Spatio-Temporal Knowledge Graph-Based Research on Agro-Meteorological Disaster Monitoring

: Currently, there is a wealth of data and expert knowledge available on monitoring agro-meteorological disasters. However, there is still a lack of technical means to organically integrate and analyze heterogeneous data sources in a collaborative manner. This paper proposes a method for monitoring agro-meteorological disasters based on a spatio-temporal knowledge graph. It employs a semantic ontology framework to achieve the organic fusion of multi-source heterogeneous data, including remote sensing data, meteorological data, farmland data, crop information, etc. And it formalizes expert knowledge and computational models into knowledge inference rules, thereby enabling monitoring, early warning, and disaster analysis of agricultural crops within the observed area. The experimental area for this research is the wheat planting region in three counties in Henan Province. The method is tested using simulation monitoring, early warning, and impact calculation of the past two occurrences of dry hot wind disasters. The experimental results demonstrate that the proposed method can provide more speciﬁc and accurate warning information and post-disaster analysis results compared to raw records. The statistical results of NDVI decline also validate the correlation between the severity of wheat damage caused by dry hot winds and the intensity and duration of their occurrences. Regarding remote sensing data, this paper proposes a method that directly incorporates remote sensing data into spatio-temporal knowledge inference calculations. By integrating remote sensing data into the regular monitoring process, the advantages of remote sensing data granted by continuous observation are utilized. This approach represents a beneﬁcial attempt to organically integrate remote sensing and meteorological data for monitoring, early warning, and evaluation analysis of agro-meteorological


Introduction
Agro-meteorological disasters refer to unfavorable weather conditions or abnormal climate events that significantly reduce crop yields or even stop crop growth during the agricultural growing process [1].These disasters have a severe impact on agricultural production and the economy.With global climate change, agro-meteorological disasters are showing an increasing trend.Therefore, conducting research on agro-meteorological disaster monitoring is of great practical significance in reducing agricultural losses caused by meteorological disasters.For the monitoring task, on one hand, it is important to timely and accurately transmit information on potential disastrous weather to relevant personnel for early warning purposes and provide them with emergency defense measures.On the other hand, it is necessary to monitor the affected area, severity, and other aspects after the disaster.
graphs have been developed in the field of agriculture.Qi et al. [11] proposed a method for constructing a Chinese meteorology and agriculture knowledge graph based on semistructured data.Liu et al. [12] described the application of crop disease and pest knowledge graphs in expert systems, search engines, and knowledge-based question-answering systems.Chen et al. [13] summarized the applications of multimodal knowledge graphs in agriculture, focusing on intelligent question answering, disease and pest recognition, and agricultural product recommendation research.Wang et al. [14] proposed a knowledge graph construction framework for the entire sweet cherry industry chain and utilized knowledge graph fusion and mining of relevant data to provide knowledge services for the development of the sweet cherry industry.Some studies have also applied knowledge graphs in the field of disaster research.Wang et al. [15] analyzed 2890 Chinese literature resources on disaster risk research in the Chinese Knowledge Resource Integrated Database from 2000 to 2017, constructing a knowledge structure graph of disaster risk research, including hotspots, co-occurrence matrices of keywords, core authors, and research institutions.Ge et al. [16,17] utilized spatio-temporal knowledge graphs for predicting natural disasters such as wildfires and landslides, achieving high accuracy.Chen et al. [18] improved the prediction of landslide disasters in areas with scarce sample data using spatio-temporal semantic reasoning.Spatio-temporal knowledge graphs have inherent advantages in organizing and managing heterogeneous spatio-temporal data, making them well suited to address the challenges faced in agro-meteorological disaster monitoring and early warning.However, no relevant research has been found in this specific area.
This paper introduces a semantic ontology framework and constructs a spatio-temporal knowledge graph for agro-meteorological disaster monitoring.It aims to organize and manage heterogeneous spatio-temporal data sources such as farmland data, crop information, meteorological data, remote sensing data, and so on.The system converts expert knowledge and calculation models into inference rules and utilizes the spatio-temporal knowledge graph to enable pre-disaster monitoring and warning as well as post-disaster impact analysis driven by data and knowledge collaboration.
The academic contributions of this paper include the following: (1) This paper constructs an agro-meteorological disaster monitoring spatio-temporal knowledge graph, facilitating the fusion of multi-source data and knowledge for pre-disaster monitoring and warning, as well as post-disaster impact analysis.(2) Compared to coarse-grained monitoring at the provincial or county levels, this paper achieves finer-scale monitoring at the level of farmland patches using remote sensing techniques, with warning information sent to specific farmland managers.
(3) This paper realizes the integration of remote sensing data into knowledge inference and calculation processes.

Study Area
Henan Province is a major agricultural province in China and the largest wheatproducing region in the country.Dry hot winds, as a typical meteorological disaster, are one of the major agro-meteorological hazards in the northern wheat-growing areas of China.Dry hot winds can be categorized into three types: the high-temperature and lowhumidity type, the withering type after rain, and the drought type [19].The meteorological indicators of dry hot winds vary in different regions.The meteorological indicators of the high-temperature and low-humidity type of dry hot winds in the northern winter wheat-growing areas are shown in Table 1 [3].
In this study, two instances of dry hot wind disasters that occurred in Yanjin County, Qi County (under the jurisdiction of Hebi City), and Wuyang County in May 2013 and May 2019 were selected for simulation calculations based on the collected dry hot wind warning records.The relevant warning records are shown in Table 2.

Henan Province
From 11 May to 13, 2013, Henan Province issued a red warning signal for three consecutive days due to the occurrence of dry hot wind [20].

Hebi City
On 22 May 2019, at 8:05 am, an orange warning signal for dry hot wind was issued, and it was expected that dry hot wind would occur in the urban area on the same day [21].

Yanjin County
On 22 May 2019, at 20:13, an orange warning signal for dry hot wind was issued again, and it was expected that dry hot wind would occur the next day [22].

Wuyang County
On 22 May 2019, at 07:39, an orange warning signal for dry hot wind was issued, and it was expected that dry hot wind would occur within the next 24 h [21].

Data Source
The historical meteorological data used in this study were obtained from the European Space Agency [23], with a spatial resolution of 1 km and a temporal resolution of hourly data.The European Space Agency meteorological data use Coordinated Universal Time (UTC).Four meteorological elements were used in this study, namely 2 m temperature, 2 m dew point temperature, wind speed u-component, and wind speed v-component.
The remote sensing satellite data used in this study were collected from the Moderateresolution Imaging Spectroradiometer (MODIS), which is operated by the Terra and Aqua satellites and provides medium-resolution imaging spectrometer data.One MODIS image with the row-column number "h27v05" covers the entire Henan Province.Among them, MOD09GA [24] is the daily land surface reflectance product provided by the Terra satellite, which includes seven bands of land surface reflectance with a spatial resolution of 500 m.The spectral characteristics of green vegetation mainly include strong absorption in the red band and strong reflection in the near-infrared band.Various vegetation indices, mainly those based on the combination of red and near-infrared channels, can enhance vegetation information and reflect the growth status of plants.Commonly used vegetation indices such as the NDVI and EVI can respond to primary damage caused by a reduction in functional leaf chlorophyll and can be used to quantitatively assess the disaster situation of crops [8].In this study, the NDVI was calculated using the red band (sur_refl_b01_1) and near-infrared band (sur_refl_b02_1) of MOD09GA, and the calculation formula is as follows: NDV where NIR represents the near-infrared band reflectance and R represents the red band reflectance.
The distribution of wheat planting in Henan in 2013 was derived from the "1 km Planting Distribution Dataset of the Three Major Grain Crops in China (2000-2019)" published by the National Ecological Science Data Center.The temporal resolution is annual, and the spatial resolution is 1 km.This dataset is extracted based on the Leaf Area Index (LAI) product from the Global Land Surface Characteristics Parameters (GLASS) product [25].
The GLASS product is a long-term, high-precision global land surface remote sensing product derived through the inversion of multi-source remote sensing data and ground measurement data [26].

Spatio-Temporal Knowledge Graph Construction
The Web Ontology Language (OWL) was chosen as the semantic expression language for constructing the spatio-temporal knowledge graph.The construction of the spatiotemporal knowledge graph primarily involved domain knowledge collection, ontology construction, knowledge extraction, knowledge fusion, formalization of inference rules, and triple storage.The research framework of this paper is illustrated in Figure 1.
The distribution of wheat planting in Henan in 2013 was derived from the "1 km Planting Distribution Dataset of the Three Major Grain Crops in China (2000-2019)" published by the National Ecological Science Data Center.The temporal resolution is annual, and the spatial resolution is 1 km.This dataset is extracted based on the Leaf Area Index (LAI) product from the Global Land Surface Characteristics Parameters (GLASS) product [25].The GLASS product is a long-term, high-precision global land surface remote sensing product derived through the inversion of multi-source remote sensing data and ground measurement data [26].

Spatio-Temporal Knowledge Graph Construction
The Web Ontology Language (OWL) was chosen as the semantic expression language for constructing the spatio-temporal knowledge graph.The construction of the spatiotemporal knowledge graph primarily involved domain knowledge collection, ontology construction, knowledge extraction, knowledge fusion, formalization of inference rules, and triple storage.The research framework of this paper is illustrated in Figure 1.

Ontology Construction
During ontology construction, we first designed an overall conceptual framework, as shown in Figure 2. The ontology is divided into semantic ontology, spatial ontology, temporal ontology, and rule objects.

Ontology Construction
During ontology construction, we first designed an overall conceptual framework, as shown in Figure 2. The ontology is divided into semantic ontology, spatial ontology, temporal ontology, and rule objects.
The semantic ontology includes geographic entities such as farmland patches, crops, farmland managers, remote sensing data, disaster reports, etc.The main entities, entity properties, and relationships between entities are shown in Figure 3. Entity properties are divided into object properties and data properties.The domain specifies the definition domain of the property, and the range specifies the value domain of the property.In particular, the spatio-temporal knowledge graph does not directly store the actual remote sensing image data.Instead, it semantically models their key characteristics, such as spatial resolution, temporal resolution, sensor type, and acquisition path.During analysis and computation, based on the computation requirements, the corresponding remote sensing image entities were inferred and computed based on semantic features such as spatial resolution, temporal resolution, sensor type, and acquisition path.The semantic ontology includes geographic entities such as farmland patches, crops farmland managers, remote sensing data, disaster reports, etc.The main entities, entity properties, and relationships between entities are shown in Figure 3. Entity properties ar divided into object properties and data properties.The domain specifies the definition domain of the property, and the range specifies the value domain of the property.In par ticular, the spatio-temporal knowledge graph does not directly store the actual remot sensing image data.Instead, it semantically models their key characteristics, such as spa tial resolution, temporal resolution, sensor type, and acquisition path.During analysi and computation, based on the computation requirements, the corresponding remot sensing image entities were inferred and computed based on semantic features such a spatial resolution, temporal resolution, sensor type, and acquisition path.The semantic ontology also includes meteorological grids, specifically Level1Hour lyMeteorologicalGrid (referred to as Level1Grid) for hourly data, Level2DailyMeteorolog  The semantic ontology includes geographic entities such as farmland patches, crops, farmland managers, remote sensing data, disaster reports, etc.The main entities, entity properties, and relationships between entities are shown in Figure 3. Entity properties are divided into object properties and data properties.The domain specifies the definition domain of the property, and the range specifies the value domain of the property.In particular, the spatio-temporal knowledge graph does not directly store the actual remote sensing image data.Instead, it semantically models their key characteristics, such as spatial resolution, temporal resolution, sensor type, and acquisition path.During analysis and computation, based on the computation requirements, the corresponding remote sensing image entities were inferred and computed based on semantic features such as spatial resolution, temporal resolution, sensor type, and acquisition path.The semantic ontology also includes meteorological grids, specifically Level1Hour-lyMeteorologicalGrid (referred to as Level1Grid) for hourly data, Level2DailyMeteorolog-icalGrid (referred to as Level2Grid) for daily data, and Level3PeriodMeteorologicalGrid (referred to as Level3Grid) for data within a specific time period.Among them, Level1Grid is further divided into various meteorological indicators.The properties of The semantic ontology also includes meteorological grids, specifically Level1Hourly MeteorologicalGrid (referred to as Level1Grid) for hourly data, Level2DailyMeteorologicalGrid (referred to as Level2Grid) for daily data, and Level3PeriodMeteorologicalGrid (referred to as Level3Grid) for data within a specific time period.Among them, Level1Grid is further divided into various meteorological indicators.The properties of different meteorological grids are shown in Table 3.The relationships between meteorological grids and other entities will be introduced in Section 2.3.4.The spatial ontology adopts the GeoSPARQL [27] spatial semantic representation specification, whose structure is shown in Figure 4.The GeoSPARQL ontology is based on the feature model of the Open Geospatial Consortium (OGC) and includes a class called geo: SpatialObject, which has two main subclasses, geo: Feature and geo: Geometry.For example, a farmland patch is a geo: Feature, which is a conceptual entity, and it also has a geo: Geometry to describe its spatial extent.The spatial ontology adopts the GeoSPARQL [27] spatial semantic repres specification, whose structure is shown in Figure 4.The GeoSPARQL ontology on the feature model of the Open Geospatial Consortium (OGC) and includes called geo: SpatialObject, which has two main subclasses, geo: Feature and geo: Ge For example, a farmland patch is a geo: Feature, which is a conceptual entity, an has a geo: Geometry to describe its spatial extent.The temporal ontology adopts the SWRL Temporal Ontology (SWRLTO) model [28], whose structure is shown in Figure 5. Objects that require the association of temporal semantic information are abstracted as "spatio-temporal entities".Instances of "spatiotemporal entities" serve as subjects in typical triples of this model and are associated with instances of the "valid time" object as objects through the predicate temporal: hasValidTime.The "valid time" object can be further divided into two subclasses: temporal: ValidInstant and temporal: ValidPeriod, which are used to represent the semantics of specific instants and time periods, respectively.mantic information are abstracted as "spatio-temporal entities".Instances of "spa poral entities" serve as subjects in typical triples of this model and are associat instances of the "valid time" object as objects through the predicate temporal: h Time.The "valid time" object can be further divided into two subclasses: tempo lidInstant and temporal: ValidPeriod, which are used to represent the semantics of instants and time periods, respectively.

Knowledge Extraction
In this paper, both the farmland data and the meteorological data go through steps, including clipping, coordinate system standardization, scaling and round conversion to GeoJSON format (as shown in Figure 6).However, only the farmla are further converted into triples because these data are updated only based on growth cycle, which can span several months, half a year, or a year.But the meteor data are updated frequently, and the proportion of abnormal data that can cause d is relatively small compared to normal data.Therefore, including all meteorolog in the database would result in a significant and meaningless storage burden.W fied abnormal data based on the meteorological indicators and their thresholds t stitute agro-meteorological disasters (e.g., the meteorological indicators and their olds for dry hot winds as shown in Table 1).For other data, we first organized them into a CSV table, as shown in Tabl then extracted triplets from them.

Knowledge Extraction
In this paper, both the farmland data and the meteorological data go through several steps, including clipping, coordinate system standardization, scaling and rounding, and conversion to GeoJSON format (as shown in Figure 6).However, only the farmland data are further converted into triples because these data are updated only based on the crop growth cycle, which can span several months, half a year, or a year.But the meteorological data are updated frequently, and the proportion of abnormal data that can cause disasters is relatively small compared to normal data.Therefore, including all meteorological data in the database would result in a significant and meaningless storage burden.We identified abnormal data based on the meteorological indicators and their thresholds that constitute agro-meteorological disasters (e.g., the meteorological indicators and their thresholds for dry hot winds as shown in Table 1).
poral entities" serve as subjects in typical triples of this model and are associated with instances of the "valid time" object as objects through the predicate temporal: hasValid Time.The "valid time" object can be further divided into two subclasses: temporal: Va lidInstant and temporal: ValidPeriod, which are used to represent the semantics of specifi instants and time periods, respectively.

Knowledge Extraction
In this paper, both the farmland data and the meteorological data go through severa steps, including clipping, coordinate system standardization, scaling and rounding, and conversion to GeoJSON format (as shown in Figure 6).However, only the farmland dat are further converted into triples because these data are updated only based on the crop growth cycle, which can span several months, half a year, or a year.But the meteorologica data are updated frequently, and the proportion of abnormal data that can cause disaster is relatively small compared to normal data.Therefore, including all meteorological dat in the database would result in a significant and meaningless storage burden.We identi fied abnormal data based on the meteorological indicators and their thresholds that con stitute agro-meteorological disasters (e.g., the meteorological indicators and their thresh olds for dry hot winds as shown in Table 1).For other data, we first organized them into a CSV table, as shown in Table 4, and then extracted triplets from them.For other data, we first organized them into a CSV table, as shown in Table 4, and then extracted triplets from them.

Knowledge Fusion
For geospatial entities with spatio-temporal attributes such as agricultural land plots and meteorological grids, in addition to the properties specified in Figure 3, additional multiscale geocoding index information is required.We calculated the geocoding based on the Web Mercator projection grid partitioning method (as shown in Figure 7) and indexed the geospatial entities in the spatio-temporal knowledge graph as follows: <Subject: Geospatial Entity Predicate: hasTileCode Object: Tile Code>.
on the Web Mercator projection grid partitioning method (as shown in Figure 7) and indexed the geospatial entities in the spatio-temporal knowledge graph as follows: <Subject: Geospatial Entity Predicate: hasTileCode Object: Tile Code>.
The grid partitioning level is generally in the range of 0-26.A geospatial entity has a series of tile codes at different scales, represented as strings.For example, "z14_x13335_y6860" indicates that the Mercator projection grid has been partitioned with a spatial x-coordinate of 13,335, a y-coordinate of 6860, and a grid code level of 14. Different types of geospatial entities insert multiscale tile codes during the construction of the knowledge graph.The matching based on the tile codes associated with feature entities supports multiscale spatial queries of spatio-temporal feature entities.

Monitoring Reasoning Engine
The reasoning engine designed in this paper consists of a series of spatio-temporal semantic reasoning rules, represented as RuleObject.Each rule is composed of an event object (TriggerObject, abbreviated as Tr) and an action object (ActionObject, abbreviated as Ac), represented as RuleObject = (Tr, Ac).Tr represents the event object included in RuleObject, Ac represents the action object included in RuleObject, and RuleObject represents the inference result.In this paper, the event TriggerObject is defined as a triplet, represented as TriggerObject = (O, T, S), where O represents the set of geospatial entities included in the event object, and T and S represent the intersection of this geospatial entity set in the temporal and spatial dimensions, respectively.A spatio-temporal co-occurrence scenario with a set of geospatial entities can be described as an event object, which serves as the definition of the condition for applying a reasoning rule.When the condition is satisfied, the reasoning engine triggers the execution of the action object to obtain the inference result.The concepts of event (or action) objects are further divided into independent events (or actions) and event (or action) combinations, which together constitute the The grid partitioning level is generally in the range of 0-26.A geospatial entity has a series of tile codes at different scales, represented as strings.For example, "z14_x13335_y6860" indicates that the Mercator projection grid has been partitioned with a spatial x-coordinate of 13,335, a y-coordinate of 6860, and a grid code level of 14. Different types of geospatial entities insert multiscale tile codes during the construction of the knowledge graph.The matching based on the tile codes associated with feature entities supports multiscale spatial queries of spatio-temporal feature entities.

Monitoring Reasoning Engine
The reasoning engine designed in this paper consists of a series of spatio-temporal semantic reasoning rules, represented as RuleObject.Each rule is composed of an event object (TriggerObject, abbreviated as Tr) and an action object (ActionObject, abbreviated as Ac), represented as RuleObject = (Tr, Ac).Tr represents the event object included in RuleObject, Ac represents the action object included in RuleObject, and RuleObject represents the inference result.In this paper, the event TriggerObject is defined as a triplet, represented as TriggerObject = (O, T, S), where O represents the set of geospatial entities included in the event object, and T and S represent the intersection of this geospatial entity set in the temporal and spatial dimensions, respectively.A spatio-temporal co-occurrence scenario with a set of geospatial entities can be described as an event object, which serves as the definition of the condition for applying a reasoning rule.When the condition is satisfied, the reasoning engine triggers the execution of the action object to obtain the inference result.The concepts of event (or action) objects are further divided into independent events (or actions) and event (or action) combinations, which together constitute the concept of knowledge inference rules [16].The formalization of the reasoning rules designed in this paper is shown in Figure 8. concept of knowledge inference rules [16].The formalization of the reasoning rules designed in this paper is shown in Figure 8.The principle of the reasoning engine executing "GiveEarlyWarning" is as follows: After obtaining meteorological data, the reasoning engine integrates it into Level1Grid, Level2Grid, and Level3Grid sequentially.Then, based on certain criteria for agro-meteorological disasters, it filters out abnormal meteorological grids.If there are abnormal meteorological grids, it calculates the spatio-temporal intersection between the abnormal grids and the farmland patches affected by this type of disaster.If the result is not empty, it indicates that certain farmland patches are at risk for this type of meteorological disaster.An early warning is issued for these farmland patches, providing the spatial extent, land management personnel, start and end dates of the disaster, severity level of the disaster, and recommended defense measures.After the meteorological data are updated, the above steps are repeated.If there are changes in the disaster information, the warning information is updated accordingly.
The principle of the reasoning engine executing "CalculateNDVIDifference" is as follows: After a disaster occurs, the reasoning engine obtains pre-and post-disaster remote sensing data and calculates the NDVI decline matrix within the farmland patches.It then returns the average and maximum values to the knowledge graph for storage, which are used to record the actual extent of the affected farmland patches.
The overall reasoning flow is shown in Figure 9, where the left half corresponds to the "GiveEarlyWarning" content mentioned above and the right half corresponds to the "CalculateNDVIDifference" content mentioned above.Green lines represent the information to be retrieved from the knowledge graph, and yellow lines represent the information to be added to the knowledge graph.The principle of the reasoning engine executing "GiveEarlyWarning" is as follows: After obtaining meteorological data, the reasoning engine integrates it into Level1Grid, Level2Grid, and Level3Grid sequentially.Then, based on certain criteria for agro-meteorological disasters, it filters out abnormal meteorological grids.If there are abnormal meteorological grids, it calculates the spatio-temporal intersection between the abnormal grids and the farmland patches affected by this type of disaster.If the result is not empty, it indicates that certain farmland patches are at risk for this type of meteorological disaster.An early warning is issued for these farmland patches, providing the spatial extent, land management personnel, start and end dates of the disaster, severity level of the disaster, and recommended defense measures.After the meteorological data are updated, the above steps are repeated.If there are changes in the disaster information, the warning information is updated accordingly.
The principle of the reasoning engine executing "CalculateNDVIDifference" is as follows: After a disaster occurs, the reasoning engine obtains pre-and post-disaster remote sensing data and calculates the NDVI decline matrix within the farmland patches.It then returns the average and maximum values to the knowledge graph for storage, which are used to record the actual extent of the affected farmland patches.
The overall reasoning flow is shown in Figure 9, where the left half corresponds to the "GiveEarlyWarning" content mentioned above and the right half corresponds to the "Cal-culateNDVIDifference" content mentioned above.Green lines represent the information to be retrieved from the knowledge graph, and yellow lines represent the information to be added to the knowledge graph.

Knowledge Storage
The factual triplets and rule triplets formed from the above steps were stored and visualized using the GraphDB database in this study.Specifically, meteorological data are only placed outside the graph database and serve as the event triggering reasoning in the knowledge graph.

Knowledge Storage
The factual triplets and rule triplets formed from the above steps were stored and visualized using the GraphDB database in this study.Specifically, meteorological data are only placed outside the graph database and serve as the event triggering reasoning in the knowledge graph.

Results
This study selected meteorological and remote sensing data from before and after the two disasters mentioned in Table 2 to monitor and assess the impact of dry hot wind disasters.

Spatio-Temporal Knowledge Graph
In this study, a pixel was considered as a farmland patch.The information of farmland patches in the spatio-temporal knowledge graph is shown in Figure 10, including three types of information: basic information such as farmland management personnel, crop types, etc.; spatial information including TileCode, GridElement, etc.; and temporal information including the validity period of the farmland patch, specifically StartTime, FinishTime, etc.

Results
This study selected meteorological and remote sensing data from before and after the two disasters mentioned in Table 2 to monitor and assess the impact of dry hot wind disasters.

Spatio-Temporal Knowledge Graph
In this study, a pixel was considered as a farmland patch.The information of farmland patches in the spatio-temporal knowledge graph is shown in Figure 10, including three types of information: basic information such as farmland management personnel, crop types, etc.; spatial information including TileCode, GridElement, etc.; and temporal information including the validity period of the farmland patch, specifically StartTime, FinishTime, etc.

Pre-Disaster Monitoring Results
Since meteorological forecast data are constantly updated, the reasoning engine, upon detecting that a farmland patch will be affected by dry hot wind starting from a certain day, adds a DisasterReport entity node to the farmland patch node, recording its start and end dates, severity level, and other relevant information.Subsequently, with the updated meteorological forecast data, the attributes of this DisasterReport entity node will also be continuously updated until the disaster is over.This process forms a historical archive of the warning results.When the farmland patch is detected to be affected by a disaster again, a new DisasterReport entity node will be added to the corresponding node.
This paper illustrates the above principle using the example of the dry hot wind disaster that occurred in Qi County in May 2013.We assumed the meteorological forecast data have a cycle of 7 days with updates occurring once daily in the following experiment.
On 5 May, meteorological forecast data from 6 May to 12 May were obtained.It was detected that a certain farmland patch would experience a dry hot wind from the 11th to the 12th, with a moderate severity on the 11th and a severe severity on the 12th.Consequently, a meteorological disaster warning was issued for this farmland patch, and a DryHotWindReport node was added.

Pre-Disaster Monitoring Results
Since meteorological forecast data are constantly updated, the reasoning engine, upon detecting that a farmland patch will be affected by dry hot wind starting from a certain day, adds a DisasterReport entity node to the farmland patch node, recording its start and end dates, severity level, and other relevant information.Subsequently, with the updated meteorological forecast data, the attributes of this DisasterReport entity node will also be continuously updated until the disaster is over.This process forms a historical archive of the warning results.When the farmland patch is detected to be affected by a disaster again, a new DisasterReport entity node will be added to the corresponding node.
This paper illustrates the above principle using the example of the dry hot wind disaster that occurred in Qi County in May 2013.We assumed the meteorological forecast data have a cycle of 7 days with updates occurring once daily in the following experiment.
On 5 May, meteorological forecast data from 6 May to 12 May were obtained.It was detected that a certain farmland patch would experience a dry hot wind from the 11th to the 12th, with a moderate severity on the 11th and a severe severity on the 12th.Consequently, a meteorological disaster warning was issued for this farmland patch, and a Dry-HotWindReport node was added.
On 6 May, meteorological forecast data from 7 May to 13 May were obtained.It was detected that the same farmland patch would experience a dry hot wind from the 11th to the 13th, with a moderate severity on the 11th and 13th, and a severe severity on the 12th.Therefore, another meteorological disaster warning was issued for this farmland patch, and the attributes of the DryHotWindReport node were modified accordingly, as shown in Figure 11.On 6 May, meteorological forecast data from 7 May to 13 May were obtained.It was detected that the same farmland patch would experience a dry hot wind from the 11th to the 13th, with a moderate severity on the 11th and 13th, and a severe severity on the 12th.Therefore, another meteorological disaster warning was issued for this farmland patch, and the attributes of the DryHotWindReport node were modified accordingly, as shown in Figure 11.Similarly, the warning information was promptly updated with the updates of meteorological forecast data.
The warning results for Qi County, Yanjin County, and Wuyang County are shown in Tables 5 and 6.Similarly, the warning information was promptly updated with the updates of meteorological forecast data.
The warning results for Qi County, Yanjin County, and Wuyang County are shown in Tables 5 and 6.The warning messages are shown in Figure 12.In the dry hot wind warning results, the information provided includes the location of the farmland patch, disaster situation, measures for preventing dry hot wind, and farmland management personnel.

Post-Disaster Monitoring Results
For each farmland patch, the reasoning engine obtains remote sensing data from the day before and the day after the disaster.It further calculates the NDVI to determine the actual extent of damage.The average decrease in the NDVI on farmland patches with different degrees of damage in each county is shown in Tables 7 and 8.The decrease in the

Post-Disaster Monitoring Results
For each farmland patch, the reasoning engine obtains remote sensing data from the day before and the day after the disaster.It further calculates the NDVI to determine the actual extent of damage.The average decrease in the NDVI on farmland patches with different degrees of damage in each county is shown in Tables 7 and 8.The decrease in the NDVI is also used as an attribute for the corresponding disaster report entity node of the affected farmland patch.The two tables above also confirm the correlation between the severity of damage to wheat caused by dry hot wind and the intensity and duration of a dry hot wind occurrence [29].Therefore, in situations where resources and manpower are limited, priority should be given to implementing disaster risk reduction measures in agricultural areas that are more severely affected by disasters.

Query Results
After the above steps, the disaster information at the farmland patch level is also added to the knowledge graph, resulting in a continuously updated agro-meteorological disaster database.Various queries can be performed using the SPARQL Protocol and RDF Query Language (SPARQL) [30] query statements.Figure 13 shows three example questions.
The query results are as follows: "3", "1322 km 2 ", and "Wuyang County".This implies that based on the existing information in the database, there are three counties in Henan Province that have experienced dry hot wind.The farmland area affected by dry hot wind in 2013 was 1322 square kilometers, and Wuyang County was the county most affected by dry hot wind in 2019.

Query Results
After the above steps, the disaster information at the farmland patch level is also added to the knowledge graph, resulting in a continuously updated agro-meteorological disaster database.Various queries can be performed using the SPARQL Protocol and RDF Query Language (SPARQL) [30] query statements.Figure 13 shows three example questions.The query results are as follows: "3", "1322 km 2 ", and "Wuyang County".This implies that based on the existing information in the database, there are three counties in Henan Province that have experienced dry hot wind.The farmland area affected by dry hot wind in 2013 was 1322 square kilometers, and Wuyang County was the county most affected by dry hot wind in 2019.

Discussion
This paper proposes a technical method based on a spatio-temporal knowledge graph for the integrated organization and management of remote sensing data, meteorological data, farmland data, crop information, expert knowledge, and computational models.The knowledge graph starts reasoning spontaneously from the input meteorological data, sequentially discovering dry hot wind occurrences, calculating pre-and post-disaster NDVI decline values, and supplementing the knowledge graph with the disaster information.For two dry hot wind disasters, the knowledge graph constructed in this study provided more specific and accurate disaster information through reasoning.Therefore,

Discussion
This paper proposes a technical method based on a spatio-temporal knowledge graph for the integrated organization and management of remote sensing data, meteorological data, farmland data, crop information, expert knowledge, and computational models.The knowledge graph starts reasoning spontaneously from the input meteorological data, sequentially discovering dry hot wind occurrences, calculating pre-and post-disaster NDVI decline values, and supplementing the knowledge graph with the disaster information.For two dry hot wind disasters, the knowledge graph constructed in this study provided more specific and accurate disaster information through reasoning.Therefore, for tasks such as detecting dry hot wind disasters and using remote sensing to assess meteorological disasters, the proposed method in this paper is a valuable technique worth exploring.
However, there are still some limitations in this research.Firstly, the spatial resolution of the wheat distribution data [25] used in this paper is relatively low, and the pixels cannot accurately represent actual farmland patches.Therefore, in practical applications, remote sensing intelligent interpretation should be used to obtain higher-resolution crop distribution data.If ground survey data can be obtained, the farmland patches can be further integrated into actual land management units (such as farms or plantations), so that warning information can be accurately communicated to the farmland managers.
Secondly, the historical meteorological data used in this paper also have a low spatial resolution.Improving the spatial resolution of these meteorological data would be more beneficial in providing more accurate warning results at the plot level.
Thirdly, for the NDVI difference matrix before and after disasters within the farmland patches, the paper only includes statistical measures such as the average value.If more domain knowledge, such as the relationship between NDVI decline and actual damage indicators (such as yield), can be obtained, it would be possible to provide indicators that are easier for farmland management personnel to understand and use after obtaining the NDVI decline values.
Lastly, in this research, remote sensing images are only used for monitoring calculations after predicting the occurrence of dry hot wind based on meteorological data.This results in the accuracy of warnings relying mainly on meteorological data.Objectively, continuous monitoring of crop growth using remote sensing images can automatically detect disasters based on their changes, forming a coupling relationship between meteorological data warning and remote sensing data monitoring to a greater extent.

Conclusions
In the field of remote sensing, the aim of Earth spatial information services is to deliver the right data/information/knowledge at the right time to the right person in the right place [31].In the context of agro-meteorological disaster monitoring tasks, this paper proposes an effective method: (1) This paper constructs a spatio-temporal knowledge graph for agro-meteorological disaster monitoring, integrating multiple heterogeneous data sources such as meteorological data, remote sensing data, farmland data, and agricultural knowledge.This enables the integration and intelligent analysis of pre-disaster meteorological disaster monitoring and post-disaster impact analysis driven by data and knowledge cooperation.
(2) This paper uses remote sensing techniques to refine the granularity of meteorological disaster monitoring and warning to the farmland patch level, providing a new approach for fine-grained agricultural management.
(3) This paper incorporates MODIS remote sensing image data directly into knowledge reasoning and computation.MODIS remote sensing data are involved in the daily monitoring process, making full use of the advantages of long-term continuous observation of crop remote sensing data.It explores the synergistic development path of the participation of remote sensing in spatio-temporal knowledge computation and reasoning.(4) The knowledge graph is supplemented with the results of monitoring, early warning, and evaluation analysis in each iteration as disaster information, achieving automatic iterative updating of the knowledge graph.Over the long term, it naturally becomes an agro-meteorological disaster database.
Based on the spatio-temporal knowledge graph and reasoning process for agrometeorological disaster monitoring and warning constructed in this paper, further research can be carried out as follows: (1) Regarding historical data, although dry hot wind meteorological disasters occur frequently, have a wide impact range, affect multiple crops, and cause significant losses, detailed and comprehensive statistical data have not yet been formed.In line with the research method proposed in this paper, historical dry hot wind meteorological disaster data could be extracted from national-level historical meteorological data over the past few years, forming a historical dry hot wind disaster database.Using this database, the spatial-temporal patterns of dry hot wind disasters and the relationship between meteorological conditions and NDVI differences can be explored.(2) Looking towards the future, once issues such as real-time acquisition of meteorological and remote sensing data are resolved, the method proposed in this paper can be directly applied to agro-meteorological disaster monitoring tasks.Additionally, since the warning in this paper relies on meteorological data, the accuracy of meteorological forecast data directly impacts the effectiveness of agro-meteorological disaster warnings.By incorporating the "spatio-temporal patterns of dry hot wind meteorological disasters" mentioned in point (1) as an auxiliary for monitoring and warning, the accuracy of warnings can be improved.

Figure 3 .
Figure 3.The main entities, entity properties, and relationships among entities in a semantic ontol ogy.

Figure 3 .
Figure 3.The main entities, entity properties, and relationships among entities in a semantic ontology.

Figure 3 .
Figure 3.The main entities, entity properties, and relationships among entities in a semantic ontology.

Figure 4 .
Figure 4. Illustration of the GeoSPARQL Spatial Ontology structure.

Figure 4 .
Figure 4. Illustration of the GeoSPARQL Spatial Ontology structure.

Figure 5 .
Figure 5. Illustration of the SWRL Temporal Ontology structure.

Figure 5 .
Figure 5. Illustration of the SWRL Temporal Ontology structure.

Figure 5 .
Figure 5. Illustration of the SWRL Temporal Ontology structure.

Figure 9 .
Figure 9. Inference process in the inference engine.

Figure 9 .
Figure 9. Inference process in the inference engine.

Figure 10 .
Figure 10.Example of farmland patches in a spatio-temporal knowledge graph.

18 Figure 11 .
Figure 11.Adding a DisasterReport instance to a farmland patch instance.

Figure 11 .
Figure 11.Adding a DisasterReport instance to a farmland patch instance.

18 Figure 12 .
Figure 12.Example of agro-meteorological disaster monitoring and early warning instance.

Figure 12 .
Figure 12.Example of agro-meteorological disaster monitoring and early warning instance.

Table 1 .
The meteorological indicators for dry hot wind days in the winter wheat region of northern China.

Table 2 .
Dry hot wind early warning records.

Table 3 .
Meteorological grid entity and properties.

Table 4 .
Example of a CSV table.

Table 5 .
Monitoring and early warning results for dry hot wind in 2013, shown as number of farmland patches of different severity levels.

Table 6 .
Monitoring and early warning results for dry hot wind in 2019, shown as number of farmland patches of different severity levels.

Table 7 .
Average NDVI decrease in differently affected farmland patches in 2013.

Table 8 .
Average NDVI decrease in differently affected farmland patches in 2019.