events at the site, equipment calibration, and field tests, and is updated in a controlled manner according to seasonality, slope condition, and changes in the monitoring network.

[0082] Step 8: Confirm flash floods, landslides, and cascading consequences, and implement asynchronous classification.

[0083] The system will write the flash flood, slope hydro-mechanical, deformation, and cascade candidates from equation (17) into the flash flood confirmed event set, landslide confirmed event set, and cascade confirmed event set, respectively. The joint disaster state vector adopts equation (18): (18) In equation (18), Indicates the first The joint disaster state vector for each calculation cycle; Indicates the state of flash floods; Indicates the state of a landslide; Indicates a landslide blockage-breakdown cascade; superscript This represents the vector transpose. The three states are calculated from the confirmed event, the current disaster stage, and the predicted arrival or instability interval, respectively; the joint early warning state is calculated from... The confirmed riverbank coupling relationship, the estimated remaining disposal time, and the location of the disaster-affected objects are mapped together.

[0084] like Figure 8 As shown, this implementation sets five internal risk levels: Normal, Attention, Alert, Severe, and Emergency. Attention indicates that the parent event of the rainstorm, prior wet conditions, or single-source candidates require intensive observation; Alert indicates that a disaster branch has gained independent support but alternative interpretations are still feasible or the evidence window has not yet closed; Severe indicates that key candidates have formed a confirmed coupling through contribution gating or flash flood and landslide branches; Emergency indicates that the flood peak or rapid instability is approaching, or a high-consequence blockage-breakdown chain has been strongly confirmed. These internal risk levels are mapped to locally applicable statutory warning signals through external interfaces, and the internal level names are not directly equated with administrative warning levels; when absolute water depth, flow velocity, displacement velocity, or on-site manual confirmation reaches the local emergency bypass conditions, personnel evacuation and containment can be implemented first, while the event map continues to undergo authenticity verification to support subsequent adjustments. Figure 8 As shown in the lower part, the flash flood state and the landslide state evolve independently along a unified time axis. The flash flood state can rise rapidly with the arrival of the flood peak and decrease relatively quickly with the receding of the flood peak, while the landslide state can continue to be maintained or rise after the rainfall ends due to infiltration, accumulation of pore water pressure and deformation development. Figure 8 The curves in the diagram are only used to represent the relative evolution trend and asynchronicity of different disaster states. They do not represent fixed time lengths, level thresholds, change rates, or functional relationships. The actual upgrades and downgrades of each state are still determined according to the corresponding confirmation events, response time windows, and recovery gating.

[0085] Table 5. Grading Status and Recommended Actions

[0086] When a landslide is confirmed to potentially enter a gully, the system generates blockage candidates based on the volume, velocity, direction of movement, distance from the gully, and proportion of the gully; it generates backwater candidates based on the valley width, dam height, material composition, flow capacity, and upstream water flow; it generates seepage and internal erosion candidates based on seepage, turbid water, fine particle migration, piping, erosion gullies, and water level changes; and it generates overtopping or breach candidates based on the dam crest freeboard, overtopping flow, and dam stability. General flash floods, floating debris blockage, and landslide dam paths are used as competing interpretations, and equations (14) to (17) are executed respectively to confirm that a landslide does not automatically confirm a breach. Figure 8The “landslide—entry into the gully / blockage—dam-front water backing—seepage erosion or overtopping—breakdown flood” shown above only represents the possible cascading candidate generation sequence and physical evolution direction; the confirmation of the previous level candidate is only one of the conditions for the generation of the next level candidate and the search for the interpretation structure. The next level candidate should still be subject to strict reduction, alternative interpretation re-search and candidate contribution gating respectively. The confirmation result of the previous level does not automatically turn the next level candidate into a confirmation event.

[0087] The landslide restoration gating adopts type (19): (19) In equation (19), Indicates the results of landslide recovery gating; It represents a restored state vector that includes at least pore water pressure, volumetric water content, displacement velocity, displacement acceleration, and stability indices calculated by the site model. This represents the recovery zone determined by the site's historical health status and testing. Indicates the duration of restoration; This indicates the event time corresponding to the current calculation period. The receding of the flood peak does not cause... Automatically select one. For example... Figure 8 As shown, even when the original rainfall and flash flood conditions have decreased, the landslide conditions can still be maintained or increased due to the continued deterioration of the slope's hydrological and mechanical conditions. Once a blockage candidate is verified and confirmed, even if the original flash flood has receded, the downstream flood conditions can still be increased again due to backwater in front of the dam, flooding, or a verified breach candidate, thus creating a secondary flood risk that is temporally separate from the original rainstorm and flash flood process.

[0088] Optimal operating mechanism one: Adjust sampling, communication, and verification resources according to risk status.

[0089] The system can be set to four operating states: normal, candidate, confirmed, and recovery. In the normal state, the summary is uploaded at a low power frequency and the device performs self-checks. In the candidate state, the sampling rate of relevant nodes is increased, the original fragments are saved, and the competitive interpretation is expanded. In the confirmed state, key nodes are uploaded at high frequency and with redundancy, the response time window, alternative interpretation structure, and estimated remaining time are continuously recalculated, and manual, UAV, or remote sensing verification is triggered. In the recovery state, the sampling frequency is gradually reduced, but pore water pressure, volumetric water content, and deformation monitoring are maintained until equation (19) is satisfied. The operating state can also adjust the model computational complexity according to the current number of candidates, the degree of interpretation competition, data integrity, and the urgency of handling. Changes in computing resources are not used as an independent basis for confirming the authenticity of candidate events.

[0090] Warning messages must include at least the location, risk level within the disaster type, joint warning status, baseline interpretation structure, main alternative interpretation structure, key independent evidence, expected arrival or instability range, data deficiencies, and recommended actions. After message release, confirmation is awaited; if no confirmation is received within the confirmation period, backup gateways, SMS, satellite, or broadcast channels are activated. Communication confirmation failure only changes the message delivery status and corresponding data quality, and does not reversely alter the authenticity of the disaster event; communication routing and disaster alternative interpretation structures are modeled separately.

[0091] Preferred Operation Mechanism Two: Establish version-based auditing, execute controlled updates, and conduct post-processing review.

[0092] Each calculation cycle saves the response time window version, condition variables, complete directed event graph, reduced directed event graph, candidate and derived dependency sets, baseline interpretation structure, alternative interpretation structure, gating variables of equations (15) to (17), disaster type status, joint early warning status, release actions, and execution receipts. Chained audit records can use equation (20): (20) In equation (20), Indicates the first Audit summary for each calculation period; Represents a hash or equivalent irreversible digest function; This indicates a summary of the audit for the previous period; Indicates the currently input evidence; This represents the current baseline interpretation structure, the alternative interpretation structure, and the set of gated results; Indicates a set of actions to be published or processed; This represents the model, response time window, rules, and data version set. When implementing the audit summary, the input fields are serialized according to a fixed field order, unified character encoding, and version number before the summary is calculated. Audits are used for retrospective analysis and version consistency, and do not replace confirmation of disaster authenticity.

[0093] Static susceptibility, topographic humidity index, lithology, slope, distance from river network, and historical disaster inventory are only used for path priors, monitoring deployment, and inspection priorities; remote sensing segmentation probability, reconstruction error, map score, and machine learning output are only used for candidate generation, interpretive structure ranking, and uncertainty estimation. Only confirmed events that are confirmed through Equation (17), are time-traceable, and have a clear cause through on-site or post-hoc observation are allowed to update the response time window, interpretive cost weight, or dynamic threshold. Each update saves the versions before and after the update and the rollback point, and uses independent time period playback, subsequent observation, or subsequent confirmed events to check the response time window hit, misassociations, and interpretive stability; if the review fails, the version before the update is restored. Mixed events only update the alternative cause library; online updates are suspended when the model's applicable domain is exceeded or the data quality is insufficient; the execution feedback of the handling action is written into the event record and triggers state recalculation, but disaster events must not be confirmed independently without independent measured evidence.

[0094] It is understood that the flash flood and landslide disaster early warning method of this invention can be executed periodically by a watershed edge gateway, a county-level early warning platform, a cloud service, or a processor connected to it. The program executes eight core stages sequentially: multi-source event recording and unified time, rainstorm parent event and dual-branch diversion and state calculation, conditional response time window, candidate event and directed event graph, baseline interpretation structure, strict removal and alternative interpretation re-search, contribution gating and candidate lifecycle, asynchronous state and cascading for different disaster types; adaptive sampling communication, versioned auditing, controlled updates, and post-processing review serve as the preferred operating layers.

[0095] Table 6 Main Loop Program Flow

[0096] Example 1: Full-chain example of torrential rain—flash flood—delayed landslide—river damming—secondary flood This embodiment uses an executable calculation example to illustrate the complete closed loop, and the following numerical values ​​are not presented as specific field measurement conclusions. Slope units are set up in the target small watershed. The main channel connected to it and upstream and downstream river sections and Disaster-bearing bodies are set up downstream. Rain gauges, radar rainfall, water level / flow rate in two river sections, multi-depth volumetric water content and pore water pressure in slope units, soil pressure at the toe of the slope, displacement at the rear edge, and video footage from the gully mouth form independent or traceable sources of observation. For example, the platform receiving time for a pore water pressure record is 10:20:08.6 seconds. The calibrated sensor response delay, edge processing delay, and communication delay are 1.2 seconds, 2.4 seconds, and 5.0 seconds, respectively. Using the receiving time backtracking mode, the event time is obtained as 10:20:00.0 seconds. The corresponding upper bounds for receiving clock, sensing, processing, and communication errors are 0.4 seconds, 0.3 seconds, 0.3 seconds, and 0.4 seconds, respectively. Conservatively summing these results in an upper bound of 1.4 seconds for the event time uncertainty.

[0097] The system identified a rainstorm parent event from 09:50 to 11:20, with a high level of humidity in the early stage. Taking one of the unified control areas and calculation periods as an example, the equivalent amount of rainfall at the event scale is 28.0 mm, the fissure flow rate is 4.2 mm, the matrix infiltration is 7.8 mm, the interception and other losses are 1.5 mm, the equivalent amount of outflow converted from the independent channel flow integral is 13.7 mm, the water storage increases by 0.6 mm, and the water consistency residual is 0.2 mm according to Equation (5), which does not exceed the tolerance of 0.8 mm determined based on the propagation of errors in each measurement; the equivalent amount of excess runoff given by the runoff generation and confluence model is also 13.7 mm and is used as the input of the flash flood response branch. To demonstrate the timescale of the complete event, this embodiment provides scene envelopes for the rainstorm parent event, ranging from 18 to 72 minutes for flash floods, 1.5 to 8 hours for slope hydrology, 4 to 11 hours for micro-vibration-mechanics, and 5 to 16 hours for deformation-instability. The actual candidate edge construction still uses the intersection criterion of the predicted response interval of the direct parent event and the sub-observation event interval according to Equation (10), and the forecast validity period is not used as the physical response delay.

[0098] In the directed event graph, the rainstorm parent event serves as the ancestor node; excess runoff, upstream river level rise, and downstream peak flow constitute flash flood tributaries; matrix infiltration and fissure flow, deep pore water pressure, slope toe earth pressure, and trailing edge displacement constitute slope hydro-mechanical tributaries; unobserved tributary inflows, river backwater, equipment drift, time synchronization anomalies, construction vibration, and remote sensing confounding serve as competing explanatory nodes. Water level rises, slopes, and abrupt change indices derived from the same water level gauge are grouped into the same data lineage dependency group and are not repeated as independent evidence. Thirty-one minutes after the start of the rainstorm, upstream water level, downstream flow, and gully mouth video enter the corresponding time matching conditions, forming flash flood candidates to be verified.

[0099] Equations (13) to (17) are executed for the candidate flash floods. In this embodiment, the weights of the six items of explanation cost—time, space, permissible mechanism, evidence quality, model applicability domain, and uncertainty—are taken as 0.25, 0.15, 0.10, 0.20, 0.10, and 0.20, respectively; the weights of the six normalized soft costs of the benchmark explanation structure are taken as 0.12, 0.15, 0.10, 0.20, 0.15, and 0.31, respectively, and the weighted average is obtained. After strictly removing candidate flash flood warnings and their derived warning scores, water level rises, and abrupt change characteristics, the best alternative explanatory structure is "unobserved tributary inflow—river channel water level / discharge change." Six soft cost examples are taken as 0.60, 0.55, 0.50, 0.65, 0.60, and 0.84 respectively, and weighted to obtain... The combined uncertainties of the benchmark and the alternative are 0.12 and 0.31, respectively, and the coverage is 0.92 and 0.55, respectively. Therefore, we obtain... , and The support scores for the three independent evidence groups—upstream water level, downstream flow rate, and ditch mouth video—are taken as 0.55, 0.50, and 0.38, respectively, and obtained according to formula (16). The three key observations are equally weighted and have valid arrivals, with quality examples of 0.96, 0.94, and 0.92, obtained according to equation (15). In this embodiment, the data integrity threshold is 0.85, and the other contribution thresholds are 0.30, 0.15, 0.25 and 0.70 respectively, all of which satisfy equation (17), and the flash flood candidate is converted into a confirmed event.

[0100] After the torrential rain subsided, the flash flood peak gradually receded, but the pore water pressure in the deep slope unit continued to rise, leading to accelerated displacement at the trailing edge. During the initial strict removal and alternative interpretation re-search of candidates for accelerated displacement, the alternative interpretation set was temporarily empty; the system did not set accordingly. Instead, it first calculates the integrity of key data. It finds that an auxiliary displacement node failed to upload data due to communication congestion, and the data integrity did not meet the threshold. Therefore, the candidate is marked as having insufficient data and is monitored at high frequency. After the backup channel is restored, the missing displacement records and timestamps arrive. After event time correction, steps five to seven are re-executed. This process illustrates that only after excluding missing or absent key observations, communication anomalies, time synchronization anomalies, and situations outside the model's applicable domain can infeasible alternative states be considered as evidence of candidate infeasible contributions.

[0101] After data completion, deep pore water pressure and slope toe earth pressure were selected as direct parent node candidates for the current deformation, and the rainstorm parent event was selected as the multi-step ancestor. The interpretation cost, comprehensive uncertainty, and coverage of the baseline interpretation structure were 0.22, 0.14, and 0.90, respectively. After strictly removing displacement acceleration candidates and their derived alarms, the optimal combination cost, uncertainty, and coverage of competing interpretations such as river recharge and construction vibration were 0.70, 0.36, and 0.48, respectively. , and The support values ​​for the three independent groups of multi-depth pore water pressure, satellite positioning displacement, and visual displacement are taken as 0.52, 0.48, and 0.36, respectively, and obtained according to equation (16). The critical data integrity score is 0.93, which meets the threshold of this embodiment, and the landslide candidate event is converted into a confirmed event. At this time, the flash flood status has decreased according to the receding flood peak, while the landslide status has increased, and the joint early warning status is dominated by landslides; the receding flash flood does not automatically trigger the landslide cancellation.

[0102] The landslide's movement direction was confirmed to be towards the main channel. Video footage and radar topographic imaging at the channel mouth showed the landslide entering the channel. Subsequently, the water level in front of the dam rose while the water level downstream of the dam briefly dropped, generating a blockage candidate. The baseline interpretation including the blockage candidate had a structural cost, uncertainty, and coverage of 0.20, 0.11, and 0.94, respectively. After strictly removing the blockage candidates and their derived states, the best competing interpretation was floating debris blockage, with a cost, uncertainty, and coverage of 0.67, 0.33, and 0.50, respectively. , and The support examples for the three independent groups—the sliding body boundary, the water level in front of / below the dam, and the ditch video—are taken as 0.60, 0.55, and 0.33, respectively, and obtained according to equation (16). The critical data integrity threshold was met, therefore the blockage candidate passed the contribution verification. It is confirmed that the landslide itself does not automatically confirm blockage; only the independent verification result of this segment is used to update the cascade status.

[0103] After blockage is confirmed, the system continues to generate candidates for upstream backwater, seepage erosion, overtopping, and breach. For each cascaded candidate, it re-fixes the corresponding target's subsequent observation set, strictly removes candidates and their derived dependencies, re-searches alternative interpretation structures, and performs contribution gating. The cascaded state only escalates when overtopping flow, dam erosion, downstream flood peak propagation, or other independent evidence meets the corresponding time window, topology, and contribution conditions. If there is only an increase in upstream water level without evidence of dam geometry, erosion, or downstream propagation, it remains on alert or has insufficient data and does not automatically confirm breach. If a validated breach candidate is established, the downstream flood state can still escalate again due to the cascade consequences, even if the original rainfall and flash flood have receded. When configuring rolling prediction or digital twin models, a two-branch approach with candidate retention and candidate deletion effects can be run under the same initial state, boundary conditions, and external inputs. The prediction difference is used as an auxiliary verification measure, but this result does not replace independent measured evidence, strict removal, alternative interpretation re-search, and data integrity checks.

[0104] Throughout the entire process described above, the system can sequentially enter normal, candidate, confirmed, and recovery operation states, adjusting sampling frequency, transmission strategy, model computing resources, and the frequency, coverage, or resolution of manual, drone, or remote sensing verification according to the risk status. Early warning messages record location, internal risk level of flash floods / landslides / cascading disaster types, joint early warning status, baseline interpretation structure, main alternative interpretation structure, key independent evidence, and expected arrival or instability intervals. After release, it awaits confirmation receipt; if no receipt is received within the set time limit, a backup communication channel is activated; backup communication only changes the message delivery method and does not serve as a disaster alternative interpretation structure. Each calculation cycle saves the event time, response time window version, complete and pruned graphs, candidate and derived dependency sets, baseline and alternative interpretation structures, contribution gating results, disaster type status, release action, and receipt.

[0105] Example 2: Flash Floods Triggered by Short-Duration Heavy Rainfall This embodiment illustrates a rapid branch road scenario characterized by "flooding without landslides." A small watershed experiences short-duration, high-intensity rainfall, with excess runoff dominating. The initial volumetric water content of the slope is low, and deep pore water pressure has not yet reached the slope's hydrological response conditions. The envelope of the flash flood scenario relative to the parent event of the rainstorm is 20 to 70 minutes. Slope hydrology and deformation windows are continuously calculated, but actual edge construction follows the interval matching criteria of the direct parent event. Water levels / flow rates from upstream and downstream equipment and gully mouth videos form flash flood candidates; higher static slope, topographic humidity index, or susceptibility only improves inspection and path priors, and does not generate separate landslide confirmations.

[0106] The baseline explanatory structure cost, uncertainty, and coverage examples for flash flood candidates are 0.20, 0.13, and 0.91, respectively. After strictly removing candidate and source-derived indices, the corresponding values ​​for unobserved tributaries joining the alternative explanatory structure are 0.62, 0.30, and 0.56, respectively. An example of the independent evidence index after source deduplication is shown below. Examples of critical data integrity are as follows If the difference before and after deletion and the threshold value reach the site-defined threshold, the flash flood candidate is confirmed; if the dynamic evidence of deep pore water pressure, displacement and micro-vibration during the same period is insufficient, the landslide candidate is not confirmed. After the flood peak recedes, the flash flood state decreases independently, while the landslide state remains normal or of concern.

[0107] Example 3: Delayed landslide events caused by post-rain pore pressure peaks and displacement acceleration This embodiment illustrates the delayed deterioration and independent recovery gating of slope after rain. When the heavy rain ends, the flash flood peak has receded, but the initial moisture level remains high. Deep pore water pressure in the middle of the slope reaches its peak 15 hours after the rain stops. Pore water pressure and soil pressure at the slope toe change sequentially, and the displacement velocity at the rear edge begins to increase continuously 18 hours after the rain stops. The flash flood rapid response window is closed, while the slope hydrological window, micro-vibration-deformation window, and post-rain recovery window remain active. The system uses deep pore water pressure as a candidate direct parent node for the current deformation and the rainstorm parent event as a multi-step ancestor, prohibiting rainstorm nodes from directly proving displacement by bypassing observable slope hydrological or mechanical direct parent nodes.

[0108] Equations (14) to (17) were executed for the displacement acceleration candidates. The cost, uncertainty, and coverage of the baseline interpretation structure were 0.21, 0.13, and 0.91, respectively; after strictly removing the candidates and their derived alarms, the optimal combination of competing interpretations such as equipment drift, river recharge, and construction vibration corresponded to 0.61, 0.34, and 0.52, respectively; calculated using independent evidence sets formed by multi-depth pore water pressure, satellite positioning displacement, and visual displacement. Examples of critical data integrity are as follows Therefore, the candidate passed the contribution gating. At this time, the flash flood state can be normal, while the landslide state is severe; only when the pore water pressure and volumetric water content enter the recovery range, the displacement velocity and acceleration no longer deteriorate, the stability index enters the recovery domain, and the example holding time is continuously met for 12 hours, is the landslide allowed to be downgraded step by step. The 12 hours mentioned is a site-specific example and not a universal fixed value.

[0109] Example 4: Rejection of Slope Hydrological Candidates Due to Competition for Interpretation of Equipment Drift This embodiment illustrates the negative falsification capability of alternative interpretation verification. During a period with no significant rainfall and stable pore water pressure, volumetric water content, groundwater level, and displacement at adjacent depths, a pore water pressure channel exhibits slow upward drift. The zero-point baseline, self-test records, most recent calibration results, and consistency with neighboring channels indicate that the equipment drift has an interpretable basis. Therefore, the system simultaneously generates candidate nodes for slope hydrology and competing causes of equipment drift. The rate of change, slope, and model alarm score derived from this channel are all classified into the same data lineage dependency group and are not repeatedly used as independent evidence.

[0110] The baseline interpretation structure with slope hydrological candidates in the complete directed event graph has a cost of 0.24, an overall uncertainty of 0.12, and a coverage of 0.90. After strictly removing candidates and their derived rates of change and alarm scores, the cost of the alternative interpretation structure for device drift is 0.28, the overall uncertainty is 0.15, and the coverage is 0.882. Therefore, we obtain... , and Independent Evidence Index: None of the submitted candidates met the confirmation threshold set by the site, and they were rejected and transferred to equipment quality handling. The system reduced the evidence quality of that channel and generated a calibration or maintenance task, restricting that channel from independently supporting severe or urgent internal risk levels until calibration is completed.

[0111] Understandably, comparative verification methods can employ historical event replay, controlled experiments, or clearly identified simulation / computation data, inputting them sequentially in chronological order, without using future observations in current candidate confirmation. At least the following comparisons should be set: using only reception time versus dual-mode event time correction, single rainfall branch versus state-related dual branch, fixed time window versus parent-child event interval matching, repeated weighting of homogeneous features versus data lineage cleanup, direct confirmation of candidate scores versus strict removal—gating of alternative explanation contributions, synchronous desynchronization versus asynchronous recovery by disaster type; and key observation missing—supplementation comparison and no feasible alternative cause comparison should be set to check whether insufficient data or insufficient model applicability domain is misjudged as an actual event.

[0112] Evaluation metrics may include the number of events with incorrect time sequence, the number of water volume consistency conflicts, response time window hit rate, false association rate, disaster type-specific confirmation accuracy and recall rate, false confirmation rate, alternative explanation separation rate, data insufficiency correct identification rate, no feasible alternative cause identification rate, event confirmation lead time, premature resolution rate, blockage-breakdown cascade identification status, single-cycle recalculation time, and message receipt rate. When no real-world field verification results are available, only the data injection method, calculation process, and evaluation method are provided; no pre-set accuracy or performance improvement percentage is given, and simulation or calculation examples are not presented as actual field test results.

[0113] In addition, such as Figure 9As shown, an embodiment of the present invention also discloses an electronic device, including: at least one processor 210; at least one memory 220 for storing at least one program; when the at least one program is executed by the at least one processor 210, it implements the flash flood and landslide disaster early warning method as in any of the preceding embodiments.

[0114] In addition, one embodiment of the present invention discloses a computer-readable storage medium storing computer-executable instructions for performing the flash flood and landslide disaster early warning method as described in any of the preceding embodiments.

[0115] The system architecture and application scenarios described in the embodiments of this invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of this invention, and do not constitute a limitation on the technical solutions provided by the embodiments of this invention. As those skilled in the art will know, with the evolution of system architecture and the emergence of new application scenarios, the technical solutions provided by the embodiments of this invention are also applicable to similar technical problems.

[0116] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0117] In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0118] The terms “component,” “module,” “system,” etc., used in this specification are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process or execution thread, and components may be located on a single computer or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, or a network, such as the Internet interacting with other systems via signals).

Claims

1. A method for early warning of flash floods and landslides, comprising: Acquire multi-source observation data of the target area, and perform event time correction, source tracing and data lineage marking on the multi-source observation data to generate event records with event time, event time uncertainty interval, original source identifier, dependency group identifier, data lineage and evidence quality; Based on rainfall input, previous wet conditions, state-related infiltration capacity, and fissure conditions, rainfall is allocated to flash flood response branches and slope hydro-mechanical response branches to generate flash flood candidate events and slope response candidate events, respectively. Based on the parent event, location, monitoring parameters, and current state, a conditional multi-layer response time window is established, and based on the conditional multi-layer response time window and spatial topology, a directed event interpretation graph containing candidate event nodes, observation nodes, derived dependency nodes, and alternative cause nodes is constructed. For a candidate event to be verified, fix the subsequent observation set of the target and search the complete directed event interpretation graph for a benchmark interpretation structure that contains the candidate event to be verified and can cover the subsequent observation set of the target. Perform a counterfactual deletion operation to remove the candidate event to be verified and the data lineage closure derived only from it from the directed event interpretation graph, forming a pruned directed event graph, and then search for alternative interpretation structures in the pruned directed event graph for subsequent observation sets of the same target. Based on the existence of alternative interpretation structures, the increment of interpretation cost, the increment of uncertainty, the loss of observation coverage, the independent evidence index, and the data integrity, the non-substitutable contribution state of the candidate event to be verified is determined, and the candidate event to be verified is confirmed or rejected based on the non-substitutable contribution state. Based on confirmed flash flood events, landslide events, and cascading events, the status of flash floods, landslides, and cascading events are updated respectively, and asynchronous graded early warnings are implemented according to disaster type.

2. The method according to claim 1, characterized in that, The event time correction of the multi-source observation data includes: When the source clock quality meets the preset conditions, the trusted source time correction mode is adopted. The event time is obtained by mapping the source device clock to a unified clock and subtracting the sensing delay. When the source clock quality does not meet the preset conditions, the receiving time backtracking mode is adopted. The event time is obtained by subtracting the sensing delay, processing delay and communication delay from the platform receiving time in sequence. The event time uncertainty interval is determined by the upper bound of the non-negative uncertainty of the clock error, sensing error, processing error, and communication error involved in the time estimation.

3. The method according to claim 1, characterized in that, The allocation of rainfall to flash flood response branches and slope hydro-mechanical response branches includes: The distribution ratio of rapid flow in the fractures is determined based on fracture aperture, connectivity, surface cover, and pre-wetting conditions. Based on the fissure rapid flow allocation ratio, the distributable rainfall after deducting prior losses is allocated as fissure rapid flow and matrix infiltration; wherein, the excess runoff is used as the input of the flash flood response branch, and the matrix infiltration and fissure rapid flow are used as the input of the slope hydro-mechanical response branch; when there are independent channel outflow observations, the event-scale water consistency residual is calculated and used as evidence of data conflict.

4. The method according to claim 1, characterized in that, The conditional multi-layered response time window includes a rainstorm and pre-rain condition window, a flash flood rapid response window, a slope hydrological response window, a micro-vibration mechanical response window, a deformation and instability response window, a cascade consequence window, and a post-rain recovery window. Among them, time matching is determined when the predicted response interval of the parent event and the event time interval of the child observed event intersect. The physical response delay and the uncertain event time interval of the parent and child events are represented separately and updated in groups according to the confirmed events in the site history.

5. The method according to claim 1, characterized in that, The construction of the directed event interpretation graph, which includes candidate event nodes, observation nodes, derived dependency nodes, and alternative cause nodes, includes: Generate candidate event nodes in the spatial topology of slope unit-valley-river channel-disaster-bearing body; Based on the conditions of time matching, spatial connectivity, propagation direction permission, physical consistency satisfaction, evidence quality reaching the edge construction threshold, and being within the model's applicable domain, directed edges are established between candidate event nodes. Derived nodes originating from the same original observation or the same model output are grouped into the same dependency group based on their original source identifier and data lineage.

6. The method according to claim 1, characterized in that, Both the baseline interpretation structure and the alternative interpretation structure consist of one or more directed paths that can cover the subsequent observation set of a fixed target. The interpretation cost of the interpretation structure is a weighted sum of time interval bias cost, spatial propagation bias cost, propagation mechanism prior bias cost, evidence quality cost, model applicability domain distance cost, and normalized uncertainty cost. Among them, topological disconnection, propagation direction disallowance, and unacceptable temporal conflicts or physical conservation conflicts are hard constraints that directly make the corresponding directed edges unusable.

7. The method according to claim 1, characterized in that, Determining the non-substitutable contribution state of the candidate event to be verified includes: Calculate the observation coverage, overall uncertainty, and key data integrity of the baseline and alternative interpretation structures; Calculate the increment of interpretation cost, the increment of uncertainty, and the loss of observation coverage before and after deleting the candidate event to be verified; The independent evidence index is calculated based on the independent evidence groups after deduplication by source and common mode correlation. When the baseline observation coverage and key data integrity reach the corresponding thresholds, and the incremental cost of no feasible alternatives or explanations, the incremental uncertainty and the loss of observation coverage all reach the corresponding thresholds, and the independent evidence index reaches the corresponding threshold, the candidate event to be verified is confirmed. When a low-cost, low-uncertainty, and sufficiently comprehensive alternative explanation structure exists, the state level of the candidate event to be verified is rejected or reduced. When key data is incomplete, time synchronization is unavailable, the model's applicable domain is exceeded, or the evidence window is not closed, the candidate event to be verified will be set to a state of insufficient data or pending review.

8. The method according to claim 1, characterized in that, The implementation of asynchronous, tiered early warning based on disaster type includes: The status of flash floods is upgraded or downgraded independently based on the propagation and receding of the flood peak; The landslide status is independently downgraded after the slope hydro-mechanical and deformation status enters the preset recovery domain and remains there for a preset time; however, the receding of flash flood peaks does not automatically trigger the landslide status to be lifted, the confirmation of a landslide event does not automatically confirm a blockage event, and the confirmation of a blockage event does not automatically confirm a breach event.

9. The method according to claim 1, characterized in that, Also includes: Only confirmed events with clear causes, traceable time, and verified through candidate event contribution are subject to controlled updates of response time windows, explanatory cost weights, or dynamic thresholds. Before updating, save the current parameter version and rollback point. After updating, use independent time period playback, subsequent observation or subsequent confirmation events for verification. If the verification result does not meet the preset conditions, restore the version before the update. After the action is taken, the execution feedback is recorded and the affected hydrological-slope state, response time window, directed event graph and disaster type level are recalculated. The execution feedback is not separated from independent measured evidence and is used as the sole basis for confirming the authenticity of the candidate event.

10. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the flash flood and landslide disaster early warning method as described in any one of claims 1 to 9.