Rice leaf age tracking method and system based on time sequence unmanned aerial vehicle images

Abstract

This invention discloses a method and system for tracking rice leaf age based on time-series UAV images, belonging to the field of agricultural informatics technology. The method includes: receiving a sequence of rice canopy orthophoto images, along with corresponding image acquisition timestamps, geographical location information, rice variety information, and time-series meteorological data. Based on the rice variety information, the method obtains the growth baseline temperature and photoperiod sensitivity parameters of the rice variety; constructs a photothermal integrated physiological time axis using the time-series meteorological data and photoperiod sensitivity parameters; and converts the image acquisition timestamps into photothermal integrated growth degree-days. Based on the photothermal integrated physiological time axis, the method resamples the rice canopy orthophoto image sequence to extract a standardized time-series image block set. This invention, based on optimized event sequence reasoning, generates a spatially consistent rice field leaf age distribution map, achieving high-precision and robust tracking of rice leaf age dynamics.

Description

Technical Field

[0001] This invention relates to the field of agricultural informatics technology, and in particular to a method and system for tracking rice leaf age based on time-series unmanned aerial vehicle (UAV) images. Background Technology

[0002] Rice leaf age is a key parameter guiding precision agronomic management. Dynamic monitoring of leaf age based on UAV time-series images has become an important research direction. Existing technologies mainly involve periodically collecting canopy images by UAVs, extracting visual or deep learning features, and constructing a regression model from image features to leaf age values, thereby achieving automated leaf age estimation and avoiding the shortcomings of manual observation.

[0003] Existing methods rely on calendar time to align time-series images, ignoring the essential growth-driving factors of light and temperature. This leads to a decline in the model's generalization ability across regions and sowing periods. The model learns the apparent statistical correlation between images and leaf age numbers, rather than the biological events of orderly leaf occurrence itself. It is difficult to handle complex situations such as leaf occlusion and lacks a biological verification mechanism. The methods mostly process single plant samples independently and fail to utilize the spatial continuity and regularity of population growth to optimize the results. This limits the spatial rationality and consistency of leaf age distribution maps and restricts the in-depth application of the technology in high-precision decision-making scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a rice leaf age tracking method based on time-series UAV images to solve the problems of poor generalization, insufficient reliability in complex scenarios, and low spatial consistency of results caused by the disconnect between existing technologies and biological laws.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for tracking the leaf age of rice based on time-series UAV images, which includes receiving a rice canopy orthophoto sequence, an image acquisition timestamp, geographical location information, rice variety information, and time-series meteorological data corresponding to the geographical location corresponding to the rice canopy orthophoto sequence. Based on rice variety information, the growth baseline temperature and photoperiod sensitivity parameters of rice varieties are obtained. A photo-temperature integrated physiological time axis is constructed using time-series meteorological data and photoperiod sensitivity parameters, and the image acquisition timestamp is converted into a photo-temperature integrated growth degree-day. Based on the photothermal integrated physiological time axis, the rice canopy orthophoto image sequence was resampled to extract a standardized time-series image block set. The standardized time-series image block set is input into the event detection neural network to obtain preliminary leaf emergence event detection results; Based on the fractal structure rules of rice leaf order, the preliminary leaf emergence event detection results are decoupled into single-plant events. Based on the field population growth wave propagation model, the spatial consistency of the decoupled events is optimized to obtain the optimized tillering-level leaf age event sequence. Based on the optimized tillering-level leaf age event sequence, leaf age statistics and spatial mapping are performed to output a spatial distribution map of leaf age in rice fields.

[0007] As a preferred embodiment of the rice leaf age tracking method based on time-series UAV images described in this invention, the method includes the following steps: receiving a rice canopy orthophoto sequence, image acquisition timestamps, geographic location information, rice variety information, and time-series meteorological data corresponding to the geographic location, along with the rice canopy orthophoto sequence. The system receives rice canopy orthophoto sequences collected and transmitted by drones via a data interface, and obtains the image acquisition timestamps that are time-synchronized with the rice canopy orthophoto sequences by parsing the drone flight logs. By reading the records from the drone's positioning device, we can obtain the geographical location information that corresponds to the rice canopy orthophoto sequence and has been spatially registered. By querying the field management database, we can obtain rice variety information corresponding to the fields photographed by the rice canopy orthophoto sequence. We can also access the meteorological data service interface to obtain time-series meteorological data that matches the area specified by the geographical location information.

[0008] As a preferred embodiment of the rice leaf age tracking method based on time-series UAV images described in this invention, the method includes the following steps: obtaining the growth baseline temperature and photoperiod sensitivity parameters of the rice variety based on rice variety information. Based on rice variety information, we query the rice variety trait database and obtain the growth baseline temperature of rice varieties from the rice variety trait database. Photoperiod sensitivity parameters of rice varieties were obtained from a rice variety trait database.

[0009] As a preferred embodiment of the rice leaf age tracking method based on time-series UAV images described in this invention, the method includes the following steps: constructing a photothermal integrated physiological time axis using time-series meteorological data and photoperiod sensitivity parameters, and converting the image acquisition timestamp into a photothermal integrated growth degree-day: Daily average temperature and sunshine duration data are extracted from time-series meteorological data. The daily average temperature is compared with the growth baseline temperature of rice varieties to evaluate the daily effective accumulated temperature that reflects the temperature-driven effect of the day. By combining sunshine duration data with photoperiod sensitivity parameters of rice varieties, the daily photoperiod influence factor of the degree of inhibition was quantified. The daily effective accumulated temperature is nonlinearly modulated by the daily photoperiod influence factor to generate a daily photothermal modulated effective accumulated temperature that integrates both photothermal and photothermal effects. The effective accumulated temperature of daily photothermal modulation after sowing is continuously summed to form a photothermal comprehensive physiological time axis based on the accumulated physiological development amount; The image acquisition timestamp is located and mapped on the photothermal integrated physiological time axis to obtain the photothermal integrated growth day corresponding to the image acquisition time.

[0010] As a preferred embodiment of the rice leaf age tracking method based on time-series UAV images described in this invention, the method includes the following steps: resampling the rice canopy orthophoto image sequence based on a photothermal integrated physiological time axis to extract a standardized time-series image block set: Target physiological sampling points are set at equal intervals on the photothermal integrated physiological time axis. Based on the position of the target physiological sampling points on the photothermal integrated physiological time axis, the corresponding image frames are matched from the rice canopy orthophoto sequence. For the corresponding matching image frame, image blocks are slidably extracted using a grid window of fixed size, and radiometric normalization and size scaling operations are performed on all extracted image blocks. All image blocks that have completed the standardization process will be integrated into a standardized temporal image block set in temporal and spatial order.

[0011] As a preferred embodiment of the rice leaf age tracking method based on time-series UAV images described in this invention, the method involves inputting a standardized time-series image block set into an event detection neural network to obtain preliminary leaf emergence event detection results, including the following steps: The standardized temporal image patch set is input into the pre-trained event detection neural network, which performs spatiotemporal feature extraction and pattern analysis on the standardized temporal image patch set. The event detection neural network outputs preliminary leaf emergence event detection results, including the spatial coordinates of the event occurrence, the combined light and temperature growth period of the event occurrence, and the initial confidence level of the event occurrence.

[0012] As a preferred embodiment of the rice leaf age tracking method based on time-series UAV images described in this invention, the method includes the following steps: decoupling single-plant events based on the preliminary leaf emergence event detection results according to the rice leaf sequence fractal structure rules: Based on the preliminary leaf emergence event detection results of spatial location clustering, a set of candidate event points for a single plant is generated. For a single plant candidate event point set, the fractal dimension, branch topology, spatial geometry and time sequence are matched and verified with the preset rice leaf order fractal structure rule template, and the multidimensional matching verification results are output. Based on the comprehensive multidimensional matching verification results, the candidate event points of individual plants are assigned to specific tillering main stems and outliers are identified, thus completing the decoupling of individual plant events.

[0013] As a preferred embodiment of the rice leaf age tracking method based on time-series UAV images described in this invention, the method includes: optimizing the spatial consistency of decoupled events based on a field population growth wave propagation model to obtain an optimized tillering-level leaf age event sequence, comprising the following steps: By using the spatial coordinates of the event point after decoupling the single-plant event and the corresponding light and temperature integrated growth degree day, a field population growth wave propagation model characterizing the propagation law of the event in the field is fitted through spatial autocorrelation analysis. Based on the field population growth wave propagation model, the spatial residual is calculated to evaluate the spatiotemporal rationality of each decoupled event point, and to generate the adjustment amount of position offset, time offset and confidence weight. Based on the adjustments to the location offset, time offset, and confidence weight, position correction, time correction, and confidence weighting are performed on the decoupled event points. By integrating all event points that have undergone location correction, time correction, and confidence weighting, an optimized tillering leaf age event sequence is generated.

[0014] As a preferred embodiment of the rice leaf age tracking method based on time-series UAV images described in this invention, the method includes the following steps: performing leaf age statistics and spatial mapping based on the optimized tillering-level leaf age event sequence to output a spatial distribution map of leaf age in the rice field. The event points of each tiller in the optimized tillering leaf age event sequence are sorted by the combined light and temperature growth degree per day; The number of event points for each tiller after sorting is defined as the current leaf age of the tiller. The spatial coordinates of each tiller and the corresponding current leaf age of the tiller are extracted to form a leaf age-spatial point dataset. Spatial interpolation is performed on the leaf age-spatial point dataset to generate a continuous field leaf age raster surface, and then a rendering output is performed to show the spatial distribution map of leaf age in the rice field.

[0015] Secondly, the present invention provides a rice leaf age tracking system based on time-series UAV images, including a preprocessing module that receives a rice canopy orthophoto sequence, an image acquisition timestamp corresponding to the rice canopy orthophoto sequence, geographical location information, rice variety information, and time-series meteorological data corresponding to the geographical location. The photothermal physiology module, based on rice variety information, obtains the growth baseline temperature and photoperiod sensitivity parameters of rice varieties, and uses time-series meteorological data and photoperiod sensitivity parameters to construct a photothermal comprehensive physiological time axis, converting image acquisition timestamps into photothermal comprehensive growth degree days. The standardization processing module resamples the rice canopy orthophoto sequence based on the photothermal integrated physiological time axis to extract a set of standardized time-series image blocks. The recognition module inputs a standardized set of time-series image blocks into the event detection neural network to obtain preliminary leaf emergence event detection results. The optimization module decouples the initial leaf emergence event detection results into single-plant events based on the fractal structure rules of rice leaf order, and optimizes the spatial consistency of the decoupled events based on the field population growth wave propagation model to obtain the optimized tillering-level leaf age event sequence. The statistics module performs leaf age statistics and spatial mapping based on the optimized tillering-level leaf age event sequence, and outputs a spatial distribution map of leaf age in rice fields.

[0016] The beneficial effects of this invention are as follows: By receiving orthophoto sequences of rice canopy collected by UAVs and their associated multi-source information, a physiological time axis integrating photothermal factors is constructed. The time-series images are uniformly mapped to a daily scale reflecting the comprehensive photothermal growth degree of real growth and development. Based on this, the image sequences are resampled and standardized on the physiological time scale. Leaf biological events are identified from the standardized images using an event detection neural network. Events at the single-plant scale are decoupled and assigned by combining the inherent leaf sequence fractal structure rules of rice. The spatiotemporal consistency of the decoupled events is optimized using a field population growth wave propagation model. Based on the optimized event sequence, a highly spatially consistent rice field leaf age distribution map is generated, achieving high-precision and robust tracking of rice leaf age dynamics. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a rice leaf age tracking method based on time-series UAV images.

[0019] Figure 2 This is a schematic diagram of a rice leaf age tracking system based on time-series drone images. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0023] Reference Figures 1-2 As one embodiment of the present invention, this embodiment provides a method for tracking the leaf age of rice based on time-series unmanned aerial vehicle (UAV) images, comprising the following steps: S1. Receive rice canopy orthophoto sequence, image acquisition timestamp, geographical location information, rice variety information, and time-series meteorological data corresponding to the geographical location.

[0024] S1.1 Receive the rice canopy orthophoto sequence collected and transmitted by the UAV through the data interface, and obtain the image acquisition timestamp that is time-synchronized with the rice canopy orthophoto sequence by parsing the UAV flight log.

[0025] Furthermore, by using a dedicated data interface pre-defined for the UAV data stream, the system receives rice canopy orthophoto sequence data packets transmitted by the UAV acquisition platform and pre-compressed and packaged. Simultaneously with receiving the rice canopy orthophoto sequence, a parallel parsing process for the flight log files of the same UAV is automatically triggered. Based on the unique exposure event identifier and GPS time recorded for each image in the flight log, a millisecond-precision image acquisition timestamp is established and bound to each image frame in the received rice canopy orthophoto sequence. This process ensures that the image data and time metadata maintain a strict correspondence at the source of generation, avoiding order disorder or loss of association in subsequent transmission and storage stages, and ultimately achieving inherent synchronization between the image and the timestamp.

[0026] Specifically, by parsing flight logs in real-time upon reception, and using globally unique, immutable identifiers of exposure events recorded by the drone's hardware as keys, a rigid binding is achieved directly at the data stream level between each frame and a timestamp accurate to the instant of exposure. This low-level binding makes time synchronization an inherent attribute of the data, rather than a subsequently added, potentially erroneous annotation. For example, even if the frame order is accidentally rearranged during data transmission, the original timing sequence can be accurately restored using the identifiers, ensuring the absolute reliability and immutability of the time information that forms the basis of timing analysis, and eliminating subsequent analysis failures caused by timing errors at the source.

[0027] S1.2. By reading the records from the UAV positioning device, obtain the geographical location information that corresponds to the rice canopy orthophoto sequence and has been spatially registered.

[0028] Furthermore, during the process of parsing the UAV flight logs to obtain image acquisition timestamps, the geographical location information recorded in the flight logs—collected by the high-precision real-time dynamic positioning device integrated into the UAV at the moment of each exposure—is simultaneously read. This geographical location information typically includes latitude and longitude coordinates and elevation. The read geographical location information has been fused and calculated by the inertial measurement unit built into the positioning device and global satellite navigation system data during the UAV's flight, and directly associated with the camera center projection position of each image. In the stitching process of generating the rice canopy orthophoto sequence, this geographical location information bound to each original image is used as control points input to the georegistration algorithm, driving the stitching process to directly produce orthophoto products with a unified geographic coordinate system and accurate spatial scale. Thus, the final rice canopy orthophoto sequence has each pixel naturally possessing spatially registered geographical location information. Specifically, this method utilizes and transmits the in-situ geospatial information stream acquired by UAV remote sensing data, elevating geographic location from a later-added reference information to a fundamental parameter driving the core processing flow. Existing technologies often treat image geometric correction and georegistration as independent post-processing steps, potentially using ground control points or low-precision UAV location recordings. This method directly employs the camera center's 3D coordinates, captured by a high-precision positioning device at the moment of exposure and strictly synchronized with the image exposure. Crucially, this geographic location information is not a static label but is directly used as an indispensable spatial control constraint in the core geometric transformation process of orthophoto stitching. This means that geographic location information participates in the entire geometric reconstruction process from the original oblique image to the orthophoto product, and its accuracy directly determines the spatial accuracy of the output image. This ensures that the geographic location information used for spatial registration is an intrinsic result of the geometric quality of the image product, rather than an externally added attribute.

[0029] S1.3. Obtain rice variety information corresponding to the fields photographed by the rice canopy orthophoto sequence by querying the field management database, and access the meteorological data service interface to obtain time-series meteorological data that matches the area specified by the geographical location information.

[0030] Furthermore, based on the spatial boundaries of the fields covered by the rice canopy orthophoto sequence, a unique spatial identifier for each field is generated. This unique spatial identifier is then used as a query key to initiate a request to the networked field management database. The field management database stores the agricultural records registered for each field during the planting season, and by querying, it can accurately return the information on the rice varieties sown in that field during the current growth cycle. Simultaneously, based on the area defined by the spatially registered geographic location information attached to the rice canopy orthophoto sequence, a data request containing both spatial boundaries and a temporal range is constructed and invoked to an authoritative meteorological data service interface. The meteorological data service interface, based on the spatial range in the request, returns time-series meteorological data covering the geographic area, within the image acquisition time period, and with a time resolution matching the image acquisition timestamp, thus completing the acquisition of time-series meteorological data that perfectly matches the area specified by the geographic location information.

[0031] Specifically, an automated and precise metadata association link from space to attributes was constructed, using the geospatial entities (fields) inherent in the image data itself as the hub connecting multi-source heterogeneous data. Existing technologies for acquiring variety and meteorological data often rely on manual input or rough regional matching, which is prone to errors where data does not correspond to specific fields. By automatically deriving unique spatial identifiers for fields through the image's geographic location, unambiguous and automated association from remote sensing observation entities to agricultural management records (variety) is achieved. Meteorological data is requested based on the image's geographic range (rather than a simple center point) and precise acquisition time windows, ensuring that the acquired meteorological data completely overlaps with the fields spatially and precisely corresponds to each flight acquisition event temporally. The time-series images acquired by UAVs are transformed from isolated image sequences into a spatiotemporal anchor point, automatically extracting two core environmental explanatory variables driving crop growth changes in the images: variety genetic characteristics and meteorological environment sequences. This provides accurate and matching input for the subsequent construction of photothermal physiological timelines, fundamentally ensuring the consistency of growth-driving factor analysis and visual observation objects in terms of spatiotemporal scales, and solving the problems of model interpretation bias and generalization failure caused by mismatch between variety or meteorological data in existing technologies.

[0032] S2. Based on rice variety information, obtain the growth baseline temperature and photoperiod sensitivity parameters of rice varieties.

[0033] S2.1. Based on the rice variety information, query the rice variety trait database to obtain the growth baseline temperature of the rice variety from the rice variety trait database.

[0034] Furthermore, the received rice variety information is used as a structured query request key to access a centralized and standardized rice variety trait database. The rice variety trait database is indexed by variety name and pre-stores standardized physiological and ecological attribute data of various varieties. The query request precisely matches the corresponding variety record in the rice variety trait database and extracts the growth baseline temperature value, which represents the threshold temperature for the onset of vegetative growth of the variety, from the record. After data extraction, the growth baseline temperature is encapsulated into a data object bound to the rice variety information for subsequent calculation processes, realizing an accurate mapping from variety identifier to specific temperature parameters.

[0035] Specifically, the acquisition of driving parameters for rice leaf age tracking models has shifted from empirically based fixed values ​​or manual literature reviews to automated and precise queries based on standardized databases. Existing technologies typically simplify models by using a single empirical growth baseline temperature for different rice varieties, ignoring genetic differences between varieties. This leads to deviations in physiological time calculations when the model is applied to different varieties. By using rice variety information as a unique key to directly connect to a structured rice variety trait database, we achieve real-time acquisition of variety-specific physiological parameters, directly linking the core physiological driving parameters of the crop model to the specific genetic background of the variety. For example, different varieties from tropical and temperate zones may have significantly different growth baseline temperatures. This method automatically adapts to these differences, ensuring that the subsequently constructed photothermal integrated physiological time axis accurately reflects the genetic characteristics of the developmental rate's response to temperature for a specific variety. This provides the model with accurate input of variety physiological characteristics, overcoming the simulation errors in developmental processes caused by inaccurate parameters in existing technologies.

[0036] S2.2 Obtain the photoperiod sensitivity parameters of rice varieties from the rice variety trait database.

[0037] Furthermore, after successfully querying the rice variety trait database and obtaining the growth baseline temperature based on the same rice variety information, the same database session and query connection are maintained. In the variety record data field corresponding to the same rice variety information, the photoperiod sensitivity parameter used to quantify the degree of response of the variety's development rate to changes in day length is located and read. The photoperiod sensitivity parameter is expressed in a specific data format, describing the modulation intensity and direction of the unit of day length change on the variety's leaf emergence rate and other developmental processes. The obtained photoperiod sensitivity parameter and the previously obtained growth baseline temperature together constitute the developmental response feature set of the variety, completing the complete retrieval from rice variety information to the key developmental driving parameter set.

[0038] Specifically, by synchronously acquiring two key developmental response parameters—temperature and photoperiod—from the same authoritative data source, the inherent consistency and matching between parameters are ensured. Continuous extraction of growth baseline temperature and photoperiod sensitivity parameters from the same rice variety trait database ensures that the two parameters originate from the same trait description system for the same variety and possess homology. For example, for a variety sensitive to photoperiod, its photoperiod sensitivity parameter clearly indicates the promoting or inhibiting effect of long-day conditions. This parameter, combined with growth baseline temperature, is necessary to fully characterize the true developmental driving patterns of the variety in a specific region and season. This achieves a complete digital profile of the variety's developmental characteristics, enabling the subsequent construction of a photo-temperature integrated physiological timeline to simultaneously integrate the variety's genetic response characteristics to temperature and photoperiod. This allows for precise simulation of the variety's developmental process under real, variable photo-temperature environments, overcoming the limitations of existing technologies that only consider accumulated temperature or use coarse photoperiod corrections. This lays a solid foundation in variety biology for precise leaf age tracking across latitudes and sowing periods.

[0039] S3. Construct a photothermal integrated physiological time axis using time-series meteorological data and photoperiod sensitivity parameters, and convert the image acquisition timestamp into a photothermal integrated growth day.

[0040] S3.1 Extract daily average temperature and sunshine duration data from time-series meteorological data, compare the daily average temperature with the growth baseline temperature of rice varieties, and evaluate the daily effective accumulated temperature that reflects the temperature-driven effect of the day.

[0041] Furthermore, the acquired time-series meteorological data is analyzed daily to extract the daily average temperature and sunshine duration data. The daily average temperature data is then compared with the growth baseline temperature of rice varieties obtained from the rice variety trait database. When the daily average temperature is higher than the growth baseline temperature of the rice variety, the difference between the two is obtained. This difference is evaluated as the daily effective accumulated temperature, which reflects the driving effect of the daily temperature on rice development. If the daily average temperature is not higher than the growth baseline temperature of the rice variety, it is determined that the daily temperature does not drive development, and the daily effective accumulated temperature is recorded as zero. This completes the daily quantitative evaluation of the temperature driving effect in the time-series meteorological data.

[0042] Specifically, the assessment of temperature-driven effects has shifted from using fixed environmental temperature thresholds to an assessment based on variety-specific physiological parameters as a dynamic benchmark. Daily average temperatures are compared with the precise growth baseline temperatures of corresponding rice varieties obtained from authoritative databases, enabling variety-customized calculations of developmental drivers. For example, a tropical rice variety may have a higher growth baseline temperature, while a temperate variety may have a lower one. Under the same spring temperatures, the former may accumulate little or no effective accumulated temperature, while the latter may have already begun active growth. This dynamic comparison ensures that the daily effective accumulated temperature accurately reflects the physiological response of a specific variety under the current temperature.

[0043] S3.2. Combining sunshine duration data with the photoperiod sensitivity parameters of rice varieties, the daily photoperiod influence factor of the degree of inhibition was quantified.

[0044] Furthermore, while obtaining the daily effective accumulated temperature, the sunshine duration data of the same day is analyzed from the time-series meteorological data, and the photoperiod sensitivity parameters of the same rice variety are obtained from the rice variety trait database. The sunshine duration data is input into a response function defined based on the photoperiod sensitivity parameters of the rice variety to describe the quantitative modulation relationship between the change in day length and the development rate, and output a value. This value is the daily photoperiod influence factor that quantifies the degree to which the day's sunshine conditions promote or inhibit rice development.

[0045] Specifically, photoperiod, an independent environmental factor, is transformed into a quantifiable and calculable developmental modulation coefficient through a variety-specific response function. Most existing technologies neglect the role of photoperiod or use only seasonal months as a rough reference, failing to accurately quantify the real-time impact of day length on developmental rates. This method combines daily measured sunshine duration with the variety's photoperiod sensitivity parameters, dynamically calculating the daily photoperiod influence factor through a predefined response function. For example, for a short-day rice variety, in the autumn when day length gradually decreases, the photoperiod influence factor may be greater than one, indicating promoted development; in the summer with longer day lengths, the factor may be less than one, indicating inhibited development. This expands the characterization of the environmental drivers of development from a single temperature dimension to the dimension of photothermal interaction, accurately capturing the real-time regulatory effect of day length changes on the variety's developmental process through the daily updated photoperiod influence factor.

[0046] S3.3. The daily effective accumulated temperature is nonlinearly modulated using the daily photoperiod influence factor to generate a daily photothermal modulated effective accumulated temperature that integrates both photothermal and photothermal effects.

[0047] Furthermore, after obtaining the daily baseline effective accumulated temperature and the daily photoperiod influence factor, the daily baseline effective accumulated temperature and the daily photoperiod influence factor are multiplied. However, this multiplication is not a simple linear multiplication, but rather a nonlinear modulation relationship. That is, the daily photoperiod influence factor acts as a dynamic weight or scaling factor on the daily baseline effective accumulated temperature. Its role is to amplify, reduce, or even reverse the temperature-driven contribution. Through this operation, a new value is generated. This value is the daily phototemperature modulated effective accumulated temperature, which integrates the dual environmental effects of daily temperature and photoperiod and jointly affects the physiological development of rice.

[0048] Specifically, the daily photoperiod influence factor is considered a modulator acting on the daily basal effective accumulated temperature. For example, under promoting photoperiods, a factor greater than one indicates a stronger development-driving effect under the same temperature conditions; under inhibiting photoperiods, a factor less than one indicates a weakened development-driving effect under the same temperature. This nonlinear modulation relationship is more consistent with the mechanism in plant physiology where photoperiod regulates growth point activity and nutrient allocation through endogenous hormone pathways, thereby affecting leaf primordia differentiation and elongation. The generated daily photo-temperature modulated effective accumulated temperature is an indicator that truly reflects the net physiological development driving force of the day under photo-temperature interaction, providing key data for constructing a high-fidelity physiological time baseline.

[0049] S3.4 Continuously sum the daily effective accumulated temperature of light and temperature modulation after sowing to form a comprehensive physiological time axis of light and temperature based on the accumulated physiological development amount.

[0050] Furthermore, the rice sowing date is determined, and starting from the first day after the sowing date, the daily effective accumulated temperature of light and temperature modulation generated each day is sequentially accumulated. The value of the first day is accumulated on the first day, the values ​​of the first and second days are accumulated on the second day, and so on. All daily effective accumulated temperatures of light and temperature modulation from sowing to the current or target date are continuously summed. This continuous summing process produces a cumulative value sequence that monotonically increases over time. This cumulative value sequence, which is measured by the cumulative physiological development amount and strictly corresponds to the real calendar time series, constitutes the light and temperature integrated physiological time axis, which characterizes the total physiological development process of rice since sowing, in units of light and temperature integrated growth degree days.

[0051] Specifically, a physiological development timeline based on the combined photothermal growth degree-day (PDG) replaces traditional calendar time or ordinary accumulated temperature time. Existing technologies mostly use calendar dates or accumulated temperatures as time references, failing to eliminate the impact of differences in photothermal combinations across different years and regions on developmental speed. By continuously accumulating the effective accumulated temperature modulated by daily photothermal modulation, which incorporates photothermal interaction effects, a physiological development isochrone is generated. For example, a rice variety experiencing a specific photothermal combination in location A accumulates to a certain combined photothermal growth degree-day; while experiencing different photothermal combinations in location B, theoretically, it should reach the same physiological development stage when accumulating the same combined photothermal growth degree-day. This unifies the variable external photothermal environment sequence into comparable physiological times driving the internal developmental process of the variety. This provides a fundamental solution for comparing rice growth and development under different environments and aligning visual-based temporal analysis to a unified physiological development phase, overcoming the bottleneck of poor model generalization ability caused by inconsistent timelines in existing technologies.

[0052] S3.5. Locate and map the image acquisition timestamp on the photothermal integrated physiological time axis to obtain the photothermal integrated growth day corresponding to the image acquisition time.

[0053] Furthermore, each image acquisition timestamp that is strictly corresponding to the rice canopy orthophoto sequence is searched and matched on the constructed photothermal integrated physiological time axis. The photothermal integrated physiological time axis itself is defined by the date sequence and the corresponding cumulative photothermal modulation effective accumulated temperature value sequence. The search process determines the specific date corresponding to the image acquisition timestamp, and then reads the cumulative photothermal modulation effective accumulated temperature value corresponding to that date from the photothermal integrated physiological time axis. This read value is the photothermal integrated growth degree-day corresponding to the image acquisition time, thus completing the mapping conversion from physical clock time to physiological development scale.

[0054] Specifically, the physical time stamps of drone-acquired images are replaced with physiological development phase stamps. Existing time-series analyses are based on the calendar date of image capture, implicitly assuming that identical time intervals represent identical developmental stages, which is inaccurate. By mapping image acquisition timestamps to a photothermal integrated physiological timeline, each image is assigned a photothermal integrated growth degree daily value. This value no longer represents when the image was captured, but rather when it was captured within the plant's entire physiological development process. For example, images taken in different years, locations, and dates, if their photothermal integrated growth degree daily values ​​are the same, indicate that the rice was at the same or highly similar internal physiological development stage at the time of capture. This mapping ensures that all subsequent analyses based on time-series images, such as event detection and growth comparison, are built on a unified and comparable physiological development phase, rather than an incomparable natural calendar, greatly enhancing the robustness and comparability of the time-series analysis model to changes in the spatiotemporal environment. S4. Based on the photothermal integrated physiological time axis, the rice canopy orthophoto sequence is resampled to extract a standardized time-series image block set.

[0055] S4.1 Set equally spaced target physiological sampling points on the photothermal integrated physiological time axis, and match the corresponding image frames from the rice canopy orthophoto sequence based on the position of the target physiological sampling points on the photothermal integrated physiological time axis.

[0056] Furthermore, on the constructed photothermal integrated physiological timeline, starting from the initial photothermal integrated growth degree-day, a series of target physiological sampling points are sequentially set with a fixed photothermal integrated growth degree-day interval as the step size. Each target physiological sampling point corresponds to a specific photothermal integrated growth degree-day value. Based on the photothermal integrated growth degree-day value corresponding to each target physiological sampling point, a search and match is performed in the converted photothermal integrated growth degree-day labels of the rice canopy orthophoto image sequence. The rice canopy orthophoto image sequence frame whose photothermal integrated growth degree-day label value is closest to the photothermal integrated growth degree-day value of the target physiological sampling point is selected as the image frame corresponding to the target physiological sampling point, thus completing the mapping from sampling points with equal physiological development intervals to actual image frames.

[0057] Specifically, the resampling benchmark for time-series images is transformed from unreliable calendar time to comparable, uniform physiological development time. Existing techniques typically select images at fixed calendar intervals (e.g., every 5 days), ignoring the non-uniform growth caused by light and temperature fluctuations in the field environment. This method sets equally spaced target physiological sampling points on the photothermal integrated physiological time axis, essentially requiring a representative image to be collected every time the crop completes the same unit of physiological development. For example, in the early growing season when temperatures are suitable and sunshine is sufficient, the crop accumulates photothermal integrated growth degrees per day relatively quickly, and the calendar date interval between two samplings may be short; when encountering low temperatures or unfavorable photoperiods, accumulation slows down, and the calendar interval automatically lengthens. The resulting image frame sequence, where each frame represents a snapshot of equal developmental progress in the growth process, allows subsequent analysis to directly address the crop's inherent, uniform developmental rhythm, completely eliminating the interference of temporal irregularities caused by environmental fluctuations on model learning, and providing the event detection network with a truly synchronous and comparable input sequence in physiological time.

[0058] S4.2 For the corresponding matching image frame, slide a fixed-size grid window to capture image blocks, and perform radiometric normalization and size scaling operations on all captured image blocks.

[0059] Furthermore, for each rice canopy orthophoto image frame obtained by matching target physiological sampling points, a rectangular grid window of fixed size is defined. Starting from the upper left corner of the image frame, the grid window is slid from top to bottom and from left to right across the entire space of the image frame with a set step size. At each sliding stop position, the image area covered by the grid window is cropped as an image block. For all image blocks obtained by sliding cropping, histogram matching or reflectivity conversion algorithm based on whiteboard reflectivity is uniformly applied for radiometric normalization processing to eliminate the differences in radiometric brightness between different image blocks caused by differences in lighting conditions and shooting angles. Then, for all radiometrically normalized image blocks, bilinear or bicubic interpolation algorithms are used to scale their size to a uniform standard pixel size, completing the standardization operation of image blocks in terms of radiometric and geometric scales.

[0060] Specifically, by combining alignment along the physiological time dimension with systematic and standardized sampling along the spatial dimension, a spatiotemporally regular primary feature unit is constructed. Existing technologies for processing field images may randomly crop or focus on local areas, lacking a systematic and comprehensive capture of the entire field's spatial information. Furthermore, radiometric correction is often performed at the whole-image level, insufficiently addressing heterogeneous illumination in local blocks. Using a fixed-size grid window for sliding cropping ensures that every spatial location in the field is sampled uniformly, without overlap or with moderate overlap, forming an image patch set. Independent radiometric normalization of each image patch is particularly crucial. For example, within the same field, cloud shadows and plant shading can cause localized uneven brightness, which global correction of the entire image cannot eliminate. Independent radiometric normalization of each small image patch effectively eliminates this localized illumination noise, ensuring that each input unit (image patch) received by the subsequent neural network operates on a consistent radiometric scale, focusing on vegetation structure rather than illumination changes. This design, which combines global spatial system sampling with local radiation fine normalization, ensures the spatial representativeness and radiation consistency of the input features, removing a major obstacle for neural networks to extract robust spatiotemporal growth features.

[0061] S4.3 Integrate all image blocks that have completed the standardization operation into a standardized temporal image block set according to time and spatial order.

[0062] Furthermore, all image blocks that have undergone radiometric normalization and size scaling are sorted temporally according to the order of the target physiological sampling points corresponding to their source image frames, i.e., organized in ascending order of photothermal growth degree. All image blocks captured within the image frame corresponding to the same target physiological sampling point are spatially sorted according to the order of grid sliding capture, such as following the row and column order from left to right and from top to bottom. All image blocks arranged in this dual temporal and spatial order are organized into a structured data set, which is the standardized temporal image block set. Each image block can be uniquely identified and accessed through its index in the time series and its position in the spatial grid.

[0063] Specifically, through rigorous spatiotemporal dual-sequence integration, the generated standardized time-series image patch set itself becomes a data structure. This structure implicitly encodes the four-dimensional coordinates of each image patch: three-dimensional physical spatiotemporal coordinates (longitude, latitude, and combined growth degree-days of light and temperature) and a one-dimensional linear position within this structured set. For example, when the event detection neural network processes this set, it can not only analyze the content of the image patch but also easily deduce, through its index within the set, which location in the field the image patch represents and which stage of crop development it represents. The spatiotemporal semantics inherent in the original images are seamlessly and losslessly transferred and solidified into the final digital representation of the input model through resampling, cropping, and sorting operations. This allows the model to simultaneously utilize appearance features and implicit spatiotemporal context features for learning and reasoning, greatly enhancing the model's understanding of the spatiotemporal patterns of events. It serves as a crucial bridge connecting the front-end physiological timeline with the back-end intelligent event detection.

[0064] S5. Input the standardized time-series image block set into the event detection neural network to obtain preliminary leaf emergence event detection results.

[0065] S5.1 Input the standardized time-series image block set into the trained event detection neural network. The event detection neural network performs spatiotemporal feature extraction and pattern analysis on the standardized time-series image block set.

[0066] Furthermore, a structured, standardized temporal image patch set is fed into an event detection neural network that has been pre-trained with a large amount of labeled temporal image data. The event detection neural network typically adopts a three-dimensional convolutional neural network or a temporal Transformer architecture. Its network layers perform convolution, pooling, or self-attention operations on the input standardized temporal image patch set layer by layer, thereby automatically extracting spatiotemporal visual features that can characterize the morphological changes of the rice canopy at continuous physiological time points from the deep feature map of the network. Based on these spatiotemporal features, the network analyzes and judges whether a specific growth pattern of new leaf emergence is occurring at each spatial location at each time point, thus completing the transformation from raw image data to high-level growth event semantics.

[0067] Specifically, the paradigm shifts from mainstream image regression to temporal event detection. Current techniques train neural networks to learn the mapping between features of single or a few static images and an abstract leaf age number. The models fail to grasp the biological essence of leaf age growth as the sequential occurrence of discrete leaf emergence events. Event detection neural networks, however, are trained with a structure designed to identify dynamic visual patterns of emerging leaves from short, physiologically time-aligned image fragments. For example, the key patterns the network needs to learn are the continuous changes in shape extension and color texture of the leaf sheath opening region over several frames, rather than the total number or length of leaves in a single image. This forces the network to focus on the growth process rather than the state, enabling it to learn more discriminative and generalizable features related to event dynamics. Because the input sequence is aligned on a photothermal integrated physiological timeline, the patterns learned by the network are synchronized with the intrinsic physiological developmental rhythm.

[0068] S5.2 The event detection neural network outputs the preliminary leaf emergence event detection results, including the spatial coordinates of the event occurrence, the combined light and temperature growth period of the event occurrence, and the initial confidence level of the event occurrence.

[0069] Furthermore, in its final output layer, the event detection neural network simultaneously generates three sets of associated output data for each spatial location and time point identified as a possible event: the first set is the planar spatial coordinates representing the actual location of the event in the field, obtained by converting the original spatial index of the input standardized time-series image patch set with georegistration parameters; the second set is the daily value of the photothermal growth degree associated with the center frame of the input time-series image corresponding to the event, which represents the physiological time of the event; the third set is the probability value, calculated by the event detection neural network through the Softmax function or a similar mechanism, characterizing the credibility of the judgment result, i.e., the initial confidence level of the event occurrence; these three sets of data together constitute a structured output unit, and the set of all such output units is the preliminary leaf emergence event detection result.

[0070] Specifically, the event detection neural network outputs a multi-dimensional, fine-grained set of attributes with clear biological and spatiotemporal semantics, rather than a single scalar. Existing networks typically only output an estimate representing leaf age, losing all process and spatial information. The network synchronously outputs a triplet of location-physiological time-confidence for each detected event, directly parsing the visual perception result into a biological event object with spatiotemporal attributes. For example, an output unit not only tells the user that an event has been detected, but also explicitly states that the event occurred at a specific location in the field, at a specific physiological stage of crop development, and with a certain confidence level. The introduction of the integrated light-temperature growth time of the event occurrence is particularly crucial. It transforms the event timestamp from a meaningless calendar time into a physiologically comparable developmental phase, allowing events occurring at different times and locations to be directly compared and ranked on a unified developmental scale. The spatial coordinates of the event occurrence and the initial confidence level of the event occurrence provide indispensable input dimensions for subsequent optimization and selection based on spatial rules (fractal structure) and statistical laws (growth wave model).

[0071] S6. Based on the fractal structure rules of rice leaf order, decouple the preliminary leaf emergence event detection results for individual plants.

[0072] S6.1. Based on the preliminary leaf emergence event detection results of spatial location clustering, generate a set of candidate event points for a single plant.

[0073] Furthermore, for all the spatial coordinates of events included in the preliminary leaf emergence event detection results, a density-based spatial clustering algorithm is applied. The algorithm automatically identifies core points and boundary points based on the distribution density of event points on the plane. Event points whose distance to each other is less than the algorithm's neighborhood radius and whose number of core points reaches the minimum threshold are grouped into the same cluster. Each cluster generated by the algorithm represents the spatial range of a rice plant. All event points within a cluster constitute a single-plant candidate event point set, thereby completing the organization from discrete event points across the entire field to a set of candidate event points based on the plant.

[0074] Specifically, it abandons the conventional approach of dividing plant areas according to fixed grids or regular geometric shapes, and instead adopts an unsupervised density clustering algorithm to adaptively define the boundaries of individual plants. Existing technologies often lead to missegmentation or merging of individual plant events due to the imperfect uniformity of rice planting spacing and the influence of perspective distortion in UAV imagery. Density-based spatial clustering algorithms can keenly capture the natural clustering patterns of event points in the field. For example, in densely growing areas, the algorithm automatically identifies smaller, more compact clusters, while in sparse or missing areas, it forms larger, more dispersed clusters. It utilizes the spatial distribution characteristics of the event points themselves to infer the location of individual plants without needing to know the precise row and plant spacing in advance. This robustly decouples global event points into independent sets of candidate individual plant event points, providing accurate spatial units for subsequent analysis based on individual plant structure.

[0075] S6.2. For the single-plant candidate event point set, perform matching verification with the preset rice leaf order fractal structure rule template from the perspectives of fractal dimension, branch topology, spatial geometry and time series, and output the multi-dimensional matching verification results.

[0076] Furthermore, for each candidate event point set, four independent matching verification processes are executed to compare against a pre-defined rice leaf order fractal structure rule template. The first process extracts the fractal dimension of the spatial distribution of the candidate event point set, estimates it using box counting, and compares it with the template value. The second process analyzes the minimum spanning tree or hierarchical clustering tree of the point set, extracts its branching pattern, and compares its graph structure similarity with the theoretical branching topology of the template. The third process measures the angle and length of the line connecting each pair of event points in the point set, and statistically analyzes the degree of agreement between their distribution and the theoretical interleaf angle and tiller distance distribution of the template. The fourth process sorts the point set according to the combined light, temperature, and growth intensity of the event occurrence, and analyzes the correlation between the time interval sequence of adjacent event points and the temporal pattern predicted by the template's co-elongation rule. Each verification process generates a quantified matching score or consistency judgment, which are then summarized to form a multi-dimensional matching verification result.

[0077] Specifically, the holistic biological law of rice leaf arrangement fractal structure is decomposed into multiple complementary and independently computable geometric, topological, and temporal constraints, forming a multi-dimensional verification system. Existing technologies lack the ability to utilize inherent crop morphogenesis rules to verify and interpret visual detection results. Fractal dimension matching captures the self-similarity and scale invariance characteristics of rice leaf spatial arrangement; branch topology matching ensures that the connection relationships between event points conform to the hierarchical derivation logic of the main stem and tillers; spatial geometric matching verifies the rationality of local geometric attributes such as leaf extension angle and inter-tiller distance; and temporal matching further examines whether the sequence of events conforms to physiological laws from a developmental dynamics perspective. For example, even if the spatial location of event points is slightly deviated due to image distortion, their attribution can still be correctly inferred through branch topology and temporal matching. This multi-dimensional cross-validation design means that the assessment of the biological rationality of detection results no longer relies on a single, fragile standard, but rather improves the robustness and credibility of decoupled decision-making through multi-evidence fusion.

[0078] S6.3. Based on the comprehensive multidimensional matching verification results, assign the candidate event points of a single plant to the specific tillering main stem and identify outliers to complete the decoupling of single plant events.

[0079] Furthermore, based on the fractal dimension matching degree, branch topology conformity, spatial geometric fit, and temporal consistency measure provided by the multidimensional matching verification results, a weighted voting or probabilistic graphical model is used for comprehensive reasoning. This process obtains the probability that each event point in the candidate event point set belongs to each possible leaf position or tiller node in the template. The attribute with the highest probability is selected as the final label for that event point. Points with an attribute probability lower than a set threshold, or points that show abnormalities in multiple dimensions of verification, are marked as outliers. Finally, the specific tiller main stem attribution label and outlier identifier for each event point are output, thus resolving the candidate event point set of a single plant into clear event sequences organized by tillers, completing the decoupling of single plant events.

[0080] Specifically, by comprehensively utilizing the matching and verification results across four dimensions—fractal, topological, geometric, and temporal—the attribution probability is calculated for each event point. For example, an event point might be spatially geometrically closer to the location of tiller A, but temporally more consistent with the predicted occurrence time of tiller B. In this case, branch topology and fractal dimension information will serve as crucial evidence in arbitration. This probabilistic fusion method allows evidence from different dimensions to participate in the decision-making process with different weights and can naturally handle conflicting evidence, ultimately identifying outliers. This enables the maximal recovery of fine structures consistent with biological laws even when detection results are imperfect, achieving substantial refinement of leaf age tracking from the plant level to the tiller level, providing unprecedented granularity of growth analysis.

[0081] S7. Based on the field population growth wave propagation model, spatial consistency optimization is performed on the decoupled events to obtain the optimized tillering-level leaf age event sequence.

[0082] S7.1 Using the spatial coordinates of the event point after decoupling the single-plant event and the corresponding light and temperature comprehensive growth degree day, a field population growth wave propagation model characterizing the propagation law of the event in the field is fitted by spatial autocorrelation analysis.

[0083] Furthermore, by utilizing the spatial coordinates of all event points obtained after decoupling individual plant events and their corresponding light-temperature-growth degree days, these data are regarded as a set of spatiotemporal sample points distributed in a two-dimensional field space, with each point having a time attribute value. Using spatial interpolation techniques, such as ordinary kriging interpolation, with spatial coordinates as independent variables and light-temperature-growth degree days of the event occurrence as dependent variables, spatial autocorrelation analysis and variogram modeling are performed based on the sample point data to fit a continuous two-dimensional surface function. This function can predict the expected value of light-temperature-growth degree days of the event occurrence at any input field spatial location coordinate.

[0084] Specifically, this method reconstructs discrete, single-plant-scale event observations into a growth field model describing the continuous, macroscopic growth trend in the field using spatial statistical methods. Existing techniques often analyze event points independently, neglecting the spatial correlation and gradient of field crop growth. This method treats the spatial location and occurrence time (physiological time) of each event point as discrete samples of an implicit, continuous spatiotemporal growth field, reconstructing it using spatial interpolation techniques such as Kriging. This captures and quantifies the spatial patterns of growth event propagation in the field; for example, due to micro-gradients in water and fertilizer, topography, or wind direction, growth wavefronts may advance from one side of the field to the other. The fitted field population growth wave propagation model is essentially a continuous spatiotemporal reference surface predicting where and when events should occur. This provides a global, data-driven reference framework for evaluating the spatiotemporal rationality of individual event points, rather than relying on subjective or fixed rules, achieving a paradigm shift from isolated point detection to collective perception.

[0085] S7.2 Based on the field population growth wave propagation model, the spatial residual is calculated to evaluate the spatiotemporal rationality of each decoupled event point, and the adjustment amounts of position offset, time offset and confidence weight are generated. Furthermore, the spatial coordinates of each decoupled event point are input into the fitted field population growth wave propagation model to obtain the predicted light-temperature-growth degree-day of the event at that location; the actual observed light-temperature-growth degree-day of the event at that location is then read; the absolute value of the spatial residual directly measures the degree of deviation between the event time and the predicted time of the field population growth wave propagation model; based on the absolute value of the spatial residual, a corresponding adjustment amount is generated through a preset mapping function: the position offset is usually calculated based on the residual sign and local spatial gradient to determine the possible position correction direction and magnitude, while the time offset is directly related to the residual and used to adjust the position offset of the event at that location. The actual observed event occurrence time of each decoupled event point is reverse-corrected, and the confidence weight is decayed based on the absolute value of the residual. The larger the residual, the lower the confidence weight, thus generating a set of quantitative adjustment parameters for each event point, including location offset, time offset, and confidence weight.

[0086] Specifically, the spatiotemporal rationality assessment of event points is transformed into a calculable, continuous numerical index, thereby driving multi-dimensional automatic correction. Existing technologies lack a quantitative assessment and feedback mechanism for detection results based on population spatial consistency. The proposed spatial residual has a clear physical meaning: a positive residual indicates that the actual event occurred later than the population trend prediction, possibly due to local stress or missed detections followed by correction; a negative residual indicates that it occurred earlier than the prediction, possibly due to false detection or dominant plant growth. The residual is not merely used for filtering but is used as a source for generating adjustment quantities. For example, an event point with a small positive residual may be fine-tuned with a negative time offset to bring it closer to the prediction surface; a point with a large residual has its confidence weight significantly reduced, but it is not necessarily deleted immediately, as it may represent real abnormal growth. The generation of refined adjustment quantities based on residuals means that subsequent optimization is no longer a simple binary decision of retention or rejection, but allows for flexible, proportional correction of the event point's time, location, and confidence, reflecting a data-driven and model-guided error correction strategy.

[0087] The expression for the spatial residual is: ; in, For the first Spatial residuals of each event point For the first The actual observed event occurrence time at each decoupled event point For the first Predicting the time of occurrence of an event at a given event point location. Index for the event point.

[0088] S7.3. Based on the adjustment of position offset, time offset and confidence weight, perform position correction, time correction and confidence weighting on the decoupled event points.

[0089] Furthermore, based on the generated adjustment amounts, optimization operations are performed on each decoupled event point: for event points with non-zero position offsets, the position offset is superimposed on their original spatial coordinates to obtain the corrected spatial coordinates; for event points with non-zero time offsets, the position offset is superimposed on their original event occurrence light-temperature combined growth degree-day. Adding the time offset, we obtain the corrected event occurrence light-temperature integrated growth degree day; for event points with a confidence weight of less than one, we multiply the original event occurrence initial confidence by the confidence weight to obtain the weighted new confidence; this completes the fine adjustment of the spatial location, occurrence time and confidence of event points, while retaining all event point data and their correlations.

[0090] Specifically, the correction operation is bidirectional: on the one hand, it moderately pulls observation points that deviate significantly from the population model back towards the model's predicted values, suppressing potential noise; on the other hand, it preserves uncertainty information by reducing the confidence weight of these points rather than deleting them directly. For example, an event point whose detection time is slightly delayed due to brief leaf occlusion has its time slightly forward corrected; a point in a field that was indeed delayed due to wind damage, although its residual is large and its weight is reduced, still retains its data and location, and may be considered a valid but low-weighted sample in subsequent imputation. This parallel approach of correction and weighting achieves a fine balance between denoising, smoothing, and information preservation, making the optimized data more consistent with overall spatial statistical patterns without excessively erasing real local variations, thus improving the internal consistency and reliability of the dataset.

[0091] S7.4 Integrate all event points that have undergone position correction, time correction, and confidence weighting to generate an optimized tillering leaf age event sequence.

[0092] Furthermore, all event points that have undergone location correction, time correction, and confidence weighting are grouped according to their respective tiller main stem affiliation tags. Within each tiller group, all event points are sorted in ascending order according to the corrected light-temperature-growth degree days of the event occurrence. This sorted sequence of event points constitutes the leaf age event sequence of that tiller. The leaf age event sequences of all tillers in the field are collected to form the final, optimized tiller-level leaf age event sequence. Each event in this sequence carries corrected spatial coordinates, corrected light-temperature-growth degree days, and weighted confidence.

[0093] Specifically, the achievement lies in generating a high-fidelity tiller-level growth profile that integrates original detection data, individual plant structure decoupling, and population spatial optimization. The resulting optimized tiller-level leaf age event sequence not only records the number of leaves grown by each tiller but also precisely records the physiological time point at which each leaf emerged, its location, and the reliability of this record. For example, this sequence can clearly show how the leaf emergence interval changes after a certain period of low temperature for a particular tiller. This sequence forms the basis for all subsequent precision agronomic analyses (such as tiller dynamics, panicle differentiation prediction, and yield composition analysis). It extracts the information contained in time-series UAV images through layer-by-layer processing and optimization, refining it into a structured, semantic, and directly applicable digital crop growth phenotypic profile for scientific analysis and decision-making. This represents a leap from remote sensing imagery to a biologically meaningful digital twin of the growth process.

[0094] S8. Based on the optimized tillering-level leaf age event sequence, perform leaf age statistics and spatial mapping, and output a spatial distribution map of leaf age in rice fields.

[0095] S8.1 Sort the event points of each tiller in the optimized tillering leaf age event sequence according to the combined light and temperature growth degree per day.

[0096] Furthermore, the optimized tiller-level leaf age event sequence is read. For each independent tiller recorded in the sequence, all event points associated with that tiller are extracted. Each event point contains the corrected event occurrence light-temperature integrated growth degree-day attribute. For the event point list of this tiller, it is sorted in ascending order from smallest to largest according to the corrected event occurrence light-temperature integrated growth degree-day value, generating the event point time sequence of this tiller arranged in the order of physiological development, thus completing the temporal organization of the growth events within each tiller.

[0097] Specifically, the simple operation of sorting is based on a rigorously calibrated physiological timeline using the combined light and temperature growth days as a benchmark. After constructing the light and temperature timeline, detecting events, fractal decoupling, and spatial optimization, the calibrated combined light and temperature growth days are used for sorting. This ensures that the sorting results reflect the pure sequence of physiological development, completely eliminating the influence of variable environmental factors. For example, even if two tillers experience leaf emergence on different dates, as long as their combined light and temperature growth day sequence is correct, it indicates their inherent developmental rhythm. Sorting based on physiological time is the foundation for accurately counting leaf age and comparing developmental stages across tillers. It ensures that subsequent counting operations have consistent physiological meaning, rather than simply being timestamp counts.

[0098] S8.2 Count the number of event points for each tiller after sorting, define it as the current leaf age of the tiller, extract the spatial coordinates of each tiller and the corresponding current leaf age of the tiller to form a leaf age-spatial point dataset.

[0099] Furthermore, for each tiller event point time series that has been sorted by light and temperature combined growth degree days, the total number of event points contained in the series is directly counted, and this total number of event points is defined as the current leaf age value of the tiller. A representative spatial coordinate is extracted from the event point data of the tiller, for example, the average of the spatial coordinates of all event points of the tiller or the spatial coordinates of the latest event point, as the spatial location of the tiller. The spatial coordinates of each tiller are paired with the current leaf age value of its corresponding tiller to form a location-leaf age data pair. The set of location-leaf age data pairs of all tillers in the field constitutes the leaf age-spatial point dataset.

[0100] Specifically, the complex results of time-series event detection and optimization are condensed into two scalars with direct agronomical significance: the current leaf age of the tiller and the spatial coordinates of the tiller. A point dataset connecting plant phenotype and geographic space is constructed. The current leaf age of the tiller is defined by statistically counting the number of optimized event points. Each leaf age value is bound to a precise spatial coordinate of the tiller, thus grounding the abstract leaf age number in a concrete location in the field. For example, the resulting leaf age-spatial point dataset clearly indicates where a 5-leaf tiller exists and where a 3-leaf tiller exists in the field. This data organization seamlessly integrates agronomic parameters (leaf age) with the geographic information system, providing a standardized input format for subsequent spatial statistical analysis and visualization. S8.3. Perform spatial interpolation on the leaf age-spatial point dataset to generate a continuous field leaf age raster surface, and render and output a spatial distribution map of leaf age in the rice field.

[0101] Furthermore, using a leaf age-spatial point dataset as input, which contains a series of discrete spatial point coordinates and the current leaf age attribute value of the corresponding tiller at each point, a spatial interpolation algorithm, such as the inverse distance weighting method or Kriging interpolation, is employed. Using the spatial point coordinates as control points and the current leaf age of the tiller as the attribute to be interpolated, leaf age values ​​are estimated at regular grid points across the entire field area. Through the spatial interpolation algorithm, a continuous digital matrix covering the entire field is generated, with each grid cell possessing an estimated leaf age value—the field leaf age raster surface. The data values ​​on the field leaf age raster surface are mapped onto a preset color gradient table, for example, using dark green to represent high leaf age and light green to represent low leaf age. Geographic reference information such as field boundaries is overlaid, and a visually intuitive, colorful, and readable spatial distribution map of leaf age in the rice field is generated using a graphics rendering engine.

[0102] Specifically, by utilizing spatial interpolation techniques, discrete, tiller-level fine observations are upscaled to generate continuous, field-level population phenotypic distribution maps, achieving a transformation of decision-making information from point to surface. Existing technologies may only provide a list of leaf ages at sampling points, failing to intuitively display spatial variation patterns. This method, through spatial interpolation of the leaf age-spatial point dataset, generates a field leaf age raster surface that reveals spatial gradients, evenness differences, and anomalous areas that are not directly visible to the naked eye. For example, the distribution map can clearly show that the leaf age is generally higher on the north side of the field due to better sunlight, or that the leaf age is lower in low-lying areas due to waterlogging stress. This visualization not only presents results but also serves as a diagnostic tool. The rendered output spatial distribution map of leaf age in rice fields delivers all the precise information obtained from complex preliminary processing to agronomists in the most intuitive form, guiding variable fertilization and precise irrigation agronomic measures. This achieves a value loop from time-series UAV images to field management prescription maps, representing the final and most crucial step in technology serving decision-making.

[0103] This embodiment also provides a rice leaf age tracking system based on time-series UAV images, including: a preprocessing module that receives rice canopy orthophoto sequences, image acquisition timestamps, geographic location information, rice variety information, and time-series meteorological data corresponding to the geographic location corresponding to the rice canopy orthophoto sequences; The photothermal physiology module, based on rice variety information, obtains the growth baseline temperature and photoperiod sensitivity parameters of rice varieties, and uses time-series meteorological data and photoperiod sensitivity parameters to construct a photothermal comprehensive physiological time axis, converting image acquisition timestamps into photothermal comprehensive growth degree days. The standardization processing module resamples the rice canopy orthophoto sequence based on the photothermal integrated physiological time axis to extract a set of standardized time-series image blocks. The recognition module inputs a standardized set of time-series image blocks into the event detection neural network to obtain preliminary leaf emergence event detection results. The optimization module decouples the initial leaf emergence event detection results into single-plant events based on the fractal structure rules of rice leaf order, and optimizes the spatial consistency of the decoupled events based on the field population growth wave propagation model to obtain the optimized tillering-level leaf age event sequence. The statistics module performs leaf age statistics and spatial mapping based on the optimized tillering-level leaf age event sequence, and outputs a spatial distribution map of leaf age in rice fields.

[0104] This embodiment also provides a computer device applicable to the rice leaf age tracking method based on time-series UAV images, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the rice leaf age tracking method based on time-series UAV images proposed in the above embodiment.

[0105] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0106] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the rice leaf age tracking method based on time-series UAV images as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0107] In summary, this invention constructs a physiological timeline that integrates photothermal factors by receiving orthophoto sequences of rice canopy collected by UAVs and their associated multi-source information. The time-series images are uniformly mapped to a daily scale reflecting the comprehensive photothermal growth degree of real growth and development. Based on this, the image sequences are resampled and standardized on the physiological timescale. Leaf biological events are identified from the standardized images using an event detection neural network. Events at the single-plant scale are decoupled and assigned by combining the inherent fractal structure rules of rice leaf order. The spatiotemporal consistency of the decoupled events is optimized using a field population growth wave propagation model. Based on the optimized event sequence, a highly spatially consistent rice field leaf age distribution map is generated, achieving high-precision and robust tracking of rice leaf age dynamics.

[0108] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for tracking rice leaf age based on time-series UAV images, characterized in that: Receive rice canopy orthophoto sequences, along with the corresponding image acquisition timestamps, geographic location information, rice variety information, and time-series meteorological data for each geographic location. Based on rice variety information, the growth baseline temperature and photoperiod sensitivity parameters of rice varieties are obtained. A photo-temperature integrated physiological time axis is constructed using time-series meteorological data and photoperiod sensitivity parameters, and the image acquisition timestamp is converted into a photo-temperature integrated growth degree-day. Based on the photothermal integrated physiological time axis, the rice canopy orthophoto image sequence was resampled to extract a standardized time-series image block set. The standardized time-series image block set is input into the event detection neural network to obtain preliminary leaf emergence event detection results; Based on the fractal structure rules of rice leaf order, the preliminary leaf emergence event detection results are decoupled into single-plant events. Based on the field population growth wave propagation model, the spatial consistency of the decoupled events is optimized to obtain the optimized tillering-level leaf age event sequence. Based on the optimized tillering-level leaf age event sequence, leaf age statistics and spatial mapping are performed to output a spatial distribution map of leaf age in rice fields.

2. The rice leaf age tracking method based on time-series UAV images as described in claim 1, characterized in that: Receive rice canopy orthophoto sequences, along with the corresponding image acquisition timestamps, geographic location information, rice variety information, and time-series meteorological data for each geographic location. This includes the following steps: The system receives rice canopy orthophoto sequences collected and transmitted by drones via a data interface, and obtains the image acquisition timestamps that are time-synchronized with the rice canopy orthophoto sequences by parsing the drone flight logs. By reading the records from the drone's positioning device, we can obtain the geographical location information that corresponds to the rice canopy orthophoto sequence and has been spatially registered. By querying the field management database, we can obtain rice variety information corresponding to the fields photographed by the rice canopy orthophoto sequence. We can also access the meteorological data service interface to obtain time-series meteorological data that matches the area specified by the geographical location information.

3. The rice leaf age tracking method based on time-series UAV images as described in claim 2, characterized in that: Based on rice variety information, the growth baseline temperature and photoperiod sensitivity parameters of rice varieties are obtained, including the following steps: Based on rice variety information, we query the rice variety trait database and obtain the growth baseline temperature of rice varieties from the rice variety trait database. Photoperiod sensitivity parameters of rice varieties were obtained from a rice variety trait database.

4. The rice leaf age tracking method based on time-series UAV images as described in claim 3, characterized in that: A photothermal integrated physiological timeline was constructed using time-series meteorological data and photoperiod sensitivity parameters. Image acquisition timestamps were converted into photothermal integrated growth days, including the following steps: Daily average temperature and sunshine duration data are extracted from time-series meteorological data. The daily average temperature is compared with the growth baseline temperature of rice varieties to evaluate the daily effective accumulated temperature that reflects the temperature-driven effect of the day. By combining sunshine duration data with photoperiod sensitivity parameters of rice varieties, the daily photoperiod influence factor of the degree of inhibition was quantified. The daily effective accumulated temperature is nonlinearly modulated by the daily photoperiod influence factor to generate a daily photothermal modulated effective accumulated temperature that integrates both photothermal and photothermal effects. The effective accumulated temperature of daily photothermal modulation after sowing is continuously summed to form a photothermal comprehensive physiological time axis based on the accumulated physiological development amount; The image acquisition timestamp is located and mapped on the photothermal integrated physiological time axis to obtain the photothermal integrated growth day corresponding to the image acquisition time.

5. The rice leaf age tracking method based on time-series UAV images as described in claim 4, characterized in that: Resampling of rice canopy orthophoto sequences based on a photothermal integrated physiological timeline yields a standardized time-series image block set, including the following steps: Target physiological sampling points are set at equal intervals on the photothermal integrated physiological time axis. Based on the position of the target physiological sampling points on the photothermal integrated physiological time axis, the corresponding image frames are matched from the rice canopy orthophoto sequence. For the corresponding matching image frame, image blocks are slidably extracted using a grid window of fixed size, and radiometric normalization and size scaling operations are performed on all extracted image blocks. All image blocks that have completed the standardization process will be integrated into a standardized temporal image block set in temporal and spatial order.

6. The rice leaf age tracking method based on time-series UAV images as described in claim 5, characterized in that: The standardized time-series image patch set is input into the event detection neural network to obtain preliminary leaf emergence event detection results, including the following steps: The standardized temporal image patch set is input into the pre-trained event detection neural network, which performs spatiotemporal feature extraction and pattern analysis on the standardized temporal image patch set. The event detection neural network outputs preliminary leaf emergence event detection results, including the spatial coordinates of the event occurrence, the combined light and temperature growth period of the event occurrence, and the initial confidence level of the event occurrence.

7. The rice leaf age tracking method based on time-series UAV images as described in claim 6, characterized in that: Based on the fractal structure rules of rice leaf order, the preliminary leaf emergence event detection results are decoupled for single-plant events, including the following steps: Based on the preliminary leaf emergence event detection results of spatial location clustering, a set of candidate event points for a single plant is generated. For a single plant candidate event point set, the fractal dimension, branch topology, spatial geometry and time sequence are matched and verified with the preset rice leaf order fractal structure rule template, and the multidimensional matching verification results are output. Based on the comprehensive multidimensional matching verification results, the candidate event points of individual plants are assigned to specific tillering main stems and outliers are identified, thus completing the decoupling of individual plant events.

8. The rice leaf age tracking method based on time-series UAV images as described in claim 7, characterized in that: Based on the field population growth wave propagation model, spatial consistency optimization is performed on the decoupled events to obtain the optimized tillering-level leaf age event sequence, including the following steps: By using the spatial coordinates of the event point after decoupling the single-plant event and the corresponding light and temperature integrated growth degree day, a field population growth wave propagation model characterizing the propagation law of the event in the field is fitted through spatial autocorrelation analysis. Based on the field population growth wave propagation model, the spatial residual is calculated to evaluate the spatiotemporal rationality of each decoupled event point, and to generate the adjustment amount of position offset, time offset and confidence weight. Based on the adjustments to the location offset, time offset, and confidence weight, position correction, time correction, and confidence weighting are performed on the decoupled event points. By integrating all event points that have undergone location correction, time correction, and confidence weighting, an optimized tillering leaf age event sequence is generated.

9. The rice leaf age tracking method based on time-series UAV images as described in claim 8, characterized in that: Based on the optimized tillering-level leaf age event sequence, leaf age statistics and spatial mapping are performed to output a spatial distribution map of leaf age in rice fields, including the following steps: The event points of each tiller in the optimized tillering leaf age event sequence are sorted by the combined light and temperature growth degree per day; The number of event points for each tiller after sorting is defined as the current leaf age of the tiller. The spatial coordinates of each tiller and the corresponding current leaf age of the tiller are extracted to form a leaf age-spatial point dataset. Spatial interpolation is performed on the leaf age-spatial point dataset to generate a continuous field leaf age raster surface, and then a rendering output is performed to show the spatial distribution map of leaf age in the rice field.

10. A rice leaf age tracking system based on time-series UAV images, based on the rice leaf age tracking method based on time-series UAV images according to any one of claims 1 to 9, characterized in that: It includes a preprocessing module that receives rice canopy orthophoto sequences, image acquisition timestamps, geographic location information, rice variety information, and time-series meteorological data corresponding to the geographic location corresponding to the rice canopy orthophoto sequences. The photothermal physiology module, based on rice variety information, obtains the growth baseline temperature and photoperiod sensitivity parameters of rice varieties, and uses time-series meteorological data and photoperiod sensitivity parameters to construct a photothermal comprehensive physiological time axis, converting image acquisition timestamps into photothermal comprehensive growth degree days. The standardization processing module resamples the rice canopy orthophoto sequence based on the photothermal integrated physiological time axis to extract a set of standardized time-series image blocks. The recognition module inputs a standardized set of time-series image blocks into the event detection neural network to obtain preliminary leaf emergence event detection results. The optimization module decouples the initial leaf emergence event detection results into single-plant events based on the fractal structure rules of rice leaf order, and optimizes the spatial consistency of the decoupled events based on the field population growth wave propagation model to obtain the optimized tillering-level leaf age event sequence. The statistics module performs leaf age statistics and spatial mapping based on the optimized tillering-level leaf age event sequence, and outputs a spatial distribution map of leaf age in rice fields.

Images