Method and system for quantifying autumn phenology under forest based on infrared camera

Abstract

The invention discloses an under-forest autumn phenology quantification method and system based on an infrared camera, and belongs to the technical field of ecological detection, and the method comprises the steps: S1, under-forest vegetation image collection: employing the infrared camera to carry out the under-forest vegetation image shooting and collection at a phenology stage; s2, obtaining a daily-scale vegetation optimal image: carrying out image feature obtaining through an under-forest vegetation daily optimal image selection algorithm fusing multi-feature clustering and hierarchical screening; s3, checking the consistency of the meteorological data of the under-forest infrared camera and the data of the canopy meteorological station; s4, vegetation phenology dynamic extraction: adopting a dynamic quantile method to obtain a time sequence change curve, and determining a change rate curvature and a remote sensing phenology period of a dynamic threshold value based on a vegetation index; according to the method, forest under-forest vegetation dynamic monitoring is carried out by combining dynamic threshold identification and infrared camera imaging, redundant or low-quality images are effectively eliminated, the vegetation index extraction efficiency is improved, and the phenological period judgment precision is improved; and meanwhile, the algorithm is lightweight and efficient.

Description

Technical Field

[0001] The present invention belongs to the technical field of ecological monitoring, and in particular relates to a method and system for quantifying autumn phenology in forest understory based on infrared cameras. Background Art

[0002] The current vegetation phenology monitoring technology system mainly includes three methods: ground observation, controlled experiments, and satellite and UAV remote sensing monitoring. Different technical means have advantages in the analysis of temporal and spatial scales and ecological processes.

[0003] Ground-based observations are a fundamental method for obtaining phenological data, dating back to the 9th century AD with records of cherry blossom blooms in Kyoto, Japan. In the early 20th century, with the systematic development of phenological research, standardized observation networks were gradually established globally, such as the US National Phenological Observatory, the China Phenological Observatory, and the European Phenological Observatory. Eddy covariance (EC) technology enables continuous quantitative characterization of ecosystem phenology through high-frequency flux monitoring. Its data support has formed international monitoring networks such as AmeriFlux and ChinaFlux.

[0004] Traditional ground-based observations are limited by their sparse spatial distribution, limited coverage, and varying standards for manual recording, making them inadequate for studying ecosystem-scale phenological dynamics. Traditional autumn phenological monitoring relies on field observers recording tree senescence. This method is more suited to monitoring individual species rather than entire ecosystems and is therefore not applicable to monitoring entire ecosystems. While ground-based observations suffer from extrapolation uncertainties due to spatial heterogeneity, their high precision provides an indispensable benchmark for validating remote sensing data.

[0005] Temperature, light, and water manipulation experiments are used to analyze the climate response mechanisms of vegetation. Most manipulation experiments are conducted on only a few species and usually only last for a few years. Such short-term warming experiments may not be sufficient to induce adaptive responses or to fully understand the long-term responses of plants to environmental changes. Secondly, these manipulation experiments are usually conducted on seedlings or young trees, and rarely on mature trees. Third, the phenological responses of plants to experimental conditions and natural conditions may be significantly different. However, there are relatively few experimental studies on autumn phenology. Despite the above limitations, manipulation experiments have certain advantages over natural observations, namely that certain conditions can be kept constant (for example, water, light, and nutrients can all be controlled), which makes them ideal for hypothesis testing and can significantly deepen our understanding of phenological responses to climate change.

[0006] The low spatial resolution of satellite data (ranging from 30 meters to several kilometers) makes it particularly difficult to extract and interpret phenological dates in communities with mixed species at different phenological stages. Similarly, in deciduous forests, spring green-up typically occurs first in the ground layer, meaning that spring green-up dates determined by remote sensing methods may reflect the green-up of herbs and shrubs rather than the dominant trees in these forests, which tend to green up later. While satellite remote sensing has broadened macroscale phenological research, it is affected by multiple factors, including poor observation conditions, low spatial and temporal resolution, sensor calibration, and bidirectional reflectance distribution function (BRDF) effects, which can reduce phenological accuracy. Furthermore, vegetation index (VI) time series are constructed using the maximum value composite (MVC) technique. MVC selects the maximum NDVI value within a specific composite period at each pixel and aims to minimize atmospheric effects, including residual clouds. The composite process typically includes cloud screening and data quality checks. MVC works well on near-Lambertian surfaces, where the main sources of pixel variation are atmospheric pollution and path length. However, due to the anisotropy of vegetation canopies and changing atmospheric conditions, MVC becomes inconsistent and difficult to predict. MVC removes noise (such as clouds and shadows) by selecting the maximum value of a vegetation index (such as NDVI) within a time window. This method only retains the "best state" value, which may obscure the full picture of vegetation dynamics. Short-term environmental disturbances (such as occasional clear weather) may cause MVC to select abnormally high values, interfering with the judgment of the true phenological stage.

[0007] Trail cameras, used to monitor animal activity, have the potential to monitor understory ecosystems. These low-cost digital cameras are designed for long-term, unsupervised outdoor use and are equipped with ambient light sensors that automatically convert images to black and white. Unlike traditional digital cameras, trail cameras capture additional information such as temperature and photoperiod through their sensors. Preliminary studies have shown that it is feasible to track changes in greenness using red, green, and blue (R, G, B) values. For example, vegetation indices such as green color coordinates (GCC), red color coordinates (RCC), and blue color coordinates (BCC), as well as the excess greenness index (ExG), the normalized green-red difference index (GRVI), and the visible atmospheric resistance index (VARI) have been employed. Although GCC is the most commonly used vegetation index in digital cameras, it is not optimal for detecting the end of the growing season (EOS) for understory plants with pronounced leaf angles. Camera images can also vary depending on automatic white balance (AWB) settings and external environmental factors. Therefore, the selection of appropriate vegetation indices and preprocessing methods for extracting understory EOS requires reconsideration.

[0008] The phenology detection methods in the prior art mainly have the following problems:

[0009] (1) Limitations of traditional phenological monitoring methods: They rely on manual timed observations or single sensor data, making it difficult to achieve continuous and automated monitoring of the dynamic changes of understory vegetation. They are also susceptible to environmental interference such as light and weather, resulting in insufficient data integrity and accuracy. Existing satellite remote sensing phenological models mostly judge phenological periods based on canopy scales or fixed thresholds, and are unable to adapt to the complex lighting conditions and dynamic growth characteristics of understory vegetation (such as local differences in leaf discoloration and leaf fall), resulting in large errors in phenological period determination.

[0010] (2) Redundancy and inefficient utilization of image data: There are a large number of invalid or low-quality data (such as overexposure, blur, and interference from non-vegetation areas) in the massive images generated by continuous shooting of infrared cameras. Traditional methods lack an effective daily-scale optimal image screening mechanism, resulting in low efficiency and poor reliability of vegetation index extraction.

[0011] (3) Impact of insufficient consistency of meteorological data: There are significant differences between the understory microenvironment and canopy meteorological station data (such as temperature). Direct use without verification may lead to deviations in phenological model parameters and affect the credibility of the results.

[0012] (4) Lack of dynamic threshold self-adaptation capability: Existing methods use global fixed quantile thresholds (such as maximum or minimum values) or simple logical functions to fit vegetation index curves, which makes it difficult to capture the dynamic fluctuation characteristics of the autumn phenological period (such as the extension of the peak period and the early dormancy period), resulting in inaccurate identification of key phenological nodes (such as the beginning of color change and the end of leaf fall). Summary of the Invention

[0013] In response to the problems mentioned in the background technology, the present invention proposes a method and system for quantifying autumn phenology in forest understory based on infrared cameras, which effectively eliminates redundant or low-quality images, reduces manual intervention, improves the efficiency of vegetation index extraction, and improves the accuracy of phenological period determination; at the same time, through the design of a lightweight algorithm based on K-means clustering and SG filter, combined with dynamic window quantile calculation, the algorithm is lightweight and efficient.

[0014] Technical solution: In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0015] A method for quantifying autumn phenology in forest understory based on infrared cameras includes the following steps:

[0016] S1: Understory vegetation image collection: Use an infrared camera to capture and collect understory vegetation images during the phenological stage;

[0017] S2: Daily-scale vegetation optimal image acquisition: Image feature acquisition is performed using the daily optimal image selection algorithm for understory vegetation that integrates multi-feature clustering and hierarchical screening;

[0018] S3: consistency check between understory infrared camera meteorological data and canopy meteorological station data;

[0019] S4: Dynamic extraction of vegetation phenology: The dynamic quantile method is used to obtain the time series change curve, and the remote sensing phenological period of the change rate curvature and dynamic threshold is determined based on the vegetation index.

[0020] As a preference, the specific process of S2 is:

[0021] S21: Image screening and clustering: K-means clustering algorithm and elbow rule are used for image selection;

[0022] S22: Daily-scale optimal image: Determine the image in the cluster that best reflects the greenness of vegetation.

[0023] As a preference, the specific process of S21 is:

[0024] S211: Filter images whose green-red vegetation index and saturation are greater than or equal to 0;

[0025] S212: Apply K-means clustering and the elbow rule to determine the optimal number of clusters based on the sum of squared distances and the Calinski-Harabasz score.

[0026] Preferably, in S212, the specific content of K-means clustering is:

[0027] Step 1, select K cluster centers;

[0028] Initialize K centroids c1, c2, c3, ..., c k , each centroid is a 7-dimensional vector, corresponding to 7 initial values of photo features, specifically:

[0029] c k =[c k1 , c k2 ,...,c k7 ], k = 1, 2, ..., K

[0030] Among them, c k represents the centroid vector;

[0031] Step 2, calculate distance and assign clusters;

[0032] For each photo's 7-dimensional feature vector X i =[x i1 ,x i2 ,…,x i7 ],

[0033] Among them, x i1 Indicates white balance; x i2 Indicates the direction of light; x i3 Indicates color temperature; x i4 Indicates contrast; x i5 Indicates grayscale; x i6 Indicates saturation; x i7 Indicates brightness;

[0034] Calculate the Euclidean distance d to all centroids, and then assign the photo to the cluster to which the nearest centroid belongs, specifically:

[0035]

[0036] Among them, X i is a 7-dimensional feature vector; c k is the centroid vector; x ij is the jth eigenvalue of the i-th photo; c kj represents the jth eigenvalue of the kth cluster center;

[0037] Step 3, update the centroid;

[0038] Recalculate the centroid of each cluster and take the mean of the feature values of all photos in the cluster;

[0039] The centroid update formula of the kth cluster for:

[0040]

[0041] Among them, C k is the photo set of the kth cluster; |C k | is the number of photos in the cluster; j represents the feature dimension; x ij is the jth eigenvalue of the i-th photo; x i is the i-th photo;

[0042] Step 4, iterate until convergence: Repeat steps 2 and 3 until the centroid position no longer changes or the maximum number of iterations is reached, and finally determine the category of each photo.

[0043] Preferably, in S212, the specific content of the elbow rule is:

[0044] The elbow rule determines the optimal number of clusters by analyzing the relationship between the evaluation index of the clustering result and the number of clusters; the SSE evaluation index is used for evaluation to adaptively select the best cluster center. The calculation formula of the SSE evaluation index is:

[0045]

[0046] Among them, K is the cluster number; C k is the photo set of the kth cluster; x ij is the jth eigenvalue of the i-th photo; c kj represents the j-th eigenvalue of the k-th cluster center.

[0047] Preferably, the specific process of S22 is:

[0048] S221: Determine the cluster that best reflects the greenness of vegetation by comparing the images in each cluster;

[0049] S222: Finally, the most suitable photo for each day is selected within the best cluster; if not, a photo in the next cluster is selected.

[0050] As a preference, the specific process of S4 is:

[0051] S41: Use SG filter to smooth the vegetation index time series;

[0052] S42: dynamic quantile thresholds were used to extend the peak growth period and dormancy period to process the original vegetation index;

[0053] S43: curve fitting using a four-parameter logistic function to simulate the dynamic fluctuations of vegetation greenness;

[0054] S44: Determine the dynamic quantiles of peak and dormant periods using Akaike information quantity;

[0055] S45: The remote sensing phenological period of the rate of change curvature and dynamic threshold is determined using the first-order logistic derivative, the third-order logistic derivative, and the maximum smoothed vegetation index value based on the threshold.

[0056] As an example, in S42, the dynamic quantile threshold Q t The calculation formula for (q) is:

[0057] Q t (q) = Quantile({X1, ..., X t},q),

[0058] Among them, q represents the quantile, X t represents the vegetation index time series, and Quantile() represents the quantile function. Preferably, in S44, the calculation formula of Akaike information content AIC is:

[0059] AIC=-2lnL+2k

[0060] Where k represents the number of parameters in the logistic model; L is the maximum likelihood estimate of the model on the vegetation index.

[0061] A forest understory autumn phenology quantification system based on an infrared camera, according to any of the above-mentioned forest understory autumn phenology quantification methods based on an infrared camera, includes an image acquisition and dynamic monitoring module, a daily scale optimal image selection module, a meteorological data consistency verification module, and a vegetation phenology dynamic extraction module.

[0062] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0063] (1) The present invention can effectively eliminate redundant or low-quality images (such as overexposure and blur), reduce manual intervention, and improve the efficiency of vegetation index extraction; by screening the best intra-cluster images, it can enhance the accuracy of vegetation greenness representation and avoid the interference of illumination changes on image analysis.

[0064] (2) The present invention can dynamically adapt to the complex environment under the forest and improve the accuracy of phenological period determination. It uses a dynamic quantile threshold to extend the growth peak period and dormancy period to process the original vegetation index and fit the vegetation index curve with a four-parameter logistic function, optimizes the quantile selection in combination with the Akaike information criterion, and uses the first-order derivative, third-order derivative equation, 30% maximum vegetation index and 50% maximum vegetation index to extract the inflection point of the vegetation index.

[0065] The dynamic quantile threshold can adapt to the vegetation growth cycle of different vegetation types and climatic conditions (such as extending the peak period and advancing the dormancy period), avoiding the misjudgment of phenological periods caused by fixed thresholds.

[0066] (3) The present invention enhances the credibility of meteorological data, eliminates the influence of microenvironment, corrects the interference of forest microenvironment (such as temperature gradient and humidity difference) on phenological model, and improves the reliability of model parameters through synchronous matching and consistency verification of meteorological data of forest infrared cameras and canopy meteorological station data; ensures that the correlation analysis of vegetation dynamic changes and meteorological conditions has a scientific basis, and avoids distorted conclusions caused by data deviation.

[0067] (4) The present invention realizes lightweight and efficient algorithm through the design of lightweight algorithm based on K-means clustering and SG filter, combined with dynamic window quantile calculation, and is suitable for large-scale deployment; at the same time, the algorithm has low complexity and can run in real time in edge computing devices (such as infrared camera embedded systems), supporting long-term monitoring of multiple nodes in the forest; data storage and transmission requirements are reduced by more than 50%, significantly reducing hardware costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flow chart of the method for quantifying autumn phenology in forest understory based on infrared cameras of the present invention;

[0069] Figure 2Schematic diagram of the research area of the present invention; wherein (a) is a schematic diagram of the position of the UVL4-CN infrared camera; (b) is a schematic diagram of the field of view captured by the camera;

[0070] Figure 3 Schematic diagram of cluster analysis for determining the optimal grouping and visualizing the results of the present invention; wherein (a) is a schematic diagram showing the optimal number of clusters using the elbow rule; (b) is a schematic diagram showing the visualization of cluster grouping;

[0071] Figure 4 Schematic diagram of the effect of the dynamic quantile of vegetation index on the logistic fitting of vegetation index shape according to the present invention; wherein, (a) is GRVI; (b) is VARI; (c) is TGI; (d) is ExG; (e) is WFI; (f) is RGRI; (g) is GCC; (h) is IKaw; (i) is the legend;

[0072] Figure 5 Schematic diagram of the Akaike Information Criterion (AIC) of the Logistic Fit Vegetation Index Shape of the Vegetation Index Dynamic Quantiles at the Beginning of the Maturity Period and the End of the Dormancy Period of the present invention;

[0073] Figure 6 Schematic diagram of the present invention using GRVI to obtain EOS data from the UVL4 forest camera logistic function fitting and phenological extraction, wherein (a) is a schematic diagram of the first-order derivative method; (b) is a schematic diagram of the third-order derivative method; (c) is a schematic diagram of the threshold-based method (30% and 50% thresholds);

[0074] Figure 7 Schematic diagram comparing understory vegetation and open environment conditions (photoperiod and temperature analysis) of the present invention; wherein (a) is a photoperiod comparison diagram; (b) is an average temperature comparison diagram; (c) is a maximum temperature comparison diagram; and (d) is a minimum temperature comparison diagram;

[0075] Figure 8 Schematic diagram of the comparison of the EOS differences between the canopy and understory based on the CDD model and TPM of the present invention; wherein, (a) is the EOS hysteresis of the canopy and understory based on the CDD model; (b) is the EOS hysteresis of the canopy and understory based on the TPM. DETAILED DESCRIPTION

[0076] The present invention will be further illustrated below with reference to specific examples. The examples are implemented based on the technical solutions of the present invention. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0077] The infrared camera-based forest understory autumn phenology quantification method provided in this embodiment performs dynamic monitoring of forest understory vegetation by combining dynamic threshold recognition with UVL4-CN infrared camera imaging.

[0078] This embodiment takes Yuexi County, Anqing City, Anhui Province, China (30°57′-31°06′N, 116°02′-116°11′E, located in Yaoluoping National Nature Reserve) as the study area. This area is less affected by human activities and has good lighting conditions. The vegetation in the reserve is mainly north subtropical deciduous-evergreen broad-leaved mixed forest, and the main protection target is the native forest ecosystem in the Dabie Mountains. The UVL4-CN infrared camera was placed on a large tree 1 meter above the ground with a vertical angle of 30°. Its field of view includes small trees, seedlings and herbaceous plants. The data was recorded during the growing season (from the 180th to the 333rd day of 2022). At this time, summer begins and plant growth enters a critical period. The ecological functions of understory vegetation (such as water and soil conservation, soil retention and nutrient cycling) become more significant. According to vegetation characteristics, understory vegetation is defined as plants with a height of less than 5 meters. The specific steps are as follows:

[0079] S1: Understory vegetation image acquisition: Use an infrared camera to capture images of understory vegetation during the phenological stage;

[0080] Infrared cameras are used to precisely track phenological stages. Timed recording (e.g., at fixed times daily) generates continuous images, capturing key phenological events such as bud break, leaf expansion, flowering, maturity, and leaf fall. Automated recording avoids frequent entry into the forest floor, disrupting the microenvironment, making it particularly suitable for long-term ecological positioning observations. Infrared cameras can directly measure vegetation surface temperature through thermal imaging, reflecting transpiration, water stress, or photosynthetic activity (e.g., increased leaf temperature may indicate drought stress). Built-in light sensors record the understory light environment and quantify the duration of sunlight for understory vegetation growth. Simultaneous monitoring of temperature and humidity helps analyze the impact of water conditions on plant phenology. Traditional methods require multiple devices to measure time, which is prone to errors due to temporal and spatial variations in microenvironments. Infrared cameras' integrated sensors ensure temporal and spatial consistency of data. Infrared cameras are typically designed for low power consumption and, when powered by solar energy, can operate continuously for months in remote forests. Their waterproof and dustproof design adapts to the high humidity and humus-rich understory environment, ensuring device durability.

[0081] S2: Daily-scale vegetation optimal image acquisition: Image feature acquisition is performed using the daily optimal image selection algorithm for understory vegetation that integrates multi-feature clustering and hierarchical screening;

[0082] Traditional remote sensing data acquisition often uses images acquired at fixed times, such as 8:00 or 16:00 (just after sunrise or sunset), for analysis. This method fails to consider the impact of image parameters on vegetation greenness monitoring. Therefore, this paper proposes an algorithm for selecting the optimal daily image of understory vegetation that integrates multi-feature clustering and hierarchical screening. Specifically:

[0083] S21: Image screening and clustering: K-means clustering algorithm and elbow rule are used for image selection;

[0084] The image quality produced by the camera is affected by factors such as automatic white balance (AWB), light direction (LD), correlated color temperature (CCT), contrast (CTR), grayscale (GL), saturation (Sat) and brightness (Bri), as shown in Table 1.

[0085] Table 1 Main factors affecting image quality and a brief description of the impact of each factor on the image

[0086]

[0087]

[0088] In this embodiment, image selection follows the following steps:

[0089] S211: Filter images whose green-red vegetation index and saturation are greater than or equal to 0;

[0090] First, use the Green-Red Vegetation Index (CRVI) and saturation to eliminate photos with CRVI and Sat below 0, and make sure they are taken during daytime.

[0091] S212: Apply K-means clustering and elbow rule, and select the appropriate K based on the sum of squared distances (SSD) and Calinski-Harabasz score (CH);

[0092] The K-means clustering algorithm is used to group images into K clusters based on seven influencing factors, minimizing the intra-cluster sum of squares. This method groups images based on automatic white balance (AWB), light direction (LD), correlated color temperature (CCT), contrast (CTR), saturation (Sat), brightness (Bri), and grayscale (GL). The optimal number of clusters (K) is determined using the elbow rule, which clusters images with similar performance characteristics by reducing intra-cluster error.

[0093] K-means is a commonly used unsupervised learning algorithm for partitioning a dataset into K clusters. Its goal is to minimize the sum of the squared distances between each cluster's data points and the cluster center. K-means is widely used in fields such as image processing, market analysis, text mining, and bioinformatics. The basic idea behind this method is:

[0094] Step 1: Randomly select K cluster centers;

[0095] Randomly initialize K centroids c1, c2, c3, ..., c k , each centroid is a 7-dimensional vector, corresponding to the initial values of 7 photo features (white balance, lighting direction, color temperature, contrast, grayscale, saturation and brightness), specifically:

[0096] c k =[c k1 , c k2 ,...,c k7 ], k = 1, 2, ..., K

[0097] Among them, c k is the centroid vector.

[0098] Step 2, calculate distance and assign clusters;

[0099] For each photo's 7-dimensional feature vector X i =[x i1 ,x i2 ,…,x i7 ], where x i1 is white balance; x i2 is the direction of light; x i3 is the color temperature; x i4 is the contrast; x i5 is grayscale; x i6 is saturation; x i7 The brightness is calculated. The Euclidean distance d to all centroids is calculated and the photo is assigned to the cluster with the nearest centroid. Specifically:

[0100]

[0101] Among them, X i is a 7-dimensional feature vector; c k is the centroid vector; x ij is the jth eigenvalue of the i-th photo; c kj Represents the jth eigenvalue of the kth cluster center; j represents the 7 characteristic dimensions of the photo. Cluter() represents the photo x i The number of the cluster to which it is finally assigned; Indicates returning the centroid number k that achieves the minimum distance.

[0102] Step 3, update the centroid: recalculate the centroid of each cluster and take the mean of all the photo feature values in the cluster. For example, the centroid update formula for the kth cluster is for:

[0103]

[0104] Among them, C k is the photo set of the kth cluster; |C k | is the number of photos in the cluster; j represents the feature dimension; x ij is the jth eigenvalue of the i-th photo; x i is the i-th photo.

[0105] Step 4, iterate until convergence: Repeat steps 2 and 3 until the centroid position no longer changes or the maximum number of iterations is reached, and finally determine the category of each photo.

[0106] The elbow rule is a classic clustering algorithm evaluation method used to determine the optimal number of clusters in a clustering algorithm. In the K-mean clustering algorithm, selecting the number of clusters that matches the sample classification is crucial for effective data analysis. The elbow rule determines the optimal number of clusters by analyzing the relationship between the evaluation index of the clustering results and the number of clusters. Its core is to calculate the tightness of the clustering results under different numbers of clusters, usually based on the clustering error between the data points and their corresponding cluster centers, that is, the SSE evaluation index. It can adaptively select the best cluster center. The point in the curve where the rate of decline of SSE slows down significantly (shaped like an elbow bend) is the optimal K (the goal is to minimize SSE). This point indicates that the marginal benefit of increasing the number of clusters on model improvement is reduced. The SSE evaluation index can be expressed as:

[0107]

[0108] Among them, K represents the cluster; C k represents the photo set of the kth cluster; x ij represents the jth eigenvalue of the i-th photo; c kj represents the jth eigenvalue of the kth cluster center; j represents the 7 feature dimensions of the photo, x i represents the i-th photo.

[0109] S22: Daily Optimal Image: Determine the image in the cluster that best reflects the greenness of vegetation, specifically:

[0110] S221: Determine the cluster that best reflects the greenness of vegetation by comparing the images in each cluster, specifically including:

[0111] (1) Prioritize images with high Tenengrad clarity within the cluster;

[0112]

[0113]

[0114] Among them, G x (i, j) and G y(i, j) are the horizontal and vertical gradients of the Sobel operator at pixel (i, j) in the photo. I is the grayscale matrix of the photo, and * represents the convolution operation.

[0115] (2) Center evaluation based on parameter features within the image cluster;

[0116] The main features that affect image quality are adopted. The requirements for intra-cluster images are: grayscale around 110; the smaller the skewness coefficient (white balance), the better; the contrast is close to 0.5 (in forest understory phenology monitoring, the recommended contrast range is 50%-70%); the greater the saturation; and the brightness is around 0.5.

[0117] Ensure that the dynamic changes in vegetation greenness of the selected clusters are dominated by phenological stages rather than short-term environmental noise fluctuations.

[0118] (3) Temporal distribution of images within a cluster;

[0119] Check the continuity of the time distribution of images within candidate clusters. If the images in cluster A are concentrated in a fixed time period of sunny weather (such as 8:00-11:00 or 13:00-16:00), while the images in cluster B are temporally dispersed and contain rainy weather data, then cluster A is preferred. Combined with meteorological data, image clusters with periods of severe light fluctuations are eliminated.

[0120] S222: Finally, the most suitable photo for each day is selected within the best cluster; if not, a photo in the next cluster is selected.

[0121] Specific image selection requirements: select images with CRVI>0 and Sat>0, and ensure that the images are taken during the day; select the appropriate K based on the maximum f″(SSD)||minimum f′(CH) (the maximum value of the second-order derivative of SSD and the minimum value of the first-order derivative of CH); compare the clusters and select the appropriate cluster.

[0122] In this example, based on the algorithm process for selecting the optimal daily image of understory vegetation, which integrates multi-feature clustering and hierarchical screening, a K value of 3 was determined, and Cluster 1 was selected as the optimal. This cluster has a maximum saturation of 0.20 and a minimum correlated color temperature of 23.15, making it suitable for tracking changes in vegetation greenness. However, external factors such as precipitation, fog, and light can introduce image noise. Combined with the K-means clustering results of the photo parameters, Cluster 3 is the suboptimal choice. If no image is available in Cluster 1 on a given day, an image is selected from Cluster 3.

[0123] S3: consistency check between understory infrared camera meteorological data and canopy meteorological station data;

[0124] The present invention compared the theoretical photoperiod and the understory photoperiod. The results showed that the average theoretical photoperiod was 12.25 hours from the beginning of July to the end of November. In contrast, the understory photoperiod was slightly shorter, ranging from 1.74 to 2.02 hours on average. As the season transitioned from summer to autumn, the difference between the theoretical photoperiod in the open environment and the actual photoperiod in the forest understory became apparent. Regarding the forest understory temperature recorded by the UVL4 understory camera, the average daily temperature (Tunderstory), daytime temperature (Tunderstory max) and nighttime temperature (Tunderstory min) at 1 meter above the ground were in the ranges of 18.09°C-18.99°C, 21.56°C-22.56°C and 16.46°C-17.42°C, respectively (see Figure 7 b to Figure 7 d) in the above.

[0125] The understory temperature was slightly higher than the mean and minimum temperatures in the open environment (Tunderstory and Tunderstorymin were 18.09℃-19.99℃ and 21.56℃-22.56℃, respectively, and Topen-air and Topen-air min were 15.9℃-16.96℃ and 10.65℃-11.81℃, respectively), but was cooler than the open environment during the day (Tunderstory max and Topen-air were in the range of 21.56℃-22.56℃ and 22.87℃-23.99℃, respectively).

[0126] We also observed significant variation in the temperature difference between the understory and open spaces (ΔT, representing the difference between the understory and open spaces). Specifically, understory temperatures were higher in both average and minimum temperature differences (ΔT average ranged from 1.98°C to 2.24°C, and ΔT min ranged from 5.51°C to 5.91°C), but lower in maximum temperature differences (ΔT max = -1.37 ± 0.14°C).

[0127] S31: The comparison of autumn phenology between the understory and canopy further demonstrates the reliability of camera monitoring in the understory;

[0128] The present invention calculated the required CDD (Cold Degree Days) based on EOS, Tunderstory, and Topen-air. CDD is based on a developmental reference temperature (Topt) estimated to be between 15 and 25°C. The CDD model in the present invention showed that the EOS of understory plants was significantly later, approximately 17 days, compared to the canopy position (Topt = 17°C). However, the lag time between the understory and the canopy varied depending on the extraction method and the vegetation index used. Similar conclusions were drawn for other developmental reference temperature values, except that the CDD values varied accordingly. This variation was attributed to the strong vertical microclimate gradient from the crown to the roots, which led to significant variations in temperature and light availability. In addition, the present invention used TPM (Temperature and Photoperiod Multiplicative) to compare the EOS between the understory and the canopy. TPM showed that leaf senescence in the understory was significantly later, with a lag time of 12.5 to 14 days. The lag time estimated by TPM is shorter than that of CDD, mainly because TPM considers the effects of photoperiod on the induction of vegetation leaf abscission meristem and frost tolerance, which makes TPM more widely applicable and robust than existing process-based autumn phenology models.

[0129] S4: Vegetation phenology extraction: remote sensing phenological periods based on vegetation index to determine the curvature of change rate and dynamic threshold;

[0130] S41: Use Savitzky–Golay (SG) filter to smooth the VI time series;

[0131] The filter parameters include a cubic polynomial (k=3) and a moving window of size 11 (m=11).

[0132] S42: Dynamic quantile values are used to extend the peak growth period and dormancy period to process the original vegetation index;

[0133] In this example, dynamic quantile values are used to extend the peak growth and dormancy periods, rather than the Maximum Value Composition (MVC) method for processing the raw vegetation index (VI). This method is applied to daily UVL4 understory camera data before smoothing. Quantile thresholding is used to extract the VI and timeline expansion.

[0134] Dynamic Quantile is a quantile threshold based on daily vegetation indices. It is used to adaptively identify extended vegetation indices during key phases of the vegetation growth cycle, such as peak and dormant periods. The core idea is to dynamically adjust the threshold based on changes in local data distribution, rather than fixing a global threshold.

[0135] Assume that the vegetation index time series in July and November is {X t}, t=1,2,...,T, the window size is w (w=30), and the quantile is q (for example, q=0.9 represents the 90% quantile). Dynamic quantile threshold Q t (q) is calculated as follows:

[0136] Q t (q) = Quantile({X1, ..., X t},q),

[0137] The extended peak period q is the 50%-100% quantile of the daily vegetation index time series in July, and the dormant period q is the 0%-50% quantile of the daily vegetation index time series in November. The Quantile() function is a quantile function.

[0138] In this embodiment, the original quantile is calculated by segmenting the vegetation index time series by month and setting a window for the key phenological stage: the quantile of the growth peak period (July) is recorded as Window size w=30, extract daily VI data {X t}, t∈[7 / 1,7 / 30]. The dormancy period (November) quantile is recorded as Window size w=30,{X t}, t∈[11 / 1,11 / 30].

[0139] Quantile time axis expansion: Extend forward [6 / 1,6 / 30], Replicate to the entire month of June (30 days) to form an extended sequence of growth peak extension periods Extended backward [12 / 1,12 / 30], Replicate to the entire month of December (30 days), forming an extended dormancy sequence

[0140] S43: For curve fitting, a four-parameter logistic function is used to simulate the dynamic fluctuation of vegetation greenness. The specific calculation formula is:

[0141]

[0142] where t is the day of the year, y(t) is the fitted vegetation index (VI) value at day t, and b1, b2, b3, and b4 are the fitting parameters.

[0143] The four-parameter logistic fitting function has upper and lower asymptotes. Extending the peak growth period by 30 days and delaying the dormancy period by 30 days helps ensure that the model fully captures the dynamic changes of vegetation indices, thereby better fitting the growth of vegetation from mature to dormant stages.

[0144] S44: Determine the dynamic quantiles of peak and dormant periods using Akaike information quantity;

[0145] The Akaike Information Criterion (AIC) is a statistic used for model selection. In the model selection process, the smaller the AIC value, the better the model. This is because the AIC not only considers the goodness of fit of the model (through maximum likelihood estimation) but also penalizes the model's complexity (number of parameters) to avoid overfitting. When comparing different models, the model with the smallest AIC value is preferentially selected as the optimal model.

[0146] The specific calculation formula of AIC is:

[0147] AIC=-2lnL+2k

[0148] Where k represents the number of parameters in the logistic model (including the intercept term); L is the maximum likelihood estimate of the model on the vegetation index; the better the model fit (i.e., the larger L), the smaller the AIC value.

[0149] S45: The remote sensing phenological period of rate of change curvature (RCC) and dynamic threshold was determined using the Logistic first-order derivative equation, Logistic third-order derivative equation, and the maximum smoothed vegetation index values of 30% and 50%.

[0150] The calculation formula of the first-order derivative equation is:

[0151]

[0152] The calculation formula of the third-order derivative equation is:

[0153]

[0154] Where t is the day of the year, and b1, b2, b3, and b4 are the fitting parameters.

[0155] In this embodiment, in order to obtain consistent and continuous EOS of understory plants, the present invention uses dynamic quantile values rather than the maximum or minimum vegetation index of the time series. This method prolongs the maturity and dormancy periods. The experiment found that the vegetation index values of different quantiles in the first 30 days have a significant effect on the shape of the logistic fitting curve of the maturity and dormancy periods. The most significant difference occurs near the peak at the beginning of the growing season (POS), followed by the end of the dormancy period (see Figure 4 ).

[0156] The AIC of dynamic quantiles is as follows Figure 5As shown. According to the fitted vegetation index time series indicators, the selected vegetation index value for the beginning of the mature period is almost the 60% quantile of the vegetation index time series in July, while the 15% quantile in November was selected as the end of the dormant period. The AIC values of GRVI, VARI, TGI, ExG, WFI, RGRI, GCC and IKaw were -1710.96, -1463.38, -1304.56, 893.63, -1315.30, -1456.77, -1840.16 and 1315.30, respectively, among which GCC had the highest accuracy. Judging from the accuracy alone, R 2 The values are 0.934, 0.928, 0.919, 0.929, 0.795, 0.795, 0.936, and 0.951, respectively, indicating that ExG is particularly suitable for high-precision extraction.

[0157] A system for quantifying autumn phenology in forest understory based on an infrared camera, according to the above-mentioned method for quantifying autumn phenology in forest understory based on an infrared camera, mainly includes:

[0158] Image acquisition and dynamic monitoring module:

[0159] Understory vegetation image acquisition: Infrared cameras are used to automatically capture time-series images of understory vegetation, covering the peak growth period to the autumn phenological stage (such as leaf color change, leaf fall, etc.).

[0160] Image storage and annotation: Temperature and lighting duration are obtained through ambient light sensors and PIR sensors.

[0161] Daily scale optimal image selection module:

[0162] Image feature screening unit: Filter invalid images based on the Green-Red Vegetation Index (GRVI) and saturation threshold (≥0).

[0163] Cluster analysis and optimal cluster selection unit: The K-means clustering algorithm and elbow rule are applied to determine the optimal number of clusters; by comparing images within the cluster, the cluster that best reflects the greenness of the vegetation is selected.

[0164] Daily optimal image extraction unit: Filter the daily optimal image from the best cluster. If there is no valid image in the cluster, the suboptimal cluster is selected.

[0165] Meteorological data consistency verification module:

[0166] Data synchronization and matching: Time-align the understory meteorological data (such as temperature) collected by the infrared camera with the canopy weather station data.

[0167] Statistical test analysis: Verify the consistency of the two types of data through correlation analysis or error test (such as root mean square error RMSE).

[0168] Vegetation phenology dynamic extraction module:

[0169] Time series smoothing unit: Use SG filter (Savitzky-Golay) to denoise and smooth the vegetation index time series.

[0170] Dynamic Quantile and Curve Fitting Unit: Calculates the dynamic quantile threshold and fits the vegetation index dynamic curve using a four-parameter logistic function. The optimal quantile threshold is selected using the Akaike Information Criterion (AIC).

[0171] Phenological period acquisition unit: Combine the first-order derivative (rate of change), third-order derivative (curvature) and dynamic threshold to determine the key nodes of autumn phenology (such as the beginning of color change and the end of leaf fall).

[0172] The present invention combines dynamic threshold recognition with infrared camera imaging to dynamically monitor forest understory vegetation, effectively eliminating redundant or low-quality images, improving the efficiency of vegetation index extraction, and increasing the accuracy of phenological period determination; at the same time, achieving lightweight and efficient algorithms.

[0173] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for quantifying autumn forest understory phenology based on infrared cameras, characterized by: The following steps are involved: S1: Understory vegetation image collection: Use an infrared camera to capture and collect understory vegetation images during the phenological stage; S2: Daily-scale vegetation optimal image acquisition: Image feature acquisition is performed using the daily optimal image selection algorithm for understory vegetation that integrates multi-feature clustering and hierarchical screening; S3: consistency check between understory infrared camera meteorological data and canopy meteorological station data; S4: Dynamic extraction of vegetation phenology: The dynamic quantile method is used to obtain the time series change curve, and the remote sensing phenological period of the change rate curvature and dynamic threshold is determined based on the vegetation index.

2. The method for quantifying autumn forest phenology based on infrared cameras according to claim 1, characterized in that: The specific process of S2 is: S21: Image screening and clustering: K-means clustering algorithm and elbow rule are used for image selection; S22: Daily-scale optimal image: Determine the image in the cluster that best reflects the greenness of vegetation.

3. The method for quantifying autumn forest phenology based on infrared cameras according to claim 2, characterized in that: The specific process of S21 is as follows: S211: Filter images whose green-red vegetation index and saturation are greater than or equal to 0; S212: Apply K-means clustering and the elbow rule to determine the optimal number of clusters based on the sum of squared distances and the Calinski-Harabasz score.

4. The method for quantifying autumn forest phenology based on infrared cameras according to claim 3, characterized in that: In S212, the specific content of K-means clustering is: Step 1, select K cluster centers; Initialize K centroids c1, c2, c3, ..., c k , each centroid is a 7-dimensional vector, corresponding to 7 initial values of photo features, specifically: c k =[c k1 ,c k2 ,…,c k7 ], Among them, c k represents the centroid vector; Step 2, calculate distance and assign clusters; For each photo's 7-dimensional feature vector X i =[x i1 ,x i2 ,…,x i7 ], Among them, x i1 Indicates white balance; x i2 Indicates the direction of light; x i3 Indicates color temperature; x i4 Indicates contrast; x i5 Indicates grayscale; x i6 Indicates saturation; x i7 Indicates brightness; Calculate the Euclidean distance d to all centroids, and then assign the photo to the cluster to which the nearest centroid belongs, specifically: Among them, X i is a 7-dimensional feature vector; c k is the centroid vector; x ij is the jth eigenvalue of the i-th photo; c kj represents the jth eigenvalue of the kth cluster center; Step 3, update the centroid; Recalculate the centroid of each cluster and take the mean of the feature values of all photos in the cluster; The centroid update formula of the kth cluster for: Among them, C k is the photo set of the kth cluster; |C k | is the number of photos in the cluster; j represents the feature dimension; x ij is the jth eigenvalue of the i-th photo; x i is the i-th photo; Step 4, iterate until convergence: Repeat steps 2 and 3 until the centroid position no longer changes or the maximum number of iterations is reached, and finally determine the category of each photo.

5. The method for quantifying autumn forest phenology based on infrared cameras according to claim 3, characterized in that: In S212, the specific content of the elbow rule is: The elbow rule determines the optimal number of clusters by analyzing the relationship between the evaluation index of the clustering result and the number of clusters; the SSE evaluation index is used for evaluation to adaptively select the best cluster center. The calculation formula of the SSE evaluation index is: Among them, K is the cluster number; C k is the photo set of the kth cluster; x ij is the jth eigenvalue of the i-th photo; c kj represents the j-th eigenvalue of the k-th cluster center.

6. The method for quantifying autumn forest phenology based on infrared cameras according to claim 2, characterized in that: The specific process of S22 is as follows: S221: Determine the cluster that best reflects the greenness of vegetation by comparing the images in each cluster; S222: Finally, the most suitable photo for each day is selected within the best cluster; if not, a photo in the next cluster is selected.

7. The method for quantifying autumn forest phenology based on infrared cameras according to claim 1, characterized in that: The specific process of S4 is: S41: Use SG filter to smooth the vegetation index time series; S42: dynamic quantile thresholds were used to extend the peak growth period and dormancy period to process the original vegetation index; S43: curve fitting using a four-parameter logistic function to simulate the dynamic fluctuations of vegetation greenness; S44: Determine the dynamic quantiles of peak and dormant periods using Akaike information quantity; S45: The remote sensing phenological period of the rate of change curvature and dynamic threshold is determined using the first-order logistic derivative, the third-order logistic derivative, and the maximum smoothed vegetation index value based on the threshold.

8. The method for quantifying autumn forest phenology based on infrared cameras according to claim 7, characterized in that: In S42, the dynamic quantile threshold Q t The calculation formula for (q) is: Q t (q)=Quantile({X1,...,X t },q), Among them, q represents the quantile, X t Represents the vegetation index time series, and Quantile() represents the quantile function.

9. The method for quantifying autumn forest phenology based on infrared cameras according to claim 7, characterized in that: In S44, the calculation formula of Akaike information content AIC is: AIC=-2lnL+2k Where k represents the number of parameters in the logistic model; L is the maximum likelihood estimate of the model on the vegetation index.

10. A system for quantifying autumn phenology in forest understory based on an infrared camera, according to the method for quantifying autumn phenology in forest understory based on an infrared camera according to any one of claims 1 to 9, characterized in that: It includes image acquisition and dynamic monitoring module, daily scale optimal image selection module, meteorological data consistency verification module, and vegetation phenology dynamic extraction module.

Images