Monitoring Autumn Phenology in Understory Plants with a Fine-Resolution Camera

Abstract

Autumn phenology plays a crucial role in shaping the capacity for carbon sequestration. However, understories, a vital yet often neglected ecosystem component, have complicated autumn phenology prediction. We address the challenge of monitoring understory phenological dynamics by using a UVL4 trail camera and selecting appropriate deriving processes and vegetation indices (VIs). We found the understory photoperiod was on average 1.88 h shorter than the canopy’s, while the understory temperature was 2.11 °C higher than the canopy’s open-air temperature. The maximum temperature inside the understories was on average 1.37 °C lower than in open-air conditions. Specifically, the 60% quantile of the daily VI in July and the 15% quantile in November effectively captured the prolonged minimum and the minimum in the VI time series when applying logistic modeling. The excess green vegetation index (ExG) outperformed other VIs in estimating understory greenness change. The cold degree days model (CDD) and low-temperature and photoperiod multiplicative model (TPM) revealed that senescence progressed from the upper crown downwards, causing over 13 days of lag in the understory. These findings offer a new perspective on quantifying autumn phenology in subtropical forests and provide insights into asynchronous changes in vertical microclimatic gradients in Earth system and vegetation models.

1. Introduction

Vegetation phenology, a key driver of ecosystem processes, profoundly influences terrestrial carbon cycling and climate feedback [1,2]. Autumn phenological events, including leaf color change and senescence, serve as critical indicators of plant climate adaptation while governing ecosystem energy fluxes and biogeochemical cycles [3,4]. Within forest ecosystems, understory plants constitute the majority component of biodiversity and significantly contribute to ecological functions [5,6]. Recent studies have shown that the autumn phenology of understory vegetation is crucial for understanding the energy balance, carbon cycling, and ecological stability and resilience [7,8]. With the increasing recognition of the importance of understory plants for the ecosystem function and biodiversity, precisely monitoring their autumn phenology is fundamental to understanding ecosystem dynamics and biodiversity conservation mechanisms.

Traditional autumn phenology monitoring relies on in situ observers recording tree senescence, such as when more than 50% of leaves change color or fall [9]. However, this method suits individual species rather than entire ecosystems. Satellite-based remote sensing has broadened macro-scale phenological studies amid global climate change [10]. Indices like the normalized-difference vegetation index (NDVI) [11], enhanced vegetation index (EVI) [12], and the near-infrared reflectance of vegetation (NIRv) [13] are employed to detect phenological events. VI-derived phenology involves improving VI time-series quality, fitting the seasonal VI, and identifying the start and end of the growing season (SOS and EOS) using derivative or threshold methods [14,15]. However, factors like poor observation conditions, a coarse spatial and temporal resolution, sensor calibration, and BRDF effects can reduce satellite-derived phenological accuracy [16,17]. Repeated digital photography at the ground level with a fine resolution offers a cost-effective method to monitor greenness [18,19]. Camera-based phenological observation networks have been established in North America (PhenoCam) [20], Europe (EuroPhen) [21], Australia (Australian PhenoCam Network) [22], Japan (Phenological Eyes Network) [23], and China (under construction) [10]. It is important to note that cameras used for phenological observations are typically mounted on flux towers, which do not allow for documenting greenness changes in the understory.

Trail cameras, primarily used to monitor animal activity, can potentially monitor understory ecosystems [19,24,25]. These low-cost digital cameras are designed for prolonged, unsupervised outdoor use and equipped with ambient light sensors that automatically adjust photograph color to monochrome. Unlike traditional digital cameras, trail cameras capture additional information like the temperature and photoperiod via their sensors. Initial studies have shown the feasibility of using red, green, and blue (R, G, and B) values to track greenness changes [26]. For example, indices like the green chromatic coordinate (GCC), red chromatic coordinate (RCC), and blue chromatic coordinate (BCC), as well as the excess green index (ExG), normalized green–red difference index (GRVI), and visible atmospherically resistant index (VARI), have been employed [18,27]. While the GCC is the most commonly used VI in digital cameras, it is not optimal for detecting the EOS in understory plants with distinct leaf angles [28]. Camera photographs may also vary due to auto white balance (AWB) settings and external environmental factors. Thus, selecting suitable VIs and preprocessing approaches for extracting the understory EOS requires serious reconsideration.

Traditional cameras often struggle to capture both vegetation phenological dynamics and related environmental factors like temperature and light. This study addresses that gap by using a UVL4 trail camera, capable of high-frequency imaging and simultaneous temperature and light monitoring. We continuously monitored understory greenness change and derived autumn phenology in an understory ecosystem containing pine and Dryopteris sp. Cyclobalanopsis glauca, among others. Our research objectives were to select suitable photographs and preprocess the VI to quantify understory greenness; to assess VI performance in tracking the understory autumn phenology using various extraction methods; and to compare the canopy and understory autumn phenology using process-based models (CDD and TPM).

2. Materials and Methods

2.1. Study Site

This study was conducted in Yuexi County, Anqing City, Anhui Province, China (30°57′N–31°06′N and 116°02′E–116°11′E, within the Yaoluoping National Nature Reserve) (Figure 1). This area features an understory with minimal human impact and a favorable light environment The reserve’s vegetation is characterized as a northern subtropical deciduous–evergreen broad-leaved mixed forest, with the primary conservation objective being the preservation of the Dabie Mountain region’s native forest ecosystem. In this study, a UVL4-CN IR camera was placed 1 m above ground level on tall trees, at a vertical angle of 30$°$. Its field of view included small trees, saplings, and herbaceous plants. Data were recorded during the growing season (days of the year 180 to 333, 2022), when summer begins and plant growth enters a critical period. During this time, the ecological functions of the understory vegetation, such as water retention, soil stabilization, and nutrient cycling, become more evident. Based on the vegetation characteristics, understory vegetation was defined as plants shorter than 5 m.

2.2. UVL4-CN IR Camera and Image Preprocessing

The UVL4-CN IR camera (Figure 2) was equipped with an ambient light sensor that triggered the infrared LED light source when the illuminance fell below 500 lux. The temperature was directly monitored by the temperature sensor, and the duration of daylight was indirectly recorded by the light sensor when capturing RGB images, with illuminance exceeding 500 lux. The camera was installed 1 m above the ground at a downward angle of 30 degrees, capturing understory vegetation and the lower parts of trees up to 5 m in height. Photographs were taken hourly, starting at 10:30 on 9 July 2022 and ending at 21:30 on 28 November 2022, resulting in 24 photos per day and a total of 2908 photographs.

To ensure a consistent image quality in our study, we considered several factors affecting the camera’s performance. Operating under automatic control, the camera produced a variable image quality influenced by factors such as the auto white balance (AWB), light intensity (LI), light direction (LD), correlated color temperature (CCT), contrast (CTR), gray level (GL), saturation (Sat), and brightness (Bri) in

Table A1. The primary factors influencing image quality, accompanied by a succinct description of each factor’s impact on the image.

Factor	Abbreviation	Characteristic
Auto White Balance	AWB	The white balance setting of the image, affecting the accurate reproduction of colors.
Light Intensity	LI	The intensity of the light source, affecting the exposure and brightness of the image.
Light Direction	LD	The direction of the light source, affecting the shadows and depth of the image.
Correlated Color Temperature	CCT	The color temperature of the light source, affecting the warmth or coolness of the image colors.
Contrast	CTR	The contrast between light and dark areas in the image, affecting its clarity and detail.
Gray Level	GL	The levels of gray in the image, affecting its tonal depth and detail.
Saturation	Sat	The purity of the colors in the image, affecting how vivid the colors appear.
Brightness	Bri	The overall brightness of the image, affecting its lightness or darkness.

[29]. These factors could impact image quality, and the digital camera applied non-linear transformations to the image data beyond user control [30]. To analyze this variability, we used K-means clustering to divide images based on eight influence factors into K clusters, minimizing the sum of squares within clusters [31]. This method grouped images by AWB, LI, LD, CCT, CTR, Sat, and Bri. The optimal number of clusters (K) was determined using the elbow method [32], which clusters images with similar performance characteristics by reducing within-cluster error. Image selection followed these processes: Initially, we used the green–red vegetation index (CRVI) and saturation to eliminate photos with a CRVI and Sat below 0, ensuring daytime capture. Then, K-means clustering and the elbow method were applied to select an appropriate K based on the sum of squared distances (SSD) and Calinski–Harabasz Score (CH). We determined the optimal cluster by comparing images within each cluster for true vegetation greenness reflection. Finally, the most suitable photo for each day within the optimal cluster was selected; if unavailable, a photo from the next cluster was chosen.

Based on these processes, we determined K to be 3 and selected Cluster 1 as optimal (Figure 3). This cluster had the highest saturation (0.20) and lowest correlated color temperature (23.15), ideal for tracking vegetation greenness changes. However, external factors like precipitation, fog, and lighting conditions might have introduced photo noise. Considering the K-means clustering results of the photo parameters, Cluster 3 was the secondary choice. If no image was available in Cluster 1 for a day, one was selected from Cluster 3 [29].

2.3. Methods

2.3.1. Vegetation Index

Vegetation indices (VIs) are essential tools for examining vegetation dynamics, as they can delineate properties such as the photosynthetic activity and canopy structure. However, the results obtained from VIs may vary depending on the chosen index, quality control procedures, and compositing algorithms. When selecting VIs, it is crucial to consider their spectral sensitivity, unique characteristics, and suitability for the specific application [33]. In this study, we utilized eight vegetation indices (VIs) listed in

Table 1. Vegetation indices used in this study.

VI	Abbrev.	Formula	Characteristic	Reference
Green–red vegetation index	GRVI	$GRVI={\displaystyle \frac{(g-r)}{(g+r)}}$	Sensitive to land cover types	[35]
Visible atmospherically resistant index	VARI	$VARI={\displaystyle \frac{(g-r)}{(g+r-b)}}$	Anthocyanin in plant leaves	[36]
Triangular greenness index	TGI	$TGI={\displaystyle \frac{(2\ast g-r-b)}{(2\ast g+r-b)}}$	Plant health and stress	[37]
Excess green	ExG	$ExG=2\ast g-r-b$	Plant greenness	[37]
Blue–red vegetation index	BRVI	$\mathrm{B}\mathrm{R}\mathrm{V}\mathrm{I}={\displaystyle \frac{(b-r)}{(b+r)}}$	Plant health and chlorophyll	[38]
Red–green ratio index	RGRI	$RGRI={\displaystyle \frac{r}{g}}$	Vegetation health and vitality	[39]
Green chromatic coordinate	GCC	$GCC={\displaystyle \frac{g}{(r+g+b)}}$	Vegetation dynamic	[18]
Kawashima index	IKaw	$IKaw={\displaystyle \frac{(r-b)}{(r+b)}}$	Leaf nitrogen concentration	[40]

Notes: $r=R/(R+G+B)$, $g=G/(R+G+B)$, and $b=B/(R+G+B)$, where red (R), green (G), and blue (B) are the reflectance values of each band.

to detect autumn phenology: the green–red vegetation index (GRVI), visible atmospherically resistant index (VARI), triangular greenness index (TGI), excess green index (ExG), blue–red vegetation index (BRVI), red–green ratio index (RGRI), green chromatic coordinate (GCC), and Kawashima index (Ikaw). Each of these indices exhibits sensitivity to leaf coloration and vegetation photosynthesis [34]. Comprehensive details are provided in Table 1.

2.3.2. Calculation of the Photoperiod

The variation in photoperiod magnitude depends on the spatial location and date. This study used the latitude and day of the year to calculate the photoperiod [41]. The formula [Equation (1)] for this calculation formula is presented as follows:

$$\left\{\begin{array}{c}axis={\displaystyle \frac{\pi }{180}}\ast 23.439\\ m=1-\tan \left(lat\right)\ast \tan \left(axis\ast \cos ({\displaystyle \frac{\pi \ast doy}{182.625}})\right)\\ b={\displaystyle \frac{{\cos }^{-1}\left(1-m\right)}{\pi }}\\ photoperiod=24\ast b\end{array}\right.$$

(1)

Here, $axis$ represents the obliquity of the ecliptic, which is the tilt of the Earth’s rotation axis relative to its orbital plane. This tilt causes the equatorial plane to be misaligned with the ecliptic plane. Our study used an angle of 23.39°, which is stable on a human timescale, changing only over millennia. $m$ indicates the exposed radius part between the Sun’s zenith and Sun’s circle. $b$ is the fraction of the Sun’s circular exposure, calculated using geographical latitude $lat$ and the day of the year $doy$. $doy$ is the day’s position within the year, ranging from 1 to 365 (366 days in leap years).

2.4. Extraction of Autumn Phenology

2.4.1. Process-Based Model

The cold degree days model (CDD) is a temperature-based model that significantly impacts plant physiological dormancy and ecological dormancy. We used CDDi [Equation (2)] to measure chilling during autumn using the minimum temperature. In evaluating the cold degree days index (CDD and CDDi), we examined the impacts of various base temperatures (daily minimum and mean temperature at 10°, 15°, and 20 °C) over a three-month period (August, September, and October). Our objective was to determine which monthly CDD period, combined with the respective base temperature, could best explain the variations in autumn phenology among deciduous tree species [42].

$${CDD}_I=\sum \left(T_b-T_i\right) or \sum \left(T_b-T_{min}\right)$$

(2)

$T_b$, $T_i$, and $T_{min}$ represent the base temperature, daily mean temperature, and daily minimum temperature, respectively.

The low temperature and photoperiod multiplicative model (TPM) [43] enhances the simulation of autumnal processes by considering the interactive effects of low temperatures and the photoperiod. The model’s core hypothesis is that plant leaf senescence begins when either the photoperiod or minimum temperature exceeds a critical threshold, with the senescence process governed by the interplay between these two factors. As autumn progresses, the daily minimum temperature drops and the photoperiod shortens, accelerating leaf senescence. The model specifies that autumn leaf phenology occurs on a specific date, Y, when the leaf senescence state (R_sen) reaches a predefined critical value (S_sen), as shown in Equation (3).

$$S_{sen}={\sum }_{i=D_{start}}^Y{R}_{sen}\left(T_i,P_i\right)=S_{sen}^{\mathrm{*}}$$

(3)

Here, the leaf senescence state ($S_{sen}$) is the sum of the daily leaf senescence rate ($R_{sen}$) from $D_{start}$ to the date when autumn phenology occurred (Y). $D_{start}$ can be determined in two ways: (1) If leaf senescence is triggered by the photoperiod, $D_{start}$ is the first day after the summer solstice (173rd day of the year) when the photoperiod falls below a predefined threshold ($P_{start}$), as described in Equation (4). (2) If leaf senescence is triggered by low temperature, $D_{start}$ is the first day after the peak multiyear average daily minimum temperature (the 200th day of year) when the daily minimum temperature drops below a temperature threshold ($T_{start}$), as detailed in Equation (5).

$$D_{start}=firstday when P_i>173<P_{start}$$

(4)

$$D_{start}=firstday when T_i>173<T_{start}$$

(5)

The daily leaf senescence rate ($R_{sen}$, arbitrary unit) in Figure 4 is an exponential function of the product of the daily minimum temperature and photoperiod, increasing as this product decreases [Equation (6)].

$$R_{sen}\left(T_i,P_i\right)=\left\{\begin{array}{ll}{\displaystyle \frac{1}{(1+\exp \left(a\times \left(T_i\times P_i-b\right)\right))}}& i\ge D_{start}\\ 0& i\ge D_{start} \end{array}\right.$$

(6)

where $T_i$ and $P_i$ are the minimum temperature and photoperiod on day $i,$ respectively. The TPM has four fitted parameters, $P_{start}/T_{start}$, a, b, and $S_{sen}^{\mathrm{*}}$, with a > 0 and b > 0.

2.4.2. VI Curve Fitted Method

We interpolate the vegetation index time-series data for specific periods: from June 15 to July 9 and from November 29 to December 31. This interpolation is based on the VI data from the temporally adjacent 30 days. However, we recognize that using constant parameters in the maximum value composite (MVC) during preprocessing might not accurately capture vegetation dynamics, as noise can affect the maximum or minimum VI values. To mitigate the impact of noise, we adopt a dynamic quantile value composite instead of the MVC procedure as a standard preprocessing method for raw VI data. This approach is applied to the daily UVL4 trail camera data before smoothing. Subsequently, the VI time series are smoothed using a Savitzky–Golay (SG) filter. The filter parameters include a polynomial degree of three (k = 3) and a moving window size of eleven (m = 11). For the curve fitting, we employ a four-parameter logistic function, as shown in Equation (7), to model the natural fluctuations in vegetation greenness.

$$y\left(t\right)={\displaystyle \frac{b_1}{(1+\exp \left(b_2+b_3\ast t\right))}}+b_4$$

(7)

where $t$ is the time on the day of the year. $y\left(t\right)$ is the fitted VI value at date $t$. $b_1$, $b_2$, $b_3$, and $b_4$ are the fitting parameters.

2.4.3. Phenology Extraction Method

The rate of change curvature (RCC) and dynamic threshold are determined using the first derivative Equation (8), third derivative Equation (9), and maximum smoothed VI values of 30% and 50% (Figure 5).

$$y^{\prime }\left(t\right)={\displaystyle \frac{b_1\ast b_3\ast \mathrm{e}\mathrm{x}\mathrm{p}(b_2+b_3\ast t)}{{(1+\exp \left(b_2+b_3\ast t\right))}^2}}$$

(8)

$$y^{‴}\left(t\right)={\displaystyle \frac{-b_1{b}_3^3\mathrm{e}\mathrm{x}\mathrm{p}(b_2+b_{3}t)(1-4\exp \left(b_{2+}{b}_{3}t\right)+(\exp \left(2\left(b_2+b_{3}t\right)\right))}{{(1+\exp \left(b_2+b_3\ast t\right))}^4}}$$

(9)

3. Results

3.1. Temperature and Photoperiod Pattern of Understory Plants

In Figure 6a, we compare the theoretical and understory photoperiods. The results revealed that the average theoretical photoperiod was 12.25 h from early July to late November. In contrast, the understory photoperiod was slightly shorter, averaging 1.88 $\pm$ 0.14 h. As the seasons transitioned from summer to autumn, a noticeable divergence emerged between the theoretical photoperiod in open-air and the actual photoperiod in the forest understory. Regarding the temperatures in the forest understory as recorded by the UVL4 trail camera, the average daily temperature (Tunderstory), daytime temperature (Tunderstory max), and nighttime temperature (Tunderstory min) at 1 m above the ground were 18.54 $\pm$ 0.45 °C, 22.06 $\pm$ 0.50 °C, and 16.94 $\pm$ 0.48 °C, respectively (Figure 6b–d). The understory temperatures were slightly higher than the average open-air temperature and the maximum temperature (Topen-air = 16.43$\pm$ 0.53 °C, Topen-air min = 11.23$\pm$ 0.58 °C), but cooler during the daytime (23.43$\pm$ 0.56 °C). We also noted a higher variability between the understory and open-air temperatures $(∆T$, representing the difference between the understory temperature and open-air temperature). Specifically, the understory temperature was warmer in both ${∆T}_{average}$ and ${∆T}_{min}$ (2.11$\pm$ 0.13 °C and 5.71$\pm$ 0.20 °C) but cooler in ${∆T}_{max}$ (−1.37$\pm$ 0.14 °C).

3.2. Comparison in Simulation Greenness Change

To obtain a consistent and sequential EOS for understory plants, we used a dynamic quantile value instead of the maximum or minimum vegetation index of the time series. This approach extended the periods of maturity and dormancy. Our findings revealed that different quantile vegetation index values from the preceding 30 days significantly influenced the shape of the logistic fitted curve during both maturity and dormancy phases. The most pronounced discrepancies occurred around the peak start of the season (POS), followed by the end dormancy (Figure 7). The AIC, BIC, RMSE, and R² of the dynamic quantile are presented in Figure A1. Based on these metrics for the fitted VI time series, the extended start maturity VI value selected was almost the 60% quantile of the July VI time series, while 15% of November was chosen as the end of dormancy. The AIC values for GRVI, VARI, TRI, 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 these, GCC exhibited the highest accuracy. When considering only accuracy, the R² values were 0.934, 0.928, 0.919, 0.929, 0.795, 0.795, 0.936, and 0.951, indicating that ExG was particularly suitable for high-precision extraction.

3.3. Comparisons of EOS Derived from VI Time Series

For the four phenological extraction methods, we identified significant disparities in the estimation of EOS across various VI time series (Figure 8 and Figure A2). Notably, the EOS extracted from the third derivative method was substantially later compared to the other three methods, with an average EOS of 279.02 days. Conversely, the first derivative method yielded the earliest EOS, averaging at 226.31 days. Regarding the variability of EOS estimates, we observed that the coefficient of variation (CV) was higher for the derivative methods (CV derivative method = 0.12) than for the threshold methods (CV threshold method = 0.05). Specifically, the first derivative method exhibited a greater mean fluctuation in EOS estimates compared to the third derivative, with CVs of 0.13 and 0.10, respectively. Additionally, the CV derived from the 30% threshold method was 0.06, which was higher than the 50% threshold CV of 0.04 (as shown in Figure A2). Post hoc tests of VIs using different EOS extraction methods (Figure 8) revealed a significant and consistent clustering pattern. The EOS extracted by TGI, ExG, and GCC showed a similar performance, outperforming the other VIs. In contrast, the stability of the EOS extracted by GRVI, VARI, and RGRI was almost identical, with VARI notably underperforming. Consequently, TGI, ExG, and GCC were more effective in capturing the EOS of understory vegetation. Among these, ExG, which reflected the highest precision of greenness change information, was able to detect the earliest changes compared to the other VIs.

Using the dynamic quantile value method, we identified the suitable quantile values for understory plant EOS in

Table 2. Comparison of the EOS extracted by four extraction methods with suitable VI time series.

VI	First Derivative	Third Derivative	30% Threshold	50% Threshold
GRVI	268.5	312.5	284	268
VARI	273.5	314.5	288	273
TGI	252.5	291.5	267	253
ExG	249.5	289.5	264	250
WFI	233.5	262.5	244	233
RCRI	269.5	313.5	285	269
GCC	250.5	293.5	266	250
Ikaw	233.5	262.5	244	233

. Notably, there was substantial variation in EOS estimates derived from different methodologies. The third derivative method demonstrated the highest sensitivity to temporal changes in the vegetation index, with a range of 52 days (from 314.5 to 262.5 days). This was closely followed by the 30% threshold method, which exhibited a range of 44 days (from 288 to 244 days). Both the first derivative method and the 50% threshold method showed a 40-day range, with the first derivative method ranging from 268.5 to 233.5 days, and the 50% threshold method from 268 to 233 days. Among the VIs, VARI yielded the latest EOS, while WFI and Ikaw provided the earliest. Comparatively, ExG and GCC showed intermediate EOS values across each extraction method when compared to other VIs.

The above results indicated that the threshold methods were relatively less sensitive to fluctuation in VI time series, whereas derivation methods were highly sensitive to noise in the raw data. The threshold methods effectively controlled the extracted EOS within a reasonable range, ensuring high accuracy. Based on this analysis, the 30% threshold method, combined with the ExG vegetation index, was deemed suitable for extracting understory plant EOS in subtropical forests.

3.4. Tracking EOS from Understory to Overstory Using the CDD Model and TPM

Thermal time, often measured in cold degree days, had proven to be a more effective predictor of phenological events compared to time of year or number of days [44,45]. We calculated the CDDs necessary from EOS, Tunderstory, and Topen-air. The CDDs were based on a developmental base temperature (Topt) estimated to between 15 and 25 °C. Our CDD model indicated that the EOS of understory plants occurred significantly later, by approximately 17 days (Figure 9), compared to canopy positions (Topt = 17 °C). However, the lag between understory and canopy varied with the extraction method and VIs used. Similar conclusions were drawn with other Topt values, with only the CDD value varying accordingly. This variation was attributed to the strong vertical microclimatic gradients from crowns to roots, leading to significant changes in temperature and light availability. In addition, we employed the TPM to compare the EOS between the understory and canopy. The TPM revealed a pronounced later leaf senescence in the understory, with a lag ranging from 12.5 to 14 days. The lag estimates by TPM were lower than those from CDD, primarily because the TPM accounted for the influence of the photoperiod on the induction of vegetation leaf abscission meristems and freezing resistance [43]. This made the TPM more broadly applicable and robust compared to the existing process-based autumn phenology models.

4. Discussion

The primary finding of our study is that trail cameras are a highly effective tool for capturing the continuous understory autumn phenology. This method is not only simple and automated but also cost-effective. Our results demonstrate a significant agreement between the understory temperature and photoperiod trends derived from trail cameras and those obtained from open-air meteorological stations and theoretical photoperiod calculations, respectively. Environmental factors, particularly illumination and weather conditions, significantly impact image quality and color representation [20]. Precipitation, in particular, can lead to an expansion of leaf cells, an increase in chlorophyll content, and a regulation of non-structural carbohydrate allocation and transformation. These physiological changes enhance the reflection and refraction of light, resulting in a greener leaf appearance, which requires careful consideration, especially during late autumn [46]. In line with previous studies, we addressed a contaminated VI by selecting the nearest day’s VI when all daytime photographs were unsatisfactory [47]. We also found that prolonged maturity and dormancy periods can significantly influence the curve-fitting shape and autumn phenological extraction from UVL4 trail camera data (Figure 7). Compared to the MVC method, using dynamic quantile values can reduce the environmental influence on the accuracy of greenness change capturing. This is attributed to the fact that while VIs tend to diminish under a cloud and aerosol influence according to the MVC hypothesis, trail camera VIs increase under environmental factors, making MVC unsuitable [48,49]. Our findings further indicate that prolonged maturity has a more significant impact on the curve shape than dormancy, as the maximum VI value predominantly influences the saturation and change rate of the curve [50,51].

Biophysical or structural parameters, such as the leaf area index (LAI), leaf angle distribution (LAD), and clumping index (CI), can greatly influence the understory light source and radiative transfer process, thereby diversifying the performance of vegetation indices in monitoring the vegetation status [33]. The eight VIs in Table 2 can effectively track the understory vegetation dynamic for phenological extraction. This is primarily because the digital camera’s visible light spectrum (R, G, and B) outperforms near-sensing infrared in assessing the understory vegetation greenness change at a close range [52]. The GCC and ExG are highly sensitive to the vegetation greenness, a direct indicator of chlorophyll activity and plant vitality. This sensitivity allows them to capture subtle vegetation changes that other indices might miss due to the soil background, light conditions, or other non-vegetative factors. The GCC calculates the proportion of green color in the RGB color model, making it less sensitive to light conditions compared to indices relying on absolute color values. This is crucial for consistent monitoring across different times of the day and varying weather conditions [53]. The ExG emphasizes green reflection while suppressing red and blue contributions, enhancing its robustness against lighting variations and making it superior at distinguishing plant material from background soil across different understory types [54].

Our study revealed that the autumn phenology predicted by the CDD model in the lower canopy area of mixed coniferous broad-leaved forest was, on average, seventeen days longer than in the upper crowns, while the TPM showed a thirteen-day difference. This aligns with findings by Zahnd, et al. [55] and Gressler, et al. [56] that leaf senescence progressed from the upper to the lower crowns. The lower canopy, in lower light conditions, senesces later than those in the strong light and the leaf lifespan is generally longer in shaded compared to sun-exposed conditions [57]. The CDD model, which focuses primarily on temperature accumulation, is influenced by the understory’s shielding from extreme temperatures by the canopy, leading to milder conditions that delay the onset of dormancy triggered by cold temperatures [58]. The TPM, integrating both low-temperature and photoperiod effects, might capture a more nuanced interaction of these factors, potentially detecting faster transitions in the understory compared to the sole temperature- based CDD model [43]. This directly demonstrates that the initiation and progression of leaf senescence are tightly regulated by multiple layers of factors, including internal and external factors such as age, phytohormones, and environmental stresses.

5. Conclusions

Leaf senescence is a pivotal stage in plant development, critically influencing plant functioning and the delivery of ecosystem services. Despite its importance, our understanding of the understory autumn phenology and microclimatic conditions within subtropical forests is currently limited. Our study employed the UVL4 trail camera to monitor the temperature, photoperiod, and greenness changes in the lower crowns of the Yaoluoping National Nature Reserve core zone. Our findings revealed that the understory experienced average and minimum temperatures that were 2.11 °C and 5.71 °C higher, respectively, compared to open-air temperature. Interestingly, the maximum temperature in the understory was only 1.37 °C lower than that in open air. Importantly, our results indicate that the ExG was the most effective in estimating changes in understory greenness, surpassing other indices in its sensitivity to vegetation greenness. The understory autumn phenology was approximately two weeks longer than that of the higher canopy positions. The TPM, which concurrently accounted for the low temperature and photoperiod, predicted an autumn phenology that was roughly four days shorter than the prediction made by the CDD model, which solely factored in the low temperature. These insights offer a novel approach to quantifying autumn phenology in understory subtropical forests and underscore the importance of considering the asynchronous changes in vertical microclimatic gradients in the Earth system and vegetation models. Our study contributes to a more nuanced understanding of the complex interactions between the phenology and microclimate, which is essential for improving the accuracy of ecological models and predicting the impacts of climate change on forest ecosystems.