Estimation of Understory Fine Dead Fuel Moisture Content in Subtropical Forests of Southern China Based on Landsat Images

Abstract

The understory fine dead fuel moisture content (DFMC) is an important reference indicator for regional forest fire warnings and risk assessments, and determining it on a large scale is a critical goal. It is difficult to estimate understory fine DFMC directly from satellite images due to canopy shading. To address this issue, we used canopy meteorology estimated by Landsat images in combination with explanatory variables to construct random forest models of in-forest meteorology, and then construct random forest models by combining the meteorological factors and explanatory variables with understory fine DFMC obtained from the monitoring device to (1) investigate the feasibility of Landsat images for estimating in-forest meteorology; (2) explore the feasibility of canopy or in-forest meteorology and explanatory variables for estimating understory fine DFMC; and (3) compare the effects of each factor on model accuracy and its effect on understory fine DFMC. The results showed that random forest models improved in-forest meteorology estimation, enhancing in-forest relative humidity, vapor pressure deficit, and temperature by 50%, 34%, and 2.2%, respectively, after adding a topography factor. For estimating understory fine DFMC, models using vapor pressure deficit improved fit by 10.2% over those using relative humidity. Using in-forest meteorology improved fits by 36.2% compared to canopy meteorology. Including topographic factors improved the average fit of understory fine DFMC models by 123.1%. The most accurate model utilized in-forest vapor pressure deficit, temperature, topographic factors, vegetation index, precipitation data, and seasonal factors. Correlations indicated that slope, in-forest vapor pressure deficit, and slope direction were most closely related to understory fine DFMC. The regional understory fine-grained DFMC distribution mapped according to our method can provide important decision support for forest fire risk early warning and fire management.

1. Introduction

The occurrence of forest fires often leads to significant ecological and resource losses, threatening human life and property [1,2,3]. According to projections by the United Nations Environment Programme (2022), global extreme wildfires will increase by 14% by 2030 and by 30% by 2050. The level of understory fine dead fuel moisture content (DFMC) directly affects the possibility of fire ignition and potential fire behavior, making it a critical factor and core indicator of fire early warning [4,5,6]. Therefore, the estimation of understory fine DFMC is a major focus of fire ecology and management research [7,8].

Current methods for estimating DFMC include equilibrium moisture content [9,10,11], process modeling [12], remote sensing estimation [13,14], and meteorological factor regression methods. Remote sensing, in particular, is an important means of assessing the spatial and temporal heterogeneity of the fuel moisture content due to its advantages of strong spatial continuity, large coverage, and the possibility of repeating observations over a region [15].

Remote sensing is primarily used to estimate the live fuel moisture content (LFMC) and assess how it affects fire activity [16,17,18,19,20]. For example, Nolan et al. [20] estimated LFMC in eastern Australian forests and woodlands using MODIS vegetation indices and empirical index models, and determined the LFMC thresholds for fire occurrence to be 156.1% and 101.5%, respectively. A single sensor exhibits obvious limitations in LFMC estimation, since the number of unknown parameters is generally greater than the number of observed spectral reflectance values. Therefore, considering multi-band data can provide more information for estimating the fuel moisture content, while multi-model coupling shows better generalization in the face of complex forest structures [15]. Given that the key parameters of the land surface (vegetation, water bodies, meteorology, soil, etc.) in nature are not independent of each other but are interdependent, the introduction of additional auxiliary variables and then estimation of LFMC by machine learning methods solves these problems to a certain extent [15,21,22]. Timeliness is the most important aspect of fire danger warning efforts. The Generalized Reduced Gradient (GRG) numerical optimization method, combined with the PROSAILH5B Radiative Transfer Model (RTM), has also been used to track the diurnal variation of LFMC [23]. These previous works assert that the use of empirical models and the incorporation of multi-source data can effectively improve the accuracy of estimations.

Meteosat Second Generation (MSG) data has been used to estimate air temperature and relative humidity, which can be used to infer DFMC. By comparing the DFMC estimated from remotely sensed meteorological data and interpolated meteorological data, the feasibility of using remotely sensed meteorological data to estimate DFMC has been validated. This approach provides a conceptual framework for remote sensing estimation of DFMC. However, the reliability of this method is affected by the lack of measured DFMC data [24]. The vapor pressure deficit has also been employed to estimate DFMC on a large scale; exponential relationships between DFMC and vapor pressure deficit have been previously found [8,25,26]. For instance, Nolan et al. [25] utilized remotely sensed estimates of vapor pressure deficit to indirectly assess the DFMC, demonstrating the model’s effectiveness across various regions. Nonetheless, the influence of any single factor on DFMC appears to be limited. In another study, Fan et al. [27] applied machine learning techniques to medium- to low-resolution remote sensing data to predict DFMC with a 1 h and 10 h lag on an hourly basis. These studies emphasize the application of medium- to low-resolution remote sensing in DFMC estimation, without involving the use of medium- to high-resolution remote sensing data for DFMC estimation. Additionally, the influence of terrain on DFMC has been largely overlooked, particularly in areas with significant topographic variations such as mountains and hills. The impact of terrain in these regions is substantial. Furthermore, we found that the meteorological parameters used in existing studies to predict understory fine DFMC are typically measured above the canopy, rather than in-forest. In-forest meteorology determines the moisture exchange between the fuels and the atmosphere, thereby influencing the spatiotemporal variability of the understory fine DFMC. Despite the similarity of forest canopy meteorology [28], there are significant differences in in-forest meteorology, influenced by vegetation cover, topography, and other factors [29,30,31,32]. Therefore, modeling accuracy for estimating understory fine DFMC may be higher using in-forest meteorology rather than canopy meteorology [26].

Factors closely associated with understory fine DFMC include temperature, relative humidity, wind, precipitation, topography, season, and vegetation conditions [7,27,33]. In general, under specific vegetation and stand conditions, the relative humidity in the forest increases and the temperature decreases after precipitation, leading to an elevation in the understory fine DFMC. However, precipitation data have not been heavily considered in previous remote sensing-based efforts to estimate understory fine DFMC. Also, it has been suggested that understory fine DFMC within stands of higher tree densities have higher and slower drying rates [32], but these relationships become complicated under different topography and vegetation types [34,35,36]. Due to the complex relationship between understory DFMC and other factors [37], the estimation of understory DFMC via consideration of only a single meteorological factor is usually not as effective as the combined effect of multiple factors. Therefore, integrating understory meteorology and other complex variables into medium- to high-resolution remote sensing for the estimation of understory fine DFMC presents a significant challenge.

Here, we (1) investigate the feasibility of Landsat images for estimating in-forest meteorology; (2) explore the feasibility of canopy meteorology or in-forest meteorology and explanatory variables for estimating understory fine DFMC; and (3) compare effects of each factor on model accuracy and its effect on understory fine DFMC.

2. Materials and Methods

2.1. Study Area

Our study area is in southern Jiangxi Province, China (Gannan region), from 113°54′ to 116°38′ East and 24°29′ to 27°09′ North (Figure 1). Gannan is located in the southern edge of the middle subtropical zone, belonging to the humid monsoon climate area of subtropical hills and mountains. The climate is mild, with an average annual temperature of 19.5 °C and an annual precipitation of 1400–1700 mm. The average altitude is 300–500 m, and the topography is mainly hilly and mountainous. The forest coverage in the Gannan is 76.23%. The forest stands dominated by flammable tree species such as Pinus massoniana, Pinus elliottii Engelm, and Cunninghamia lanceolata, accounting for over 40% of the regional forest area.

Gannan has a high population density (228 people per square kilometer) and a well-developed road network (115.3 km per 100 square kilometers). The Gannan region is prone to forest fires with the majority of fire causes being related to humans. According to statistics from the Ganzhou Forestry Bureau, there were 96 forest fires between 2016 and 2020, with a total burned area of 1057 hectares.

Understory fine DFMC monitoring devices (designed and manufactured by Northeast Forestry University, Harbin, Heilongjiang Province, China) were deployed in the forests of Yinkeng Ecological Forestry region in Yudu County and Nanshan Forest Park in Nankang District, Gannan for long-term positional monitoring of in-forest meteorology and understory fine DFMC.

2.2. Research Framework

Thirty-two L1-level images of Landsat-8 and -9 (no clouds over the sample area) were obtained from July 2021 through May 2023 from the U.S. Geological Survey (USGS, https://earthexplorer.usgs.gov/, accessed on 1 July 2023). Remote sensing image pre-processing was performed in the ENVI environment, and forest canopy meteorological factors and vegetation indices were calculated based on infrared band data (

Table 1. Types of variables used and their abbreviations.

Data Type	Variable	Abbreviate
Vegetation index	Normalized difference vegetation index	NDVI
Topography data	Slope (°)	SLP
	Aspect	ASP
	Elevation (m)	EVL
Regional-scale meteorological data	Precipitation (mm)	P
	Days since last precipitation (day)	DSLP
Sample-scale meteorological data	In-forest temperature (°)	IFT
	Canopy temperature (°)	CT
	In-forest vapor pressure deficit (kPa)	IFD
	Canopy vapor pressure deficit (kPa)	CD
	In-forest relative humidity (%)	IFRH
	Canopy relative humidity (%)	CRH
Season	Four seasons (1–12)	SEAS

). We used the NASA DEM_HGT V001 dataset [38] to obtain topography data, and we used European Center for Medium-Range Weather Forecasts ERA5-Land hourly data (https://cds.climate.copernicus.eu/, accessed on 1 July 2023) to obtain precipitation data. Extrapolation of in-forest meteorological factors from the sample plot scale to the regional scale was then realized by building random forest models of in-forest meteorological factors monitored by the device and canopy meteorology estimated by remote sensing, as well as other explanatory variables. Secondly, sample-scale understory fine DFMC random forest models were constructed based on sample-scale meteorology (in-forest and canopy), topography, vegetation, precipitation, and seasonal factors. Finally, the optimal estimation model was screened and the understory fine DFMC distribution was mapped. We then obtained the contribution of each factor to the optimal model and explored the correlation of each factor to fine DFMC. The technical route is shown in Figure 2.

2.3. Sample Plot Layout

Thirty-nine monitoring devices were deployed in Gannan to monitor understory fine DFMC and in-forest meteorology at a time step of 2 h. Considering the damage of equipment or data anomalies, 25 of these devices were selected to monitor data (

Table 2. Overview of the sampling.

ID	ASP	SLP (°)	EVL (m)	Vegetation Form	ID	ASP	SLP (°)	EVL (m)	Vegetation Form
1	12.5	8.7	150	Coniferous forest	14	171	15	43	Broadleaf forest
2	11.3	8.4	168	Mixed forest	15	151	21	78	Broadleaf forest
3	16.2	22	215	Mixed forest	16	151	20	126	Broadleaf forest
4	5	12.5	172	Mixed forest	17	164	13	97	Coniferous forest
5	20	15.2	184	Broadleaf forest	18	238	4	126	Coniferous forest
6	42	15.9	176	Mixed forest	19	258	5	92	Mixed forest
7	67.8	23.9	182	Coniferous forest	20	232	4	165	Coniferous forest
8	4	17	183	Broadleaf forest	21	260	4.5	53.7	Coniferous forest
9	42	18	185	Broadleaf forest	22	314	7.8	132	Coniferous forest
10	87	5.5	151	Coniferous forest	23	293	7.1	135	Coniferous forest
11	81	11	154	Broadleaf forest	24	218	6.8	85	Coniferous forest
12	135	2	149	Coniferous forest	25	271	5.1	81	Mixed forest
13	149	4	84	Coniferous forest

and Abbreviations Section).

The understory fine DFMC monitoring device consists of a weighing device, a solar powered device, a meteorological device, and data storage (Figure 3). Sample conditions and device operation were regularly checked through the field, and data were accessed through an Android cell phone. The tensile force sensor connected with weighing tray can carry out real-time weighing of fuels; the meteorological element device automatically obtains relative humidity (%), air temperature (°C), light intensity (lux), and wind speed (m/s), etc., of which the maximum weighing range is 0.1–1000 g, the accuracy is ±0.1 g, the temperature error is ±0.5 °C, the relative humidity error is ± 0.1%, wind speed error is ± 0.1 m/s, and light intensity error is ± 0.1 lux.

A 50 cm × 50 cm sample plot was set up to collect fine dead fuels from the forest understory in each sample plot (Figure A1 in Appendix A). The collected samples were brought back to the laboratory and put into an oven for drying until the weight of the samples no longer changed. Then, the dry weight of the samples was recorded and put back into the field monitoring device. The monitoring device was then set to obtain the weight of the fuels. The DFMC is calculated as follows (Equation (1)):

$$\mathrm{D}\mathrm{F}\mathrm{M}\mathrm{C}\left(\mathrm{\%}\right)={\displaystyle \frac{W_{tw}-W_{td}}{W_{td}}}\times 100$$

(1)

where $W_{tw}$ is the weight of dead fuel and $W_{td}$ is the weight of dried dead fuel (

Table A1. Summary of the stand information for the sample plots.

ID	Mean Height (m)	Mean Diameter at Breast (cm)	Canopy Density (%)	Major Tree Species
1	8.23	12.29	40	Pinus
2	6.68	8.59	70	Pinus, Schima superba
3	11.33	16.8	80	Pinus, Liquidambar formosana Hance
4	9.25	9.35	60	Pinus, Schima superba
5	15.6	18.07	55	Liquidambar formosana Hance
6	8.05	11.65	60	Pinus, Schima superba
7	5.82	8.27	20	Pinus
8	9.16	10.6	40	Schima superba
9	9.2	7.38	60	Schima superba
10	8.87	11.9	20	Pinus
11	11.78	19.07	80	Schima superba
12	9.36	13.78	50	Pinus
13	8.36	11.97	25	Pinus
14	11.38	14.18	60	Schima superba
15	6.44	9.89	50	Schima superba
16	14.33	21.48	70	Schima superba
17	12.58	17.06	50	Pinus, Cunninghamia lanceolata
18	5.36	7.29	5	Pinus,Cunninghamia lanceolata
19	7.11	12	10	Pinus, Schima superba
20	4.09	8.15	20	Pinus
21	13.7	17.57	30	Pinus
22	11.69	18.04	60	Pinus
23	9.88	16.25	75	Pinus
24	6.32	10.11	30	Pinus, Cunninghamia lanceolata
25	8.62	12.02	60	Pinus, Liquidambar formosana Hance, Schima superba, Cunninghamia lanceolata

in Appendix A).

2.4. Remote Sensing Estimation of Meteorological Data

2.4.1. Temperature

The two thermal infrared bands of Landsat-8 and Landsat-9 are widely used in the estimation of temperature. When estimating temperature from Landsat imagery, the atmospheric effect on ground thermal radiation is first estimated, and then this atmospheric effect is subtracted from the total thermal radiation observed by the satellite sensors to obtain the ground thermal radiation intensity. This is then subtracted from the total amount of thermal radiation observed by the satellite sensors to obtain the ground thermal radiation intensity, which is then converted into the corresponding ground temperature by a specific algorithm [39]. The brightness L_λ value of thermal infrared radiation received by the satellite sensor comprises three components: the brightness of atmospheric upward radiation, the energy of the actual ground radiance that reaches the satellite sensor after passing through the atmosphere, and the energy of atmospheric downward radiation that reaches the ground and is reflected back. The thermal infrared radiation brightness value L_λ received by the satellite sensor can be expressed using the radiation transfer equation (Equations (2)–(4)):

$$L_{\lambda }=\left[\epsilon B\left(T_s\right)+\left(1-\epsilon \right)L\downarrow \right]\tau +L\uparrow$$

(2)

where ε is the surface-specific radiance, T_S is the real temperature of the surface (K), L↓ is the radiant luminance projected downward to the surface by the atmosphere, L↑ is the radiant luminance emitted from the surface and reaching the satellite sensors after passing through the atmosphere, and B(T_S) is the thermal radiant luminance of the blackbody, which is the transmittance of the atmosphere in the thermal infrared band. The radiant brightness B(T_S) of the blackbody with temperature T in the thermal infrared band is then

$$B\left(T_S\right)={\displaystyle \frac{\left[L_{\lambda }-L\uparrow -\tau \left(1-\epsilon \right)L\downarrow \right]}{\tau \epsilon }}$$

(3)

T_s can be obtained as a function of Planck’s formula:

$$T_S={\displaystyle \frac{K_2}{ln\left(\frac{K_1}{B\left(T_S\right)}+1\right)}}$$

(4)

• to TM, K₁ = 607.76 W/(m²·µmsr), K₂ = 1260.56 K;
• to ETM+, K₁= 666.09 W/(m²·µm·sr), K₂ = 1282.71 K;
• to TIRS Band10, K₁ = 774.89 W/(m²·µm·sr), K₂ = 1321.08 K.

2.4.2. Relative Humidity and Vapor Pressure Deficit

Atmospheric water vapor content is closely related to relative humidity and vapor pressure deficit. The split-window covariance–variance ratio (SWCVR) method is often used to calculate the atmospheric water vapor content [40]. The SWCVR method can also be used for Landsat data [41]; according to this approach, the atmospheric conditions and specific emissivity do not vary, only the surface temperature does, according to Equations (5)–(7):

$$W=a\left({\displaystyle \frac{{\tau }_j}{{\tau }_i}}\right)+b$$

(5)

$${\displaystyle \frac{{\mathsf{\tau }}_{\mathrm{j}}}{{\mathsf{\tau }}_{\mathrm{i}}}}={\displaystyle \frac{{\mathsf{\epsilon }}_{\mathrm{i}}}{{\mathsf{\epsilon }}_{\mathrm{j}}}}\times {\mathrm{R}}_{\mathrm{j},\mathrm{i}}$$

(6)

$$R_{j,i}={\displaystyle \frac{{\sum }_{K=1}^N\left(T_{i,k}-\overline{T_i}\right)\left(T_{j,k}-\overline{T_j}\right)}{{{\sum }_{K=1}^N\left(T_{j,k}-\overline{T_j}\right)}^2}}$$

(7)

where W is the water vapor content, a and b are the coefficients of the water vapor content model, τ_i is the atmospheric transmittance of the i-band, τ_j is the atmospheric transmittance of the j-band, ε_i is the surface-specific emissivity of the i-band, and ε_j is the surface-specific emissivity of the j-band, k denotes the kth image element in N, T_i,k is the luminance temperature of the i-band for the kth image element (K), T_j,k is the luminance temperature of the j-band for the kth image element (K), $\overline{T_i}$ is the average luminance temperature of the i-band of the Nth image element, and $\overline{T_j}$ is the average luminance temperature of the j-band of the Nth image element. For Landsat-8 and 9 data, i and j are 10 and 11 bands, respectively. Here, the SWCVR method was used to calculate the water vapor content of 14 Landsat-8 and 9 images and then resampled to a resolution of 30 m to obtain the water vapor in the canopy.

The required meteorological factors can be derived from Equations (8)–(11) including the actual water vapor pressure (e_a), where λ varies continuously with latitude and season [42], g is the acceleration of gravity, and δ is the ratio of the specific gas constants of water vapor and dry air (0.622).

$$e_a=\mathrm{g}{\displaystyle \frac{W\left(\lambda +1\right)}{\delta }}$$

(8)

$$e_s=0.6108 exp\left({\displaystyle \frac{17.27T}{T+237.3}}\right)$$

(9)

$$D=e_s-e_a$$

(10)

$$RH={\displaystyle \frac{e_a}{e_s}}\times 100$$

(11)

The saturated water vapor pressure (e_s) estimated by remote sensing was obtained by combining the canopy temperature estimated by remote sensing with Equation (9); T in Equation (9) is the canopy temperature estimated by remote sensing. Finally, the remotely sensed estimated vapor pressure deficit (D) was obtained by Equation (10); RH is relative humidity. We calculated the device-monitored e_s and e_a from the temperature obtained by the monitoring device according to Equations (9) and (11), and then, according to Equation (10), we obtained the device-monitored vapor pressure deficit.

2.5. Explanatory Variables

Explanatory variables include topography data, NDVI, precipitation data, and seasonal information. We extracted slope, aspect, and elevation from NASA DEMHGT V001 data. Aspect was reclassified as 0–180, with 0 for due north and 180 for due south.

NDVI, calculated from reflectance in the near-infrared and infrared bands (Equation (12)), is a commonly used metric for characterizing vegetation growth and cover [43]:

$$NDVI={\displaystyle \frac{{\rho }_{NIR}-{\rho }_{red}}{{\rho }_{NIR}+{\rho }_{red}}}$$

(12)

where NDVI is normalized vegetation index, NIR is near-infrared band reflectance, and red is infrared band reflectance.

The ERA5-Land hourly data were acquired for a 0.1° × 0.1° precipitation grid data. Hourly precipitation data were extracted and converted into day-to-day 24 h cumulative precipitation data. Since understory fine DFMC has a time lag in precipitation, we tracked daily precipitation data at the time of fuel monitoring and the number of days since the last precipitation event based on a criterion of having precipitation greater than 0.1 mm in the previous day.

Understory fine DFMCs usually show seasonal patterns of change [44]. For example, in spring and summer, Gannan has a long precipitation period and a high precipitation frequency. This indicates that fuels typically have a high moisture content during spring and summer, while in fall and winter, there is more dead fuel on the ground surface and the air is drier, leading to fuels having a lower moisture content. However, there is also variability in regional climatic conditions within seasons, e.g., in Gannan, precipitation is mainly concentrated in early rather than late summer. Therefore, the seasons were categorized monthly.

2.6. Model Construction and Evaluation

The temperature, relative humidity, precipitation, vapor pressure deficit, topography, and vegetation index information for each monitoring plot were extracted using the raster package [45] in the R 4.3.1 [46], totaling 283 records. We then constructed estimation schemes for in-forest meteorology as well as understory fine DFMC (

Table 3. Scheme construction.

Classification Scheme			Type of Feature
In-forest meteorological estimates	IFRH-scheme	IFRH-scheme1	CRH+NDVI+P+DSLP+SEAS+SLP+ASP+EVL
		IFRH-scheme2	CRH+NDVI+P+DSLP+SEAS
	IFT-scheme	IFT-scheme1	CT+NDVI+P+DSLP+SEAS+ SLP+ASP+EVL
		IFT-scheme2	CT+NDVI+P+DSLP+SEAS
	IFD-scheme	IFD-scheme1	CD+NDVI+P+DSLP+SEAS+ SLP+ASP+EVL
		IFD-scheme2	CD+NDVI+P+DSLP+SEAS
Understory fine DFMC estimation	D-DFMC -scheme	D-DFMC-scheme1	CD+CT+NDVI+P+DSLP+ SEAS+SLP+ASP+EVL
		D-DFMC-scheme2	IFD+IFT+NDVI+P+DSLP+ SEAS+SLP+ASP+EVL
		D-DFMC-scheme3	CD+CT+NDVI+P+DSLP+SEAS
		D-DFMC-scheme4	IFD+IFT+NDVI+P+DSLP+SEAS
	RH-DFMC -scheme	RH-DFMC-scheme1	CRH+CT+NDVI+P+DSLP+ SEAS+SLP+ASP+EVL
		RH-DFMC-scheme2	IFRH+IFT+NDVI+P+DSLP+ SEAS+SLP+ASP+EVL
		RH-DFMC-scheme3	CRH+CT+NDVI+P+DSLP+ SEAS
		RH-DFMC-scheme4	IFRH+IFT+NDVI+P+DSLP+SEAS

). The in-forest meteorological estimation scheme was differentiated by whether topography was considered or not. The understory fine DFMC scheme was divided into two categories (relative humidity and vapor pressure deficit) based on different representations of water vapor to explore the effects of using in-forest meteorology and topography on the estimation of fine DFMC under the same conditions and to filter out the optimal combination for the estimation of understory fine DFMC.

We used caret package in R 4.3.1 [47] for random forest model construction and ten-fold cross-validation to select the optimal model. Ten-fold cross-validation reduces the variance in model evaluation by dividing the dataset into ten subsets and using a different subset as the validation set in each iteration, while the remaining subsets serve as the training set. This methodology enhances the stability and reliability of the model’s performance estimates. It is well-suited for the random forest algorithm, which integrates decision tree models with bagging [48,49] and random feature selection to generate and aggregate predictions from numerous decision trees [50]. This ensemble technique allows the random forest to effectively capture non-linear patterns in the data, making it highly suitable for analyzing complex, high-dimensional datasets. To further ensure the robustness and generalizability of the model, 70% of the data was allocated for training, with the remaining 30% held out as an independent test dataset. Subsequently, the optimal random forest model was employed for the inversion of in-forest meteorological conditions and the estimation of understory fine DFMC. The Coefficient of Determination (R²), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE) were used to evaluate the model fit and prediction accuracy. A higher R² value indicates a better model fit, while lower MAE and RMSE values indicate smaller model errors.

The feature importance was calculated by random forest and the importance was normalized to ensure a sum of 100. In the characteristic importance of the random forest, the larger its value, the higher its contribution to the model. Pearson correlation analysis was conducted to examine the relationships between each factor and understory fine DFMC. A higher absolute value of the correlation coefficient indicates a stronger relationship, and a smaller p-value indicates greater significance.

3. Results

3.1. Evaluation of the Accuracy of In-Forest Meteorological Estimates

Seasonal differences in meteorological factors in the forest canopy calculated by remote sensing were significant (

Table 4. Forest canopy and in-forest meteorological estimates.

Season	N	CRH	IFRH	CD	IFD	CT	IFT
Spring	71	52.57%	89%	1.62 kPa	0.3 kPa	25.96°	24.67°
Summer	64	67.50%	93%	1.69 kPa	0.39 kPa	31°	33.7°
Autumn	111	50.06%	78%	2.16 kPa	0.95 kPa	28.95°	28°
Winter	37	38.65%	79%	1.27 kPa	0.35 kPa	17.85°	14°

). In all four seasons, the vapor pressure deficit within the forest was significantly lower than the canopy vapor pressure deficit, and the relative humidity within the forest was significantly higher than the relative humidity in the canopy. The temperature within the forest was lower than the canopy in all three seasons except summer.

It was found that the accuracy of each scheme of in-forest meteorology was improved after the introduction of topography (

Table 5. Accuracy results of in-forest meteorological inversion models.

Type	Scheme	R2	MAE	RMSE
IFRH-scheme	IFRH-scheme1	0.54	10.43%	13.04%
	IFRH-scheme2	0.36	14.97%	11.87%
IFT-scheme	IFT-scheme1	0.92	1.76°	2.34°
	IFT-scheme2	0.9	1.91°	2.53°
IFD-scheme	IFD-scheme1	0.47	0.47 kPa	0.65 kPa
	IFD-scheme2	0.35	0.49 kPa	0.73 kPa

and Figure 4). The IFRH-scheme1 fitting accuracy improved by 50%; the IFD-scheme1 fitting accuracy improved by 34%; and the IFT-scheme1 accuracy improved by 2.2%. The IFT-scheme1 had the best prediction in the in-forest temperature scheme (R² = 0.92, MAE = 1.76°, and RMSE = 2.34°); the in-forest relative humidity and vapor pressure deficit models had slightly worse predictions. The IFRH-scheme1 had an R² of 0.54, a MAE of 10.43%, and an RMSE of 13.04% for the in-forest relative humidity scheme, and the IFD-scheme1 had an R² of 0.47, a MAE of 0.47 kPa, and an RMSE of 0.65 kPa for the in-forest vapor pressure deficit scheme.

3.2. Evaluation of Understory Fine DFMC Model Accuracy

The results of the understory fine DFMC model accuracy (

Table 6. Accuracy results of understory fine DFMC estimation models.

Type	Scheme	R2	MAE (%)	RMSE (%)
RH-DFMC-scheme	RH-DFMC-scheme1	0.34	6.71	9.12
	RH-DFMC-scheme2	0.5	6.42	8.5
	RH-DFMC-scheme3	0.16	7.82	10.43
	RH-DFMC-scheme4	0.22	7.32	9.86
D-DFMC-scheme	D-DFMC-scheme1	0.4	6.47	8.92
	D-DFMC-scheme2	0.53	6.15	8.25
	D-DFMC-scheme3	0.18	7.55	10.20
	D-DFMC-scheme4	0.23	7.17	9.75

and Figure 5) show that the D-DFMC-Scheme2 had the best fit (R² = 0.53, MAE = 8.25%, and RMSE = 6.15%).

The models that used vapor pressure deficit, in-forest meteorology, and topography all showed significant improvements in accuracy compared to models that did not use these factors. The fitting accuracy of the vapor pressure deficit schemes improved by an average of 10.2% compared to the relative humidity schemes. The fitting accuracy of the schemes with in-forest meteorology improved by 36.2% on average compared to the schemes with forest canopy, and the fitting accuracy of RH-DFMC-scheme4 improved by 37.5% compared to RH-DFMC-scheme3 in RH-DFMC-scheme without topography. And RH-DFMC-scheme2 improved by 47.1% compared to RH-DFMC-scheme1 in RH-DFMC-scheme with topography; in D-DFMC-scheme without topography, D-DFMC-scheme4 improved the fitting accuracy by 27.8% compared to D-DFMC-scheme3, and in D-DFMC-scheme with topography, D-DFMC-scheme2 improved the fitting accuracy by 32.5% compared to D-DFMC-scheme1. The fitting accuracy of the schemes with topography improved by 123.1% on average compared to the schemes without topography; in RH-DFMC-scheme with forest canopy meteorology, the fitting accuracy of RH-DFMC-scheme1 improved by 112.5% compared to RH-DFMC-scheme3, and in RH-DFMC-scheme with forest canopy meteorology, the fitting accuracy of RH-DFMC-scheme2 improved the fitting accuracy by 127.3% compared to RH-DFMC-scheme4; in D-DFMC-scheme using forest canopy meteorology, D-DFMC-scheme1 improved the fitting accuracy by 122.2% compared to D-DFMC-scheme3, and in D-DFMC-scheme using forest canopy meteorology, D-DFMC-scheme1 improved the accuracy by 122.2% compared to D-DFMC-scheme3. D-DFMC-scheme2 improved the fitting accuracy by 130.4% compared to D-DFMC-scheme4.

3.3. Modeling Factor

We screened the three in-forest meteorological estimation schemes with the best results as the IFRH-scheme1, the IFD-scheme1, and the IFT-scheme1. Except for elevation (EVL), which contributed most to the model of the IFRH-scheme1, the factors that contributed most to the model of the other two in-forest meteorological estimation schemes were the meteorological factors of the canopy. The relative importance (Figure 6) of the three factors contributing most to the model in IFRH-scheme1 were elevation, slope (SLP), and time since the last precipitation event (DSLP) (29.82%, 26.72%, and 20.45%, respectively). Precipitation (P) had the least effect on IFRH-scheme1. Canopy temperature (CT) in IFT-scheme1 contributed much more to the model than the other factors (70.67%), each season had a 24.37% effect, and the rest of the factors had smaller contributions. Canopy vapor pressure deficit (CD), slope, and elevation in IFD-scheme1 had the highest contribution to the model (27.22%, 24.48%, and 19.22%, respectively), and time since last precipitation event had the smallest effect on it.

Overall, D-DFMC-scheme2 offered the best results. In-forest vapor pressure deficit (IFD), slope, and in-forest temperature (IFT) had the highest contribution to the model, 24.99%, 21.43%, and 15.77%, respectively. Pearson’s correlation analysis showed that the understory fine DFMC was significantly correlated (p < 0.001) with slope, in-forest vapor pressure deficit, and aspect (ASP), with correlation coefficients of 0.32, −0.23, and −0.2, respectively. NDVI as well as precipitation had a small effect on the understory fine DFMC, and the correlation was not significant (Figure 7).

4. Discussion

4.1. In-Forest Meteorology

The relationships between in-forest and corresponding remotely sensed calculated canopy temperature, relative humidity, and vapor pressure deficit obtained by deploying field monitoring devices were extrapolated from point to point to obtain in-forest temperature (R² = 0.92), relative humidity (R² = 0.53), and vapor pressure deficit (R² = 0.47) for the study area. The higher R² of the model for the relationship between canopy temperature and in-forest temperature is attributed to the strong direct linear relationship as temperature is directly transferred from the canopy to the in-forest area by thermal radiation. Remote sensing relies on water vapor to calculate vapor pressure deficit and relative humidity. In contrast to temperature, water vapor is affected by surface environmental conditions (e.g., the climatic environments of ridges and valleys differ markedly), but differences in radiative transfer and aerodynamic properties between forest canopies can also lead to changes of in-story meteorological conditions [8,51]. This information is difficult to obtain directly from remote sensing data, and the estimation is likely more difficult, which may be the reason why the R² of the in-forest relative humidity and vapor pressure deficit estimation model was less significant. Other studies found that when topography was introduced into the estimation of in-forest meteorology, the accuracy of in-forest meteorology substantially improved, which suggests that topography has a greater influence on meteorology. The factor contributions also verified this finding. Except for in-forest temperature, the total contribution of the topography factor in the schemes of in-forest vapor pressure deficit and relative humidity was more than 50%. In general, in-forest meteorology is affected slightly differently under topography conditions. For example, the slope affects the angle of the sun’s rays, which in turn affects the radiant energy that the surface receives from the sun; different slope directions affect the time of the sun’s rays, which affects its temperature and relative humidity. Additionally, the complexity of the influence of different topography on atmospheric movements leads to strong spatial variability in localized areas [52,53], which means topography is important in the construction of in-forest meteorological schemes.

4.2. Understory Fine DFMC

Vapor pressure deficit, slope, and aspect showed high contribution to the model and high correlation with understory fine DFMC [54]. However, single factors are not sufficient to uncover their relationships with DFMC, and multiple factors together offer better results (Figure 8).

The in-forest vapor pressure deficit schemes performed better in estimating the understory fine DFMC while the results of the in-forest relative humidity schemes were slightly poorer, so we plotted the required in-forest vapor pressure deficit and in-forest temperature based on D-DFMC-scheme2 (Figure 9) and the distribution of the surface fine DFMC (Figure 10). The maps show that the spatial and temporal variations in forest meteorology and understory fine DFMC were very significant.

The understory fine DFMC decreased as the in-forest vapor pressure deficit became larger. Vapor pressure deficit it is a direct reflection of the driving force in the environment that prompts the evaporation of water from the surface of an object, characterizing the degree of dryness [55]. A larger vapor pressure deficit within the forest indicates that the water content of the air is farther from saturation, and the drier the air is, the stronger the extraction of moisture within the understory dead fuels, resulting in a decline in the understory fine DFMC. Relative humidity characterizes the ratio between saturated and actual water vapor pressure, so even if the relative humidity is the same, its water vapor content will be different. The understory fine DFMC is closely related to its airborne water vapor and vapor pressure deficit, and as a standardized physical quantity, it can be directly compared in different regions and times. The in-forest vapor pressure deficit can better reflect the understory fine DFMC.

In general, the higher the in-forest temperature, the stronger the evapotranspiration and the lower the understory fine DFMC, but as the temperature becomes higher, more water vapor can be held in the environment [56], so the effect of temperature on the understory fine DFMC is not straightforward. This also explains why lower understory fine DFMC were observed in fall and winter rather than summer. The lower the temperature of the season toward winter, the lower the water vapor content that can be held in the atmosphere [57], and the drier the understory fine DFMC. The precipitation is the most direct and rapid factor affecting understory fine DFMC, but it did not show significant correlations (0.06). This may be since the effect of precipitation on understory fine DFMC is greatly affected by local evapotranspiration which has a small effect on the effect of precipitation. The time to the last precipitation event showed strong correlation: the longer the period of time, the greater the effect of the drought on the local area. For example, in 2022, a severe drought occurred in Gannan, which led to frequent fires. This in turn led to direct effect on the understory fine DFMC as well as indirect effects; for example, the longer the drought, the greater the degree of aridity of the soil, which in turn affects the understory fine DFMC, and the understory fine DFMC becomes lower as the distance from the last precipitation becomes longer [35]. NDVI, as the smallest factor affecting the understory fine DFMC, was insignificantly correlated (0.01) and did not show a close relationship with the understory fine DFMC, so it may be insufficient to use this characterization of vegetation cover as one of the factors for the estimation of the understory fine DFMC [35].

Topography factors showed strong correlation with understory fine DFMC [54]. The understory fine DFMC decreased with the slope direction to the south [58]. In addition, the slope direction affects the wind direction and wind speed. The prevalent winds in the Gannan area are southeasterly, and the larger the wind speed, the lower the understory fine DFMC [59]. Slope direction also indirectly affects the correlation between slope and fine DFMC. Some areas have steep slopes, and in general, water in the soil moves toward the slopes that are gentler due to gravity, thus affecting the understory fine DFMC. However, since solar radiation decreases with increasing slope [60], and because many of the sample plots that have steeper slopes are on shaded slopes, the understory fine DFMC is still high. Also, the higher the elevation, the less shielding from solar radiation, the longer the exposure and windier it is, so the DFMC decreases with elevation.

4.3. Significance and Future Research Directions

Currently, fire early warning systems mainly rely on meteorological forecasts, but high fire risk in meteorological forecasts does not necessarily lead to forest fires. It is the understory fine DFMC that directly determines whether forest fuels can ignite and spread into forest fires. Remote sensing has the advantages of large coverage and strong information acquisition, which can provide refined information on the distribution of understory fine DFMC to assist early fire warning. In addition, managers can spot and remove dead fuel to reduce fuel loads. Human activities can also be restricted in low understory fine DFMC areas to better control ignition sources that cause forest fires.

The current remote sensing estimation of DFMC focuses on low and medium resolution images, and there are few estimations for medium and high resolution. We here overcame that limitation by the estimation of understory fine DFMC from medium–high resolution Landsat images. However, Landsat imagery assessment of forest fuel moisture content has the disadvantage of having too long of a revisit cycle. Some satellite images have low spatial resolution but high temporal resolution. Combining high spatial resolution and temporal resolution images will be able to effectively serve the refined mapping of understory fine DFMC.

Despite the progress made, there remains substantial scope for enhancing the accuracy of the model. Several factors that influence the understory fine DFMC, such as solar radiation and the physical and chemical properties of the soil, have not been fully incorporated [28,61]. These factors have been shown to have a significant effect on the understory fine DFMC, and their inclusion would likely improve the model’s precision. Furthermore, the heterogeneity of ecological environments, combined with the inherent limitations of empirical machine learning models like random forest, can result in reduced model applicability across diverse topography, biomes, and vegetation types. Our study is primarily focused on low mountains and hilly areas, where the topographical variability has a more pronounced impact on in-forest meteorological conditions and the understory fine DFMC compared to flatter topography. Thus, the role of topography must be re-evaluated for different land forms.

Additionally, the climatic differences between the Mediterranean (wet winters and dry summers) and the subtropical monsoon climate of our study area introduce variations in seasonal effects, which must be accounted for when developing the model. The diversity of vegetation types, characterized by differing heights, densities, and species, also plays a crucial role in influencing in-forest meteorological conditions and the understory fine DFMC [62,63]. As a result, the model, based on the methodology outlined in this study, should be adapted to the specific characteristics of different regions to improve its applicability.

5. Conclusions

We investigated the feasibility of estimating understory fine DFMC based on medium- and high-resolution remote sensing. With the assistance of multiple factors (topography, meteorology, vegetation, etc.), the estimation of understory fine DFMC based on meteorological parameters calculated from Landsat data showed significant results. Moreover, the accuracy of the understory DFMC estimation model was significantly improved by using the estimated meteorological factors in the forest and introducing the topography factors, considering the differences in meteorology between the forest and the canopy. The mapping of understory fine DFMC can serve the regional forest fire danger warning work and provide decision support for the management of dead fuels and fire source management.