Estimation of NPP in Huangshan District Based on Deep Learning and CASA Model

Abstract

Net primary productivity (NPP) is a key indicator of the health of forest ecosystems that offers important information about the net carbon sequestration capacity of these systems. Precise assessment of NPP is crucial for measuring carbon fixation and assessing the general well-being of forest ecosystems. Due to the distinct ecological characteristics of various forest types, accurately understanding and delineating the distribution of these types is crucial for studying NPP. Therefore, an accurate forest-type classification is necessary prior to NPP calculation to ensure the accuracy and reliability of the research findings. This study introduced deep learning technology and constructed an HRNet-CASA framework that integrates the HRNet deep learning model and the CASA model to achieve accurate estimation of forest NPP in Huangshan District, Huangshan City, Anhui Province. Firstly, based on VHR remote sensing images, we utilized the HRNet to classify the study area into six forest types and obtained the forest type distribution map of the study area. Then, combined with climate data and forest type distribution data, the CASA model was used to estimate the NPP of forest types in the study area, and the comparison with the field data proved that the HRNet-CASA framework simulated the NPP of the study area well. The experimental findings show that the HRNet-CASA framework offers a novel approach to precise forest NPP estimation. Introducing deep learning technology not only enables precise classification of forest types but also allows for accurate estimation of NPP for different types of forests. This provides a more effective tool for forest ecological research and environmental protection.

1. Introduction

Nowadays, global warming has become an issue that cannot be ignored. With the acceleration of urbanization, more land is being cultivated, and more forests are being cut down, leading to a rapid increase in carbon dioxide emissions [1]. In view of the above situation, the urgent task is to alleviate environmental pollution and take measures to monitor and reduce carbon emissions [2]. As the largest ecosystem on land, forest plays an important role in maintaining the stability of the biosphere and improving the ecological environment by regulating the climate and maintaining biodiversity [3,4]. However, due to human activities and natural disasters, as well as climate change, the area of forests has been shrinking. This has led to a continuous weakening of the functioning of forest ecosystems [5,6]. Therefore, it is crucial to adopt strict monitoring and preservation of forest ecosystems in order to preserve the stability and sustainable growth of the global environment.

Net primary productivity(NPP) is defined as the total amount of organic matter produced by photosynthesis per unit area at a given time after subtracting the organic matter consumed by its autotrophic respiration [7]. As a fundamental ecological variable, NPP can reflect the growth status of plants and simulate the amount of carbon fixation in vegetation [8,9]. This is important for determining the ecosystem quality status. By estimating the NPP of forest vegetation, it is possible to quantitatively assess the sequestration capacity of forest ecosystems as well as to provide a scientific basis for sustainable development and the rational use of natural resources [10].

For the estimation of NPP, traditional measurement methods include direct harvesting and chlorophyll determination, but they severely damage the vegetation. Additionally, the complex terrain and limitations of measurement methods make it difficult to conduct measurements on a large scale, requiring substantial time, resources, and manpower [11,12]. With the continuous development of science and technology, it has become a mainstream trend to calculate NPP through relevant models [13]. According to the model mechanism and structure, the estimation of NPP models mainly includes the following three types: statistical models; process models; and light-use efficiency models [14]. The research principles of the three models are quite different, so the advantages and disadvantages are also different. Statistical models estimate NPP based on the correlation between a large amount of climatic data and the measured data of the sample plots. The advantage of this method lies in its simple structure and easy parameter acquisition. However, due to the lack of consideration for factors such as changes in vegetation itself and site conditions, the estimated results are only approximate and cannot accurately estimate the change in NPP [12,15]. Process models simulate the processes of photosynthesis, respiration, and soil water loss of forest vegetation based on rich ecological theories and clear biological mechanisms and simulate the NPP in large-scale areas in combination with climate change. However, they also have the shortcomings of having many parameters that are difficult to obtain and convert to regional scales [16,17]. The principle of the light-use efficiency model is based on the effective conversion and utilization of solar radiation by vegetation, using vegetation index, photosynthetic efficiency, and photosynthetically active radiation as the basis for estimating NPP [18,19]. Among the numerous models for photosynthetic efficiency, the Carnegie–Ames–Stanford Approach(CASA) model is the most widely applied and representative one [20]. The CASA model utilizes remote sensing technology to acquire remote sensing data of the research area. It processes complex environmental and rugged terrain images on a pixel-by-pixel basis, overcoming the spatial and temporal limitations of traditional field measurement methods. This enhances research efficiency and reduces errors [21,22]. Xiao et al.estimated the NPP of the Yellow River Basin for the 20-year period from 2001 to 2020 based on an improved CASA model using MODIS remote sensing data and climate data and compared them with MODIS NPP data and found that the accuracy of CASA model estimation was satisfactory in estimating changes in NPP within the study area [23]. Zhou et al. conducted a study in Yangzhou, China, using an improved CASA model and geographical detector analysis to investigate the impact of climate, social, and ecological factors on the variation in NPP in the study area [24].

When calculating NPP, it is necessary to consider the unique characteristics of different types of vegetation, such as leaf structure, photosynthesis rate, growth cycle, etc. This can vary depending on the type of vegetation and affect the estimation of NPP [25]. If vegetation classification maps with low accuracy are used, some important vegetation types may be overlooked, resulting in biased estimation results. High-accuracy vegetation classification maps can provide more ecological information, such as species composition and community structure. These are of great significance for understanding the functioning and dynamics of ecosystems [26,27]. To conduct a more precise study of the NPP in the experimental area, it is necessary to use a high-precision vegetation classification map. With the development of remote sensing technology, vegetation classification maps are mostly derived from remote sensing image processing, and the current processing methods mainly include traditional image segmentation methods and deep learning methods [28]. The traditional method of image segmentation is a technique that divides the image into multiple different and non-overlapping subsets based on intrinsic factors such as grayscale, brightness, texture, and geometric features. It has the advantages of high segmentation efficiency and simple calculation. However, it has certain limitations in practical application, such as requiring a large amount of manpower and time investment and being susceptible to noise interference. Therefore, its utility needs to be improved [29,30]. Recent years have seen a deepening of deep learning research in the field of image classification due to the constant improvement in GPU performance. Deep learning has gradually taken up an important position in the variety of computer vision applications, particularly attaining noteworthy results in the semantic segmentation of remote sensing images [31]. Meanwhile, deep learning algorithms have also been widely used in the field of forestry. He et al. classified the forest types of Qingyuan County based on Sentinel-2 images using the ResNet algorithm and achieved a high validation accuracy of 87.90% [32]. Lee et al. used U-Net and SegNet algorithms to classify landscapes affected by human deforestation in North Gyeongsang Province, and the overall accuracy of the U-Net model was 74.8%, which was 11.5% higher than that of SegNet (63.3%) [33]. Wu et al. constructed semantic segmentation data of economic fruit forests based on UAV multispectral remote sensing images and proposed an improved ISDU-Net model with recognition accuracies in terms of pixel accuracy, average intersection ratio, frequency weight intersection ratio, and Kappa coefficients of 87.73%, 70.68%, 78.69%, and 0.84%, respectively [34].

Therefore, this study estimated the forest NPP in the Huangshan area by constructing an integrated HRNet-CASA framework. This framework combines the HRNet deep learning model with the CASA model. Firstly, based on prior experiments, we constructed a dataset for forest-type classification on the basis of VHR images through visual interpretation. HRNet was then used to classify the forest types in the study area. Secondly, based on the results of the classification of forest types, the NPP of the forests in the study area was estimated by using the CASA model through the coupling of the climatic data with the solar radiation data and by coupling the climatic data with the solar radiation data. The CASA model was used to estimate the forest NPP in the study area. During validation, this study compared the field data of the sample plots with the simulated values to ensure the reliability of the final results.

2. Materials

2.1. Study Area

The study area is located in the Huangshan District of Huangshan City in the south of Anhui Province, with geographic coordinates between 117°50′–118°21′ E and 30°00′–30°32′ N. The climate of Huangshan District is a subtropical monsoon humid climate, with four distinct seasons, abundant rainfall, an average annual temperature of 15.4 °C, and an annual precipitation of 1500–1600 mm, mostly concentrated in the spring and summer [35]. The primary types of forests found in Huangshan District include mixed coniferous and broad-leaved forests, bamboo forests, shrub forests, and broad-leaved forests. The district’s total forest area is 133,697.46 square hectometers (hm²), with a coverage rate of 78.69% and a forest stock of 8,688,119 m³. The region has a great ecological environment and is abundant in natural resources. However, long-term human activities have caused issues like an uneven distribution of forest resources and a monophyletic forest structure, which limit the area’s forests’ ability to develop sustainably.

2.2. Data Sources

2.2.1. Remote Sensing Data

This study constructed a training and testing dataset for forest-type classification based on Very High-Resolution (VHR) remote sensing images. Given the extensive coverage of 133,697.46 hm² and the terrain distribution features of Huangshan District, it was that the research area should be split into six regions to facilitate the gathering of complete forest information by aerial remote sensing data. A total flight duration of 18.5 h was conducted using a Cessna 208B aircraft, chosen for its suitability as a fixed-wing manned platform for this data collection endeavor. The specific flight parameters are detailed in

Table 1. Flight Parameters.

Regions	1	2	3	4	5	6
Flight speed (km/h)	260	260	260	240	240	240
Flight altitude (m)	2140	2190	2300	1840	1930	2050
Datum elevation (m)	80	130	240	110	200	320

.

The VHR images were captured using the IXU-RS1000 digital camera (manufactured by Phase One, Frederiksberg, Denmark), with its specific parameters also detailed in

Table 2. IXU-RS1000 Digital Camera Parameters.

IXU-RS1000 Digital Camera Parameters
Heading overlap (%)	65
Side-by-side overlap (%)	20
Ground resolution (m)	0.19
Field of view °	56.2 × 43.7
pixels	11,608 × 8708

. The collected images included three bands: red; green; and blue. The processing of the images required the following steps: First, check the completeness of the image data to ensure that the heading and side-by-side overlaps are consistent with the pre-flight planning design. Second, perform alignment operations on the digital orthophoto images by matching feature points between adjacent images and optimizing the calibration to obtain the stitched images that meet the requirements.

This study used Sentinel-2 remote sensing images from January 2018 to December 2018 to obtain the NDVI data needed for the CASA model. The data were obtained by calculating the red and near-infrared bands of the images to obtain the NDVI values, and the maximum NDVI value for each month was extracted using the maximum synthesis method to reduce the images of clouds, atmosphere, etc.

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}\mathrm{D}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{E}\mathrm{D}}}$$

(1)

Meteorological data were used from meteorological stations in and around the study area in 2018, which were obtained from the China Science Data Centre (https://data.cma.cn/metadata/#/summary, accessed on 15 August 2023), including precipitation, average temperature, and solar radiation. According to a priori experiments, these meteorological data were processed into raster images by inverse distance weighted (IDW) interpolation. All the data were finally unified to 25.80 m grid accuracy by resampling.

2.2.2. Field Data

The sample data were obtained between October and November of 2018. There were 303 sample plots in total, comprising 168 broad-leaved forest sample plots, 50 coniferous forest sample plots, 20 coniferous and broad mixed forest sample plots, 41 bamboo sample plots, and 24 shrub forest sample plots surveyed (Figure 1). Each individual sample plot measured 25.8 m by 25.8 m (area 0.067 hm²) square. Sample attributes include forest type, average diameter at breast height, average tree height, degree of depression, species composition, and other survey factors. The NPP field data for validation could be calculated from the sample data.

3. Methods

3.1. Classification of Forest Types Based on the HRET Algorithm

3.1.1. HRNET Algorithm

Deep learning provides a wide range of applications in the disciplines of image identification, target detection, and semantic segmentation. However, with the growth of image resolution, it has become an academic issue to effectively process detailed information in high-resolution images while maintaining computational efficiency. To address this problem, SUN et al. proposed HRNet (high-resolution deep neural network) in 2019, which has demonstrated promising results in the application in the field of image recognition [36]. Unlike the majority of deep learning networks, which typically have a sequential structure, HRNet employs a distinctive parallel structure that enables it to preserve high resolution in the feature layer throughout training. This allows HRNet to better retain detailed information when processing images, especially in the semantic recognition of high-resolution images. This characteristic provides significant advantages for image processing tasks.

The deep learning model used in this study is HRNet-V2, as illustrated in Figure 2. This model begins with a high-resolution subnet in the first stage and progressively adds low-resolution subnets over four stages. By employing multi-resolution group convolution and multi-resolution convolution, the network integrates data from feature maps of different resolutions in the second, third, and fourth stages. Multi-resolution group convolution refers to dividing the input channels into multiple channel subsets and performing regular convolutions on each subset at different spatial resolutions. Multi-resolution convolution fully connects the input and output subsets, with each connection being a regular convolution, and the output of each output channel subset is the sum of the convolution outputs from each input channel subset.

In the output section, HRNet-V2 differs from the traditional HRNet network. It upsamples the low-resolution subsets to the same resolution as the high-resolution subset through bilinear interpolation and then fuses them with the high-resolution subset to obtain high-resolution results [37]. This improvement adds only a small amount of computation in the output stage but fully leverages the advantages of multi-resolution convolution, resulting in better performance in semantic segmentation applications for HRNet-V2.

3.1.2. Construction of a Semantic Segmentation Sample Datasets for Tree Species

In this study, based on the real vegetation situation in the study area, the main forest types were the broad-leaved forest, the coniferous forest, the mixed forest, the shrub forest, and the bamboo forest (Figure 3). The sample dataset for semantic segmentation of tree species was constructed based on 0.2 m resolution remote sensing image data and sample plot data. The image labels were obtained by combining the field data with the visual interpretation of the remote-sensing images. It is required to segment the entire scene picture into image chunks before doing deep learning since the vast size of the complete scene image would cause a memory overflow issue if it is fed straight into the deep learning network. To achieve accurate correspondence cropping between the original image and the raster labels, a regular grid algorithm was constructed using the Python language to crop the image accurately and empirically cut the image into small chunks with a size of 300 × 300. After trimming and visual interpretation, a total of 11,375 images and their corresponding image label graphics were obtained, with 10,237 images used for model training and 1138 images used for model validation.

3.1.3. Accuracy Evaluation of Forest Types Classification

The validation data in this study were chosen by random sampling to ensure the non-differential nature of the experimental results, as it is challenging to ensure the accuracy and reliability of the classification results for the entire study area if relying solely on the validation set for validation. The actual forest classes where the random validation points were located were visually deciphered and compared with those classified by the HRNet network. The confusion matrix was calculated to measure the classification accuracy and filter the most appropriate classification model. This process involved creating 1000 random validation points and combining them with the field survey dataset. Overall accuracy (OA), Producer’s Accuracy(PA), User’s Accuracy(UA), F1-score(F1), and Kappa coefficient were used as evaluation parameters to indicate the accuracy of classification. They are calculated as follows:

$$\mathrm{O}\mathrm{A}={\displaystyle \frac{\mathrm{T}\mathrm{N}+\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{N}+\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}+\mathrm{F}\mathrm{P}}}$$

(2)

$$\mathrm{P}\mathrm{A}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(3)

$$\mathrm{U}\mathrm{A}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(4)

$$\mathrm{F}1={\displaystyle \frac{2\mathrm{T}\mathrm{P}}{2\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(5)

where TN refers to the negative sample that is predicted to be a negative sample. TP refers to the positive sample that is predicted to be a positive sample. FP refers to the negative sample that is predicted to be a positive sample. FN refers to the positive sample that is predicted to be a negative sample.

The Kappa coefficient is a ratio that indicates the proportion of the reduction in errors produced by a classification versus a completely random one and is a more desirable measure of the classification result, which is calculated as follows:

$$\mathrm{K}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{a}={\displaystyle \frac{\mathrm{O}\mathrm{A}-{\mathrm{p}}_{\mathrm{e}}}{1-{\mathrm{p}}_{\mathrm{e}}}}$$

(6)

$${\mathrm{p}}_{\mathrm{e}}={\displaystyle \frac{{\mathrm{a}}_1·{\mathrm{b}}_1+{\mathrm{a}}_2·{\mathrm{b}}_2+\cdots +{\mathrm{a}}_{\mathrm{n}}·{\mathrm{b}}_{\mathrm{n}}}{{\mathrm{n}}^2}}$$

(7)

where OA denotes the overall classification accuracy; p_e denotes the expected agreement rate; n denotes the total number of samples; a denotes the actual number of categories; and b denotes the number of correctly predicted categories.

3.2. NPP Estimation Model

The CASA model, which calculates the NPP based on a variety of data types, such as average monthly temperature, monthly precipitation, solar radiation, and remote sensing image data, was employed in this work [38]. Its calculation formula is given below:

$$\mathrm{N}\mathrm{P}\mathrm{P}\left(\mathrm{x},\mathrm{t}\right)=\mathrm{A}\mathrm{P}\mathrm{A}\mathrm{R}\left(\mathrm{x},\mathrm{t}\right)\times \mathsf{\epsilon }\left(\mathrm{x},\mathrm{t}\right)$$

(8)

where NPP (x, t) is the NPP at image x in the month of t (gC/m²); APAR (x, t) represents the photosynthetically active radiation absorbed by image x in the month of t (MJ/m²); $\mathsf{\epsilon }\left(\mathrm{x},\mathrm{t}\right)$ represents the actual light utilization efficiency (LUE, gC/MJ) of the vegetation.

3.2.1. Actual Light Energy Utilization Estimation

The light energy utilization rate is the ratio of the chemical energy contained in the dry matter produced per unit area at a given time to the photosynthetically active energy and is mainly influenced by temperature, moisture, and vegetation type [39]. The actual light energy utilization rate in this study can be expressed as follows:

$$\mathsf{\epsilon }(\mathrm{x},\mathrm{t})={\mathrm{T}}_1(\mathrm{x},\mathrm{t})\times {\mathrm{T}}_2(\mathrm{x},\mathrm{t})\times {\mathrm{W}}_{\mathsf{\epsilon }}(\mathrm{x},\mathrm{t})\times {\mathsf{\epsilon }}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$

(9)

$${\mathrm{T}}_1(\mathrm{x},\mathrm{t})=0.8+0.02\times {\mathrm{T}}_{\mathrm{o}\mathrm{p}\mathrm{t}}(\mathrm{x})-0.0005\times {{\mathrm{T}}_{\mathrm{o}\mathrm{p}\mathrm{t}}\left(\mathrm{x}\right)}^2$$

(10)

$${\mathrm{T}}_2(\mathrm{x},\mathrm{t})={\displaystyle \frac{1.184}{1+{\mathrm{e}}^{0.2\times \left[{\mathrm{T}}_{\mathrm{o}\mathrm{p}\mathrm{t}}\left(\mathrm{x}\right)-10-\mathrm{T}(\mathrm{x},\mathrm{t})\right]}}}\times {\displaystyle \frac{1}{1+{\mathrm{e}}^{0.3\times \left[{\mathrm{T}(\mathrm{x},\mathrm{t})-\mathrm{T}}_{\mathrm{o}\mathrm{p}\mathrm{t}}\left(\mathrm{x}\right)-10\right]}}}$$

(11)

$${\mathrm{W}}_{\mathsf{\epsilon }}(\mathrm{x},\mathrm{t})=0.5+0.5\times {\displaystyle \frac{\mathrm{E}\mathrm{E}\mathrm{T}(\mathrm{x},\mathrm{t})}{\mathrm{P}\mathrm{E}\mathrm{P}(\mathrm{x},\mathrm{t})}}$$

(12)

where T₁, T₂ denote the stress coefficients of maximum temperature (T_max) and minimum temperature (T_min) on the actual light energy utilization; ${\mathrm{W}}_{\mathsf{\epsilon }}(\mathrm{x},\mathrm{t})$ denotes the coefficient of water stress; T(x) denotes the optimum temperature for vegetation growth; EET denotes the actual regional evapotranspiration.

${\mathsf{\epsilon }}_{\mathrm{m}\mathrm{a}\mathrm{x}}$ refers to the utilization rate of photosynthetically active radiation by vegetation under ideal conditions, which is a physiological property of vegetation itself. Using remote sensing data, meteorological data, and vegetation NPP measured data, Zhu et al. conducted a systematic simulation of the maximum light utilization rate of different vegetation to obtain the maximum light energy utilization rate of vegetation types under different classification accuracies [40].

3.2.2. APAR Estimation

APAR is the photosynthetically active radiation absorbed by the vegetation within the image element. It is determined by the ratio of the total photosynthetically active radiation received by the image element and the absorption of photosynthetically active radiation by the vegetation. It is calculated using the following formula:

$$\mathrm{A}\mathrm{P}\mathrm{A}\mathrm{R}=\mathrm{S}\mathrm{O}\mathrm{L}\left(\mathrm{x},\mathrm{t}\right)\times \mathrm{F}\mathrm{P}\mathrm{A}\mathrm{R}(\mathrm{x},\mathrm{t})\times 0.5$$

(13)

where SOL(x, t) represents the total solar radiation (MJ/m²) of image x in month t; FPAR(x, t) represents the absorption ratio of photosynthetically active radiation of image x in time period t; the constant 0.5 represents the proportion of total solar radiation that can be utilized by the vegetation. The detailed calculation steps of these parameters are described in the previous study [41].

3.3. Calculation of Forest NPP Based on Field Data

Forest NPP is composed of community growth and annual litter, and the forest NPP of each sample plot was obtained based on the relationship between biomass and stocking of different forest types and the functional relationship between biomass and community growth and annual withering [42]. The calculation formula of the specific process is shown in

Table 3. Relationship between forest volume, biomass, and NPP of typical tree species.

Forest Type	Relationship between Biomass and Volume	Relationship between Biomass and Community Growth	Relationship between Biomass and Annual Withering
Mixed forest	$\mathrm{B}=\mathrm{V}/(0.011731+0.0018\ast \mathrm{V})$	$\mathrm{Y}=\mathrm{B}/(0.001038\ast \mathrm{A}+0.000761\ast \mathrm{B})$	$\mathrm{L}=346$
Broad-leaved forest	$\mathrm{B}=\mathrm{V}/(0.005788+0.000020\ast \mathrm{V})$	$\mathrm{Y}=\mathrm{B}/(0.003018\ast \mathrm{A}+0.000331\ast \mathrm{B})$	$\mathrm{L}=\mathrm{B}/(0.091028+0.000575\ast \mathrm{B})$
Coniferous forest	$\mathrm{B}=\mathrm{V}/(0.012917+0.000022\ast \mathrm{V})$	$\mathrm{Y}=\mathrm{B}/(0.004598\ast \mathrm{A}+0.000069\ast \mathrm{B})$	$\mathrm{L}=\mathrm{B}/(0.101320+0.000874\ast \mathrm{B})$

where B is biomass in g/m²; V is volume in m³; Y is community growth gC/m²; L is annual withering in gC/m²; A is average age.

.

In estimating the NPP of bamboo forests, the sample trees in the survey sample plots were dominated by moso bamboo. The biomass of moso bamboo was first calculated according to the moso bamboo aboveground biomass model, and then the NPP of bamboo forests was calculated using the plant dieback model established by Zhang [43]. The formulas are as follows:

$$\mathrm{W}=0.1959\times {\mathrm{D}}^{1.5761}\times {\mathrm{H}}^{0.3612}\times \mathrm{N}$$

(14)

$$\mathrm{N}\mathrm{P}\mathrm{P}=\mathrm{W}÷0.565$$

(15)

where W is the aboveground biomass of bamboo; D is the diameter at breast height (cm); H is the height of bamboo (cm); N is the number of bamboo forest stocks.

3.4. Evaluation of NPP Estimation Accuracy

This study focuses on model accuracy evaluation by calculating the coefficient of determination (R²) and mean absolute error (MAE) of the model. The calculation formulas are as follows:

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{\mathrm{n}}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\left|{\mathrm{y}}_{\mathrm{i}}-{\mathrm{x}}_{\mathrm{i}}\right|$$

(16)

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-{\mathrm{x}}_{\mathrm{i}}\right)}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-\overline{{\mathrm{y}}_{\mathrm{i}}}\right)}^2}}$$

(17)

where ${\mathrm{y}}_{\mathrm{i}}$ is the actual observed value; ${\mathrm{x}}_{\mathrm{i}}$ is the model estimate; $\overline{{\mathrm{y}}_{\mathrm{i}}}$ is the mean of the actual values; $\mathrm{n}$ is the sample size.

3.5. Framework of the Analysis

Figure 4 shows the workflow of the current analysis.

4. Results

4.1. Validation of Forest Types Classification

Based on the HRNet algorithm, the remote-sensing images of the study area were classified. The confusion matrix heat map is shown in Figure 5.

Table 4. Results of HRNet classification.

Forest-Type	PA (%)	F1-Score	OA (%)	Kappa Coefficient
Bamboo	82.50	0.8354	80.60	0.7351
Coniferous	80.29	0.7612
Broad-leaved	86.89	0.8679
Mixed	55.93	0.6346
Shrub	84.51	0.6250
Non-forest	69.33	0.8159

shows that the overall classification accuracy was 80.60%, and the Kappa coefficient was 0.7351, indicating that the model’s classification results were generally regarded as trustworthy in the study area. Among them, the producer accuracy of broad-leaved forests was 86.89%, and the F1-score was 0.8679, which was the highest among all forest types, indicating that the model performed well in identifying broad-leaved forests, followed by bamboo forest and coniferous forests, with producer accuracies of 82.50% and 80.29%, respectively, and F1-scores of 0.8354 and 0.7612, which were slightly lower compared to broadleaved forests, but still showed high classification accuracy. The producer accuracy of shrub forests was 84.51%, while the F1 score was 0.6250. With an F1-score of 0.82, the producer accuracy for non-forested land was 69.33%. The mixed forest’s producer accuracy was 55.93%, with an F1 score of 0.6346. There is some degree of misclassification during the classification process between mixed and broad-leaved forests; the percentage of misclassified mixed forests as broad-leaved forests is 35.59%. Additionally, there is a percentage of 22.70% misclassified non-forested land as shrub forests.

4.2. Forest Types Classification Results

According to the accuracy test results, it is proved that the classification accuracy achieved by using the HRNet deep learning model for forest classification in Huangshan District performs well. Based on the above results, we derived an accurate and reliable forest-type distribution map of Huangshan District (Figure 6).

According to the forest type distribution map, it is evident that the majority of forests in the study area consist of broad-leaved forests, coniferous forests, and mixed forests. These are predominantly located in the northwestern and southeastern parts of the study area. Shrub forests are primarily found in the northwestern and central parts of the study area. Bamboo forests, on the other hand, have a more scattered and extensive distribution.

In terms of the percentage of area of different types of forests, the total area of forests in the study area is 147,440.35 hm², of which broadleaved forests are 76,879.41 hm², accounting for 52.14%; shrub forests are 27,815.40 hm², accounting for 18.87%; bamboo forests are 19,942.73 hm², accounting for 13.53%; coniferous forests are 17,462.04 hm², accounting for 11.84%, and mixed forests are 5340.77 hm², accounting for 3.62%.

4.3. NPP Simulation Results

In this study, the forest classification accuracy is 80.6%, so the maximum light energy use rate of vegetation with 80% classification accuracy is used (

Table 5. Maximum light energy utilization values of typical Chinese vegetation at 80% classification accuracy.

Vegetation Type	εmax (gC MJ−1)
Broad-leaved forest	0.808
Coniferous forest	0.406
Mixed forests	0.484
Bamboo forest	0.577
Shrub	0.448

).

The simulations were carried out for the study area, and the results and distribution maps of the forest NPP were obtained as follows.

Table 6. Results of NPP simulation.

Forest Type	NPP Mean (gCm−2a−1)	NPP MAX (gCm−2a−1)	NPP Min (gCm−2a−1)
Broad-leaved	1025.94	1539.54	240.04
Mixed	867.99	1275.14	154.95
Coniferous	512.97	710.33	132.61
Bamboo	788.85	986.05	268.91
Shrub	440.21	647.48	127.06

showed that the mean NPP in the study area was 784.49 gCm⁻²a⁻¹, of which 1025.94 gCm⁻²a⁻¹ was for broadleaved forests, 867.99 gCm⁻²a⁻¹ for mixed forests, 512.97 gCm⁻²a⁻¹ for coniferous forests, 788.85 gCm⁻²a⁻¹ for bamboo forests, 440.21 gCm⁻²a⁻¹ for shrub forests.

We divided the vegetation in the study area into five intervals based on the size of NPP using the natural break point method: low-value zone (NPP less than 300); lower-value zone (NPP between 300–600); medium-value zone (NPP between 600–900); higher-value zone (NPP between 900–1200); and high-value zone (NPP greater than 1200). From the NPP distribution map (Figure 7), it can be seen that the low-value zone is mainly distributed in the northwestern part of the study area; the high-value zone is mainly distributed in the southeastern part of the study area as well as in the western part of the study area, and three remaining intervals are more widely distributed, encompassing the majority of the study area.

4.4. Validation of NPP Simulation

It is required to compare the simulated data from the CASA model with the actual values of vegetation NPP in order to confirm the validity of the forest NPP estimations. The following results were attained in this study after comparing the measured data of the sample plots with the simulated NPP values (

Table 7. Comparison of NPP simulation and field data.

	Broad-Leaved	Mixed	Coniferous	Bamboo
Field data (gCm−2a−1)	943.97	1001.39	673.31	772.83
NPP simulation (gCm−2a−1)	994.94	927.82	665.75	748.97

).

The simulated NPP was compared with the field data, and the results are shown in Figure 8. Among them, the simulation of NPP in coniferous forest performed the best, with the model R² value of 0.7501 and MAE of 100.21 gCm⁻²a⁻¹, and the simulated results were very close to the measured values. Following this was the prediction for mixed forests, with an R² value of 0.7415 and MAE of 95.72 gCm⁻²a⁻¹, indicating that the model had high predictive accuracy when simulating NPP in complex mixed forests with diverse tree species composition. The NPP simulation effect of broad-leaved forest was also satisfying, with the model R² value of 0.7202 and MAE of 134.88 gCm⁻²a⁻¹, demonstrating reliable results for broad-leaved forest NPP simulation. Accurate simulation of the NPP of broad-leaved forest is particularly significant due to its large proportion in the study area and its impact on understanding the carbon cycle within the ecosystems. However, the model performance was slightly insufficient in the NPP simulation of bamboo forest, with an R² value of 0.6397 and an MAE of 143.65 gCm⁻²a⁻¹, which is a lower simulation accuracy compared with other forest types. As a special vegetation type, the growth pattern and ecosystem structure of bamboo forests differ from those of coniferous forests, mixed forests, and broad-leaved forests, which may be one of the reasons for the slightly lower simulation effect of the model. In terms of the measurement of shrub forests, due to the lack of means of field measurement, we compared the results of this research with previous studies and found that the simulated shrub forests NPP in this study was 440.21 gCm⁻²a⁻¹, while the NPP calculated in previous studies was 420.47 gCm⁻²a⁻¹ [44]. This may be due to the fact that the study area has a subtropical monsoon climate with abundant light as well as precipitation, and the NPP of the shrub forests here is much greater than elsewhere.

It should be noted that in the estimation of NPP in broadleaved forests, there were some low-value areas where the simulated values were significantly higher than the real values. This phenomenon could potentially be attributed to the distinct ecological characteristics of broadleaved forests, such as their tall tree morphology and expansive canopy. These characteristics may cause an overestimation of vegetation cover in remote sensing imagery, resulting in a significant deviation between the simulated and actual values.

Overall, the HRNet-casa framework considers the diversity of forest types when simulating vegetation in the study area. The simulated NPP values obtained show a good fit with the actual measured values from field samples, providing a scientific basis for future ecological environment protection in this region. At the same time, we ought to observe that while the framework demonstrates good simulation performance in most cases, there are some specific plant kinds or climatic variables that might further enhance the model’s validity and accuracy.

5. Discussion

In the subject of environmental research, NPP estimates are crucial. Nonetheless, the bulk of NPP research to date has been on standardizing diverse forest types and categorizing distinct land use patterns. This calls for the use of standardized parameters and calculating methods. A more refined research methodology must be employed in order to compute the NPP of distinct forest types more precisely, and the first task is to accurately categorize different types of forests and then calculate the NPP for each type of forest on this basis. As a result, this study proposes the HRNnet-CASA framework, which uses the HRNet algorithm to classify the forest types in the study area first. The CASA model then uses the reliable forest-type distribution maps that were obtained to estimate the NPP of the forests in the study area quantitatively. Additionally, the majority of the validation process for the current NPP data depends on comparing relevant MODIS products. However, in the validation phase of this research, we used the field data collected between October and November 2018 as the validation data. Through parameters such as tree height and breast diameter, we calculated the true values of forest NPP. Since the forests in this area are not in the juvenile stages, their annual growth rates are relatively slow, and their annual variations in breast diameter and tree height parameters are small, so the field data collected in different seasons in the same area are little changed. The annual NPP values calculated based on these field data can be regarded as reliable true values.

5.1. Classification of Forest Types

When it came to the stage of classifying forest types, this research conducted a more detailed division based on the actual composition of forest types in the study area. This included broadleaved forests, coniferous forests, mixed forests, bamboo forests, shrub forests, and non-forested land. Finally, a distribution map of the different forest types in the study area was created. Broadleaved forests had the highest classification accuracy of the five different types of forests. This may be attributed to the fact that broadleaved forests, as the predominant forest type in the Huangshan District, possess a distinct advantage in terms of quantity over other forest types. Moreover, they carry greater weight in the dataset construction process and typically involve larger sample sizes, which can enhance the accuracy of the final classification results [45]. Additionally, broadleaved forests are more prominent in high-precision remote sensing images, resulting in more accurate classification results based on visual interpretation of these forests. In contrast, mixed forests had a relatively low classification accuracy. This phenomenon might be explained by the fact that mixed forests are made up of different kinds of trees. Due to the close connection to coniferous and broadleaved forests in terms of tree height and canopy form, mixed forests have a vegetation pattern that makes it challenging to distinguish them from other types of forests. Simultaneously, the proportion of different tree species in the mixed forest also increased the difficulty of classification [46,47]. Following validation, the overall accuracy of this classification reached 80.60%, demonstrating HRNet’s strong classification performance in the research domain—particularly when it comes to classifying shrub and broad-leaved forests. On the other hand, there was still potential for increasing the categorization accuracy for mixed forests and non-forested terrain. Considering the relatively small proportion of non-forested land and mixed forests in the study area of this experiment, the impact of classification error on the overall results is relatively limited. In conclusion, this study’s classification results are quite trustworthy, demonstrating the usefulness of HRNet for classifying the forest types in the research region and its ability to supply precise data support for additional NPP estimation.

5.2. Study Area NPP Simulation

The simulation step of this work was based on the CASA model and involves combining data on radiation, climate, NDVI, and forest type distribution to calculate the NPP of the different forest types in the Huangshan District. Broad-leaved forests had the highest mean NPP among them, while shrub forests had the lowest. The low-value area of NPP is primarily distributed in the northwest of the study area, as can be seen from the spatial distribution map. This might be a result of the increased frequency of human activity in the region, which causes a relatively sparse distribution of forests and a decreased ecological function, both of which lower the NPP of the vegetation. On the other hand, the high-value area is mostly found in the southeastern and the western local areas. These regions have less forest cover and are less impacted by human activity than the western part of the research area. The majority of the high-value areas are found in the west and southeast of the study area, where there is more forest cover and less anthropogenic disturbance, leading to higher NPP.

Previous research indicates a close correlation between NPP and topographic factors, with slope exerting a particularly significant influence on NPP [48,49]. Therefore, we compared the NPP values of the study area with the slope values, and one thousand random points were selected to analyze the NPP values with the slope values. The result indicated that the NPP value is significantly influenced by slope, with a Pearson correlation coefficient of 0.42. This indicated a positive correlation between NPP value and slope, meaning that areas with steeper slopes tend to have higher NPP values. According to previous research, the occurrence of this phenomenon is closely related to human activities [50]. It is more difficult to develop and utilize areas with larger slopes, and people prefer to carry out farming and other activities in areas with flat terrain, which will make areas with large slopes less affected by human activities, and the forest ecosystem can be preserved more intact, thus ensuring a stable output of NPP.

5.3. Shortcomings and Prospects

In this study, we proposed the HRNet-CASA framework and explored its feasibility in forest-type classification and NPP estimation. During the research process, we found that there was a phenomenon of misclassifying farmland as shrubland. This could be attributed to the fact that the VHR remote sensing image used in this study is a single temporal image containing only three bands: red; green; and blue. These three bands are insufficient to adequately depict the characteristics of shrubland and arable land [51].

For future research, it is recommended to consider introducing multi-temporal multispectral remote sensing images. Multi-temporal remote sensing images can capture vegetation characteristics and dynamic changes over different time periods, while multispectral remote sensing images provide rich information on surface features through multiple spectral bands [52,53]. The use of multi-temporal multispectral remote sensing images can effectively improve classification accuracy.

6. Conclusions

In this study, we successfully constructed an HRNet-CASA framework that integrates the HRNet deep learning model and the CASA model, which is designed to classify the forest types in the study area with high accuracy and to accurately estimate the NPP of each forest type based on the classification results. To the best of our knowledge, this is the first attempt to improve the accuracy of NPP estimation compared to classification accuracy. The main contributions of this study are as follows:

Firstly, based on the VHR remote sensing images, the forest types in Huangshan District were finely classified using the HRNet deep learning network, and the forest type distribution map was drawn accordingly. Following a thorough classification process, the forest types in Huangshan District were divided into six categories: broad-leaved forest; coniferous forest; mixed forest; bamboo forest; shrub forest; and non-forest land. In terms of classification effect, the overall accuracy reached 80.6%, with broad-leaved forests having the most significant classification effect and mixed forests having relatively low classification accuracy;

Secondly, this study further combined the results of forest type classification with climate and NDVI data and substituted them into the CASA model to estimate the vegetation NPP in the study area. The average NPP of forest in the study area was found to be 784.49 gCm⁻²a⁻¹, among which the average NPP of broadleaved forest was the highest (1025.94 gCm⁻²a⁻¹), and that of shrub forest was the lowest (440.21 gCm⁻²a⁻¹). Moreover, this study’s validation data were derived from field-collected sample plot data, ensuring the data’s validity and dependability. The experiment’s NPP estimation results exhibit a high degree of dependability after validation.

The experiment demonstrated that the HRNet-CASA framework combines the advanced technology of deep learning in image processing with the advantages of the CASA model in the accurate estimation of NPP. The framework indicates high accuracy in NPP estimation for different types of forests. It also offers strong data support for assessing the ecological service functions of forest ecosystems in a number of areas, including carbon cycling, water conservation, and soil retention. In future research, the introduction of multi-temporal and multi-band remote sensing images can be considered. This approach will help to reduce classification errors caused by seasonal changes and further improve the accuracy of NPP estimation.