Next Article in Journal
Spatio-Temporal Analysis of Erosion Risk Assessment Using GIS-Based AHP Method: A Case Study of Doğancı Dam Watershed in Bursa (Türkiye)
Previous Article in Journal
Evaluation of the Restoration Effects of Rooftop Greening Areas Created by Applying an Ecological Restoration Method
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Forests
Volume 15
Issue 7
10.3390/f15071132
forests-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Forest Canopy Height Retrieval Model Based on a Dual Attention Mechanism Deep Network

by
Zongze Zhao
1,*,
Baogui Jiang
1,
Hongtao Wang
1 and
Cheng Wang
2
1
School of Surveying and Land Information Engineering, Henan Polytechnic University, Jiaozuo 454003, China
2
Key Laboratory of Digital Earth Science, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China
*
Author to whom correspondence should be addressed.
Forests 2024, 15(7), 1132; https://doi.org/10.3390/f15071132
Submission received: 5 June 2024 / Revised: 22 June 2024 / Accepted: 27 June 2024 / Published: 28 June 2024
(This article belongs to the Section Forest Inventory, Modeling and Remote Sensing)
Browse Figures
Versions Notes

Abstract

:
Accurate estimation of forest canopy height is crucial for biomass inversion, carbon storage assessment, and forestry management. However, deep learning methods are underutilized compared to machine learning. This paper introduces the convolutional neural network–bidirectional long short-term memory (CNN-BiLSTM) model and proposes a Convolutional Neural network–spatial channel attention–bidirectional long short-term memory (CNN-SCA-BiLSTM) model, incorporating dual attention mechanisms for richer feature extraction. A dataset comprising vegetation indices and canopy height data from forest regions in Luoyang, specifically within the 8–20 m range, is used for a comparative analysis of multiple models, with accuracy evaluated based on the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R2). The results demonstrate that (1) the CNN-BiLSTM model exhibits strong potential (MAE = 1.6554 m, RMSE = 2.2393 m, R2 = 0.9115) and (2) the CNN-SCA-BiLSTM model, while slightly less efficient (<1%), demonstrates improved performance. It reduces the MAE by 0.3047 m, the RMSE by 0.6420 m, and increases the R2 value by 0.0495. Furthermore, the model is utilized to generate a canopy height map (MAE = 5.2332 m, RMSE = 7.0426 m) for Henan in the Yellow River Basin for the year 2022. The canopy height is primarily distributed around 5–20 m, approaching the accuracy levels of global maps (MAE = 4.0 m, RMSE = 6.0 m).

1. Introduction

Although nearly 70% of forests have undergone intensive management in recent years, our understanding of how this management, in combination with climate change, impacts the overall role of forests as carbon sinks is still limited [1]. Forests, often referred to as the “lungs of the Earth”, cover approximately 30% of the Earth’s land surface and play crucial roles in maintaining biodiversity, purifying the environment, providing forest resources, and conserving water and soil [2,3]. Forest ecosystems are among the most important ecosystems on land and play a significant role in mitigating global climate change [4]. Forests maintain more than half of the terrestrial carbon stock, and their carbon sequestration potential is a research focus of major international and domestic scientific programs, such as the International Geosphere-Biosphere Program and China’s “973 Program” [5]. Canopy height is a structural parameter that reflects the health and productivity of forest ecosystems to a certain extent [6,7]. Inverting regional forest canopy height using vegetation indices is vital for monitoring regional forest ecosystems and ecosystem restoration.
Most current studies use optical images, light detection and ranging (LiDAR), and other remote sensing data for forest canopy height retrieval. Optical remote sensing images provide rich spectral information that can be used to calculate indices, such as the leaf area index and vegetation index. The main data acquisition platforms include the Landsat series [8,9,10], Moderate Resolution Imaging Spectroradiometer (MODIS) [11], and Sentinel series [12,13] satellites. MODIS provides wide-ranging remote sensing images with a short revisit period (1–2 days) and strong real-time capabilities; however, it has lower spatial resolution (250 m–1000 m). The Sentinel satellite series (Sentinel-2) has a data acquisition period of 5 days, weaker real-time capabilities, a higher spatial resolution (10 m–20 m), and 12 spectral bands. The Landsat satellite series has a longer revisit period (16 days), a 30 m spatial resolution, and nine spectral bands. Overall, Sentinel-2 is suitable for high-precision and high-frequency vegetation monitoring, while Landsat-8 is better suited for historical change analysis over large areas and long time series. Information saturation is unavoidable during forest canopy height retrieval using remote optical sensing images. When spectral indices reach saturation, they cannot reflect changes in the forest canopy height. However, LiDAR can penetrate vegetation and overcome information saturation. Field measurements, which are time consuming and labor intensive, are the main methods used for traditional forest canopy height measurements. With the rapid development of science and technology, ground-based, airborne, and spaceborne platforms, including airborne LiDAR [14,15,16,17,18] and spaceborne LiDAR [19,20], have been widely used to retrieve canopy height. Different data platforms have different limitations. For example, (1) airborne LiDAR data have high accuracy but are only suitable for small-area measurement studies; (2) spaceborne LiDAR is suitable for large-area forest canopy height inversion but its accuracy is slightly lower; and (3) airborne and spaceborne platform data are redundant, requiring a considerable amount of time for data preprocessing. Many researchers have combined different data acquisition platforms to obtain high-precision canopy height data and to overcome the influence of data acquisition platforms on forest canopy height inversion results. The aim was to improve the effectiveness of forest canopy height inversion using different modeling methods. In recent years, a new generation of Ice, Cloud, and land Elevation Satellite-2 (ICESat-2) [20,21,22,23] and Global Ecosystem Dynamics Investigation (GEDI) multi-beam LiDAR satellites have been launched [24,25]. These satellites rapidly acquire large-area canopy height data. Many scholars have found that GEDI L2A data are more suitable for forest canopy height retrieval (MAE = 0.31 m, RMSE = 0.87 m) [26]. This study utilized the advantages of rich spectral information in optical remote sensing data for horizontal forest structures and the wide coverage of spaceborne LiDAR data for vertical forest structures. Landsat-8 Operational Land Imager (OLI) and GEDI L2A data were used to invert the regional forest canopy height.
With the rapid development of science and technology, various data acquisition methods have emerged, and forest canopy height retrieval models and methods have become increasingly mature, primarily through machine learning or deep learning methods for processing and analysis. Popescu et al. [27] established linear and nonlinear models to retrieve canopy height. Lefsky et al. [28] developed a forest canopy height extrapolation model based on forest canopy height and spectral information from remote sensing images. Building on Lefsky et al.’s work, Simard et al. [28] added input variables, such as climate data and canopy closure, and used a random forest (RF) model to invert the canopy height. Shufan et al. [29] proposed a multi-modal canopy height retrieval model based on the RF model method [21], and Tamiminia et al. [29,30,31,32], among others, used the RF model to construct canopy height extrapolation models and plotted reliable spatial distribution maps of forest canopy height. Gleason et al. [33] studied methods for inverting the canopy height using support vector machine (SVM), RF, and linear mixed effects (LME) models. Guo et al. [26] conducted a consistency analysis of the Advanced Topographic Laser Altimeter System (ATLAS) and GEDI and developed a consistency model based on RF and stepwise regression. Li et al. and Zhang et al. mapped forest canopy heights in Inner Mongolia and Alaska using deep learning and RF model methods. Lin et al. [34], Fu Ying, and Lin et al. [35] fitted vegetation indices and canopy heights to construct a canopy height model (CHM) dataset and inverted canopy height using a backpropagation–artificial neural network (BP-ANN) model. Rajit et al. [35] inverted the forest canopy height using seven machine learning models: RF, SVM, extreme gradient boosting (Xgbtree), and multivariate adaptive regression splines (MARSs). There are various methods for inverting the forest canopy height. Table 1 shows the common data sources and methods for forest canopy height retrieval. Although the models and methods for canopy height retrieval are relatively mature in terms of machine learning, the inversion accuracy is not stable and is generally low, and most model methods have a weak generalization ability. Although deep learning can improve the accuracy of forest canopy height retrieval, relatively few studies have been conducted on forest canopy height inversion based on deep learning model methods, with the majority of models exhibiting poor generalization ability.
This study proposes a forest canopy height retrieval model, convolutional neural network–spatial channel attention–bidirectional long short-term memory (CNN-SCA-BiLSTM), which leverages spatial and channel attention mechanisms, to address the aforementioned issues. This model aims to rapidly invert forest canopy height over large areas. A discrete CHM dataset was constructed by combining the GEDI forest canopy height data with vegetation index information computed from Landsat-8 OLI remote sensing imagery. The robustness of the model was tested, and its generalization ability was validated over larger geographical areas. Additionally, the model aimed to provide data support for understanding spatial variations in forest canopy height in the Henan Province of the Yellow River Basin and to evaluate the forest ecological environment in the Henan region of the Yellow River Basin. The aim of this research endeavor was to provide valuable insights into the layout and implementation of forest protection projects.
The rest of this paper is structured as follows: Section 2 introduces the study area and experimental data, and describes the proposed model’s framework. Section 3 presents the experimental results and their analysis. Section 4 offers a discussion, and Section 5 concludes the paper.

2. Materials and Methods

2.1. Summary of the Research Region

General Secretary Xi Jinping emphasized the importance of ecological conservation and high-quality development in the Yellow River Basin as major national strategies. Henan Province, located in the middle of the Yellow River Basin, is crucial for the basin’s ecological development. This research focuses on the Henan section, with latitudes ranging from 33°29′33.60″ N to 36°12′35.70″ N and longitudes from 110°14′9.10″ E to 116°6′48.80″ E. This area features mountains, hills, and river valleys. Luoyang City, within this section, is characterized by “Four Mountains” (Jingzi, Qingyao, Mangshan, Yushan) and “Three Rivers” (Qing, Zhen, Jian), with steep mountains, ridges, and fragmented valleys. The region experiences a warm temperate continental monsoon climate with four distinct seasons, an average annual temperature of 14.2 °C, a frost-free period of 216 days, and an average annual precipitation of 642.4 mm. The terrain is diverse, comprising natural and plantation forests with species such as poplar, locust, sophora, Chinese juniper, soapberry, Chinese toon, and sandalwood [36]. Figure 1 illustrates the geographical location of the study area, with red dots representing the selected GEDI L2A photon trajectory points after screening.

2.2. Experimental Data

2.2.1. GEDI L2A

GEDI provides global forest height data with a high resolution of 25 m, as well as a three-dimensional leaf area index and surface biomass data products [37]. In this study, an L2A data product consisting of six sets of data-processing algorithms was utilized. Each algorithm produces 100 relative height metrics that describe the waveform collected by the GEDI. Han et al. [38] verified that the accuracy of the a4 algorithm group in the L2A data product is the highest, making it suitable for canopy height inversion studies. The data used in this paper were obtained for Luoyang City in the Yellow River Basin and the Henan section of the Yellow River Basin, sourced from the Earth Observing System of the National Aeronautics and Space Administration (NASA) (https://search.earthdata.nasa.gov, accessed on 4 June 2024). Detailed information is presented in Table 2, with 58 and 26 files for Luoyang City and the Henan section of the Yellow River Basin, respectively.

2.2.2. Landsat-8 OLI

Landsat-8 OLI data provide remote sensing images with multispectral capabilities and high resolution. Data collected in February 2013 were used in this study. Landsat-8, launched by NASA, represents the eighth generation of Earth land observation satellites. It orbits at an altitude of 705 km with a revisit period of 16 days and has an image swath width of 185 km. The OLI sensor onboard Landsat-8 passively captures spectral images of target objects [39]. Landsat-8 OLI remote sensing images were separately acquired for Luoyang City in the Yellow River Basin and the Henan section of the Yellow River Basin to minimize uncertainties in forest canopy height retrieval caused by temporal differences. The images had cloud coverage of 1% and were obtained from the Geographic Spatial Data Cloud (https://www.gscloud.cn/search, accessed on 4 June 2024). Detailed information is provided in Table 3.

2.3. Experimental Methods

Based on the GEDI L2A data product, photon cloud data in the a4 algorithm group in the Luoyang City area in the Yellow River Basin were extracted, including parameters such as latitude, longitude, coverage, and canopy height. The GEDI L2A photon cloud was resampled to a 30 m × 30 m point cloud dataset containing latitude, longitude, and canopy height information. Landsat-8 OLI remote sensing image data were obtained and calibrated. The land cover classification was performed in the study area using the maximum likelihood supervised classification method. Forest boundaries were extracted through regional mask processing, and forest research areas were cropped. Forest vegetation indices (such as the brightness vegetation index (BI), normalized difference vegetation index (NDVI), and soil-adjusted vegetation index (SAVI)) were calculated based on the bands from the Landsat-8 OLI and combined with GEDI forest photon elevation data to construct a discrete CHM dataset. The training and testing sets were divided into a 7:3 ratio, and three different experimental comparison schemes were designed to validate the rationality of the CNN-SCA-BiLSTM model. The first group comprised a multi-model comparison experiment, including the RF, SVM, BP-ANN, and convolutional neural network–bidirectional long short-term memory (CNN-BiLSTM) model methods. The second group involved precision comparison experiments between the CNN-SCA-BiLSTM model with added spatial attention mechanism (SAM) and channel attention mechanism (CAM) modules and the basic model (CNN-BiLSTM). The aim of the third group was to explore the generalizability of the model using transfer learning tests. Figure 2 illustrates the experimental flowchart for estimating the forest canopy height based on GEDI L2A data products and Landsat-8 OLI remote sensing image data.

2.3.1. Making a CHM Dataset

Forest vegetation indices were obtained by integrating GEDI L2A data for canopy height and Landsat-8 OLI remote sensing imagery. A discrete CHM dataset was generated by utilizing the latitude and longitude information of canopy photons. The parameters of the photon cloud from the a4 algorithm group in the GEDI L2A data, which exhibited superior accuracy in canopy height retrieval, are extracted and presented in Table 4.
According to the following criteria [40], (1) select photons with good quality (quality_flag_a4 = 1); (2) select ascending orbit photon data (degrade_flag = 0); and (3) select photon data with canopy coverage of 0.95 or higher (sensitivity_a4 ≥ 0.95). The Landsat-8 OLI remote sensing imagery has a resolution of 30 m. The filtered photon cloud data were then resampled to a resolution of 30 m. Consequently, there are a total of 5099 photons for the Luoyang area in the Yellow River Basin and 18,231 photons for the Henan section of the Yellow River Basin.
The Landsat-8 OLI remote sensing images of one scene in the Luoyang area in the Yellow River Basin and six scenes in the Henan section of the Yellow River Basin were subjected to radiometric and atmospheric corrections. These were then cropped according to the study area. The study area was categorized into various land cover types utilizing the maximum likelihood supervised classification technique. Forest boundaries were obtained by masking the forest areas. The cropped and corrected images were used to delineate the forest-covered areas of the study region. Forest vegetation indices were calculated based on the data in Table 5.
The corresponding vegetation indices and photon elevations were extracted based on the ground photon positions resampled from the GEDI L2A a4 algorithm group data. A discrete CHM dataset was constructed using ten vegetation indices and canopy heights.

2.3.2. CNN-BiLSTM

A convolutional neural network (CNN) simulates and approximates biological neural networks comprising interconnected neurons that form an adaptive nonlinear dynamic system [41]. CNNs are commonly used for deep learning tasks and typically consist of input, hidden, and output layers. The hidden layers predominantly include the convolutional and pooling layers. The input layer, positioned at the beginning of the CNN structure, provides the data to the network. These data are typically in the form of two-dimensional pixel matrices representing images or text. The convolutional layers processed the input data by extracting features from the pixel matrices. These layers alter the structure of the input data, with the initial layers focusing on extracting low-level features and subsequent layers extracting more complex features. The pooling layers follow the convolutional layers and perform operations such as max pooling. These operations reduce the dimensionality of the extracted features and prevent model overfitting by summarizing the most important information. The output layer, which is positioned at the end of the CNN, receives the processed features from the convolutional and pooling layers. This information was combined using learned weights and biases to produce the final output of the network [42]. The structure and operations of the pooling and output layers are determined based on the results obtained from the preceding layers and the associated weights and biases of each neuron.
The bidirectional long short-term memory (BiLSTM) network, as previously described [43,44], consists of both a forward long short-term memory (LSTM) and a backward LSTM. Each LSTM unit is composed of three gates: forget, input, and output [45]. The forget gate determines the amount of feature information to pass through, thereby filtering irrelevant information. The input gate primarily updates information in the cell state using sigmoid functions and tanh layers, thereby incorporating the latest feature information. Finally, the output gate outputs the current feature information.
The CNN-BiLSTM model integrates the CNN network structure with the BiLSTM network structure [46,47,48]. A CNN emulates the human brain’s processing of visual information through convolutional, pooling, and fully connected layers, thereby facilitating more effective feature extraction. In contrast, BiLSTM introduces a bidirectional information flow atop LSTM, housing two directions of LSTM units: one processing the input sequence chronologically and the other processing it in reverse [49]. This bidirectional architecture enables the network to simultaneously capture past and future information in a sequence, thereby providing a more comprehensive understanding of the input sequence. The core structure of BiLSTM consists of two directions of LSTM units, each in a hidden state. For each time step, the forward LSTM is processed from the beginning of the sequence, and the backward LSTM is processed from the end. The final hidden state is the concatenation of the hidden states in these two directions. A CNN significantly reduces the number of parameters that need to be learned when handling large-scale data, thereby enhancing the training efficiency. Its multilayer convolutional structure can extract increasingly abstract features from low to high levels, enhancing the hierarchical representation capabilities of the model. The three gates of BiLSTM regulate the information flow through learned parameters, enabling the network to better capture and utilize the long-term dependencies in time-series data. By combining the CNN and BiLSTM networks, the feature extraction capabilities of the CNN and the advantages of BiLSTM in spatial and temporal modeling can be fully exploited to render the model more suitable for complex tasks and diverse data types.
The model established in this study, as depicted in Figure 3, utilizes various neural network layers to construct a deep learning network. The network structure included normalization, convolution, a rectified linear unit (ReLU), batch normalization (BN), and max pooling, as well as a flattened BiLSTM with fully connected regression layers.Parameters such as numFeatures, numResponses, and numHiddenUnits define the structure and training parameters of the CNN-BiLSTM. Here, numHiddenUnits denotes the number of neurons in the hidden layer of BiLSTM. The convolutional, ReLU, and BN layers were configured as one module with two such modules.

2.3.3. CNN-SCA-BiLSTM

The attention mechanism simulates the attention of the human brain, drawing inspiration from how the human brain focuses on specific areas at a particular moment while reducing or even ignoring attention to other parts [50]. It assigns different weights and biases to the input features of the model, emphasizing more crucial influencing factors and thereby improving the model’s decision-making ability. More importantly, it does not increase the computational or storage overheads of the model [51]. Incorporating the attention mechanism into the CNN-BiLSTM model focuses on highlighting the factors that affect canopy height retrieval, thereby enhancing the accuracy of canopy height retrieval.
In this study, we introduced a novel model for forest canopy height retrieval, termed the CNN-SCA-BiLSTM, by integrating an attention mechanism into the CNN-BiLSTM network model. In this enhanced network, the CNN simulates the human brain’s processing of visual information through multiple neural network layers, significantly reducing the number of parameters to be learned and thus improving training efficiency. The CNN’s layered structure extracts increasingly abstract features from low to high levels, thereby enhancing the model’s hierarchical representation capability for effective feature extraction. The bidirectional structure of BiLSTM enables the network to capture the dependencies between past and future information in a sequence, thereby gaining a more comprehensive understanding of the input sequence. The attention mechanism can emphasize more critical influencing factors, improving the model’s decision-making ability without increasing its space and time complexities. By combining the CNN, attention mechanism, and BiLSTM network, this model leverages the CNN’s feature extraction capabilities, the attention mechanism’s ability to identify key influencing factors, and the advantages of the BiLSTM network in spatial and temporal modeling. An attention mechanism layer is added after the flattening layer, as shown in Figure 4, which combines the SAM and channel attention mechanism (SCM) to enhance the representational capability of the CNN features. The SAM performs spatial dimension maximum and average pooling on the input features, thereby compressing the spatial size to facilitate the learning of spatial features [2]. The obtained feature maps were concatenated and subjected to convolutional operations, followed by the ReLU activation function, to obtain the spatial attention weight matrix. The SAM helps the network better understand the importance of different features and enhances its perception of local information. The SCM performs maximum and average pooling on the input features in the channel dimensions, thereby compressing the channel size. The results of the max pooling and average pooling were input into a multilayer perceptron (MLP). The outputs of the MLP are added together and mapped through the ReLU activation function to obtain the channel attention weight matrix [52]. The SCM helps the network better understand the correlation between different channels, enabling the extraction of more discriminative features. By weighing different channel features, the network’s focus on important features is enhanced, thus improving its representation capability [53]. The SAM and SCM weight matrices were multiplied to form a new spatial channel attention (SCA) matrix. The attention mechanism layer integrates the advantages of the spatial and channel attention mechanisms, considering both the importance of different spatial positions and the correlation between different channels, effectively enhancing the network’s representation capability and improving its performance.

2.4. Precision Evaluation Index

The mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R2) were selected as the accuracy evaluation metrics to evaluate the predictive performance of the model. The formulae for these three accuracy evaluation metrics are shown in Equations (1)–(3). The MAE and RMSE are commonly used to evaluate the prediction accuracy of models. Smaller MAE and RMSE values indicated smaller prediction errors and better prediction performance. The RMSE amplifies the differences with larger errors; therefore, a smaller RMSE also indicates better prediction performance. The R2 is a statistic used to evaluate the goodness of fit of the regression models. This represents the degree of fit of the model to the data, with values ranging from 0 to 1. A value closer to 1 indicates a better fit of the model to the data, while a value closer to 0 indicates a poorer fit of the model to the data [54].
A E = 1 n ∑ i = 1 n y i ^ − y i
M S E = 1 n ∑ i = 1 n y i ^ − y i 2
R 2 = 1 − ∑ i = 1 n y i ^ − y i 2 ∑ i = 1 n y i − y i ¯ 2
In the formula, n represents the number of samples, y i represents the i th true value, y i ^ represents the i th predicted value, and y i ¯ represents the mean of the true values.

3. Experimental Results and Analysis

3.1. CHM Result Analysis

After radiometric and atmospheric correction, Landsat-8 OLI remote sensing images were fused and cropped according to the Luoyang area in the Yellow River Basin and the section of Henan Province in the Yellow River Basin. The supervised maximum likelihood classification method was used for land cover classification in the study area (Figure 5a,b). Subsequently, forest boundaries were obtained through forest area masking processing, and the corrected images were cropped to obtain the forest coverage of the study area (Figure 5b, red area).
During land cover classification, the overall accuracy of the maximum likelihood supervised classification method reached 97.3%, with a kappa coefficient of 0.9417, whereas the overall accuracy of the SVM classification method reached 98.23%, with a kappa coefficient of 0.9673. Although the SVM classification method achieved higher accuracy, the maximum likelihood supervised classification method completed the classification in about 1 h compared to the 5 h required by the SVM method, making it more time efficient. Moreover, both classification methods showed similar accuracy levels, indicating that the maximum likelihood supervised classification method satisfied the experimental requirements. Comparing the land cover classification results illustrated that the distribution of discrete canopy heights in the Henan section of the Yellow River Basin was mainly concentrated in the vicinity of Luoyang City, Nanyang City, and Sanmenxia City (Figure 5a). Figure 5b shows that forests in the Luoyang area in the Yellow River Basin were mainly distributed around Xin’an County, Luanchuan County, and Luoning County. The forest distribution in the light-blue boundary area (red area) in Figure 5a is consistent with the forest distribution in the red area in Figure 5b. The terrain and landforms in the Henan section of the Yellow River Basin exhibited strong diversity, with more than six types of land cover in both the Luoyang area and the Henan section of the Yellow River Basin.
Based on the land cover classification results, a regional mask was created to obtain forest boundaries. Remote sensing image data within the forest area were extracted, and forest vegetation indices were calculated using these forest boundaries. Figure 6 shows the partial vegetation index results for the Luoyang area in the Yellow River Basin.
Based on the GEDI L2A forest canopy height and Landsat-8 OLI-calculated forest vegetation index, a discrete CHM dataset for the Luoyang area in the Yellow River Basin was constructed. The spatial distribution of the discrete CHM dataset for the Luoyang area in the Yellow River Basin is shown in Figure 7. There were 1442 points between 0 m and 8 m in canopy height, 1622 points between 8 m and 13 m, 1222 points between 13 m and 20 m, 592 points between 20 m and 29 m, and 221 points greater than 29 m, totaling 5099 discrete canopy photon points. The discrete CHM dataset was mainly concentrated between 3 m and 30 m in height.

3.2. Comparison of Multi-Model Results

By splitting the discrete CHM dataset from the Luoyang area in the Yellow River Basin into training and testing sets at a 7:3 ratio, the CNN-BiLSTM and CNN-SCA-BiLSTM models were trained. A comparative analysis was then conducted to assess the feasibility of machine learning methods, like RF and SVM, as well as deep learning methods, such as the BP-ANN and CNN-BiLSTM models. This was followed by an evaluation of the CNN-SCA-BiLSTM model’s effectiveness in enhancing canopy height inversion relative to the base model. Finally, the model’s generalizability was tested in a larger region within the Henan section of the Yellow River Basin.
The forest canopy height retrieval results of the CNN-BiLSTM model were compared with those of the RF, SVM, and BP-ANN models. The model accuracy results are shown in Table 6, indicating the following. (1) The machine learning methods, RF and SVM, had almost identical training and testing accuracies, while the neural network model, BP-ANN, outperformed the machine learning methods with a decrease in the MAE by approximately 0.29 m and the RMSE by approximately 0.46 m, and an increase in the R2 to around 0.90. (2) The CNN-BiLSTM deep learning model outperformed the RF, SVM, and BP-ANN models, with reductions in the MAE and RMSE by around 0.56 m and an increase in the R2 to approximately 0.90. Figure 8a–d provide a more visual representation of the forest canopy height retrieval results of the four models. Figure 8a,b shows the inversion results of the RF and SVM, respectively, with the prediction performance of these two models being similar when combined with the results in Table 5. Figure 8c shows a better overlap between the true and predicted values, indicating the stronger inversion capability of the BP-ANN model. Figure 8d illustrates the forest canopy height inversion effect of the CNN-BiLSTM model, showing a higher degree of fit between the true and predicted values. The experiment demonstrated that the forest canopy height retrieval effect of the CNN-BiLSTM model was superior, indicating its strong feasibility (Table 5).
The SAM and SCM were integrated based on the CNN-BiLSTM model to construct a new model called the CNN-SCA-BiLSTM. Table 5 presents the accuracy results and model training time of the CNN-SCA-BiLSTM compared with those of the base model. The incorporation of the attention mechanism into the CNN-SCA-BiLSTM model led to a significant improvement in the inversion performance, with a runtime speed reduction of less than 1%. The MAE decreased by approximately 0.30474 m, the RMSE decreased by approximately 0.477956 m, and the R2 increased to approximately 0.94, representing an improvement of approximately 0.05 compared with the base model. The addition of the attention mechanism had a minimal impact on the model’s efficiency. Figure 8e illustrates the canopy inversion effect of the CNN-SCA-BiLSTM model. Figure 8d, e shows a noticeable enhancement in the inversion capability of the CNN-SCA-BiLSTM model, with superior inversion results.
A comparative analysis of the five major canopy height retrieval models in the Luoyang area in the Yellow River Basin showed that both the true and predicted values were distributed near the fitting line, indicating good canopy height inversion performance. Machine learning methods, such as RF and SVM, show almost identical results, whereas the neural network model, BP-ANN, shows some improvement in inversion accuracy. Many researchers have explored the inversion capabilities of the RF, SVM, and BP-ANN models for canopy height estimation, and this study attempted to use the CNN-BiLSTM model. This model has been successfully applied to tasks such as text classification, sentiment analysis, and named entity recognition. Compared to machine learning models, such as RF and SVM, and the deep neural network, BP-ANN, despite longer training times, the CNN-BiLSTM model shows significantly improved accuracy with a decrease of around 0.5 m in the MAE and RMSE and an increase of around 0.05 in the R2, highlighting its superiority in canopy height inversion. The CNN-BiLSTM model outperformed the RF, SVM, and BP-ANN models in terms of the canopy height inversion. By incorporating spatial and channel attention mechanisms into the CNN-BiLSTM model, the runtime speed of the model was reduced by less than 1%, and the prediction accuracy significantly improved. Overall, the CNN-SCA-BiLSTM model demonstrated superior performance in terms of canopy height inversion.

3.3. Transfer Learning

After a precision comparison, the CNN-SCA-BiLSTM model exhibited the best performance in forest canopy height retrieval. The CNN-SCA-BiLSTM model was applied to predict the canopy height in the forest area of the broader Yellow River Basin’s Henan section to verify its generalizability. The aim was to provide data support for revealing the spatial changes in forest canopy height in the Henan Province of the Yellow River Basin and to evaluate the forest ecological environment in the Henan region, thus offering a reference for the layout and implementation of forest protection projects. The research area overlaps with the Yellow River Basin map and the administrative map of the Henan Province. Based on the same data processing principles and methods, vegetation indices were calculated using Landsat-8 OLI remote sensing images. The forest canopy height in the forested areas in the Yellow River Basin was inverted using the CNN-SCA-BiLSTM model, resulting in a continuous CHM map of the Henan section of the Yellow River Basin, as shown in Figure 9.
Figure 9 illustrates the continuous distribution of forest canopy height in the Henan section of the Yellow River Basin. The forest canopy height is mainly concentrated between 5 m and 20 m, with an average canopy height of approximately 5.6237 m. Although the forest canopy height is not high, the distribution of forested areas is extensive. Forest canopy height is closely related to forest biomass and carbon storage, indicating rich forest biomass and carbon reserves in the region. The forest resources in this area are mainly distributed in low-latitude regions, with forest resources becoming more abundant as latitude decreases. In the upstream of the Henan section of the Yellow River Basin, forest canopy height is primarily concentrated between 10 m and 20 m, indicating abundant forest volume. In the downstream areas, forest canopy height is mainly concentrated below 10 m, indicating relatively scarce forest volume and reflecting the better ecological environment in the upstream compared to the downstream areas that require further restoration.
Using GEDI L2A canopy height point cloud data collected in 2022, the accuracy of the canopy height map (5 m–25 m) of the Henan section of the Yellow River Basin was evaluated, resulting in a MAE of 5.2332 m and an RMSE of 7.0426 m. A comparison with the canopy height map (7 m–28 m) generated by Lang et al. [55] using a probabilistic deep learning model at a 10 m resolution (MAE = 4.0 m, RMSE = 6.0 m) showed a difference of approximately 1 m in both the MAE and RMSE. This study utilized 30 m resolution Landsat-8 OLI and GEDI data, employing optical imagery vegetation indices for regional forest canopy height inversion. In contrast, Lang et al. used 10 m resolution Sentinel-2 and GEDI discrete canopy height point cloud data to create a global canopy height map through interpolation methods. The variance in data sources and the differences in deep network models are the main factors contributing to the variance in accuracy. Despite the relatively lower accuracy of the canopy height map in the Henan section of the Yellow River Basin, achieving a level of accuracy close to the global canopy height map was possible using only a single vegetation index dataset. These results indicate that the CNN-SCA-BiLSTM model is suitable for regional forest canopy height inversion.

4. Discussion

This study first selected GEDI L2A data within the scope of Luoyang City in the Yellow River Basin based on specific screening criteria and then cropped it to ensure the accuracy of the forest canopy height obtained from GEDI L2A. A discrete CHM dataset was constructed by combining forest vegetation indices calculated from the Landsat-8 OLI remote sensing image data. Landsat-8 OLI remote sensing images underwent radiometric and atmospheric correction preprocessing to obtain vegetation indices. Land cover classification was performed using the maximum likelihood supervised classification method, while forest vegetation indices were derived through band calculations. A comparative analysis was conducted between the SVM and maximum likelihood supervised classification methods during the land cover classification stage. Despite the higher accuracy of the SVM classification method, it is time consuming and provides only a marginal improvement in classification accuracy. The maximum likelihood supervised classification method was selected for classifying land cover. Forest region masks were obtained based on the forest area, revealing that (1) forests in Luoyang City in the Yellow River Basin were mainly distributed in the areas surrounding Xin’an, Luanchuan, and Luoning Counties. (2) Forested areas in the Henan section of the Yellow River Basin were primarily located in Luoyang City, Nanyang City, Sanmenxia City, and surrounding areas, with canopy heights mainly concentrated between 5 m and 20 m, decreasing gradually with increasing latitude. A comparative analysis of forest distribution within Luoyang City and the Henan section of the Yellow River Basin revealed similar distribution patterns. The classification results of land cover indicated a wide distribution of forests in the Henan section of the Yellow River Basin, indicating that the ecological restoration effect in the vicinity of the Yellow River Basin has been effective in recent years.
This study used the BP-ANN method for forest canopy height retrieval, which shows higher accuracy than the study by Lin et al. [34], who also used the BP-ANN method (RMSE = 3.42 m, R2 = 0.786). This difference in accuracy can be attributed to the choice of the data source. Ying et al. primarily used ICESat-2/ATLAS data, whereas this study utilized more precise canopy height data from the a4 algorithm group in the GEDI L2A data, improving canopy height prediction accuracy. Throughout the experimental process, the study area had diverse and complex terrain. The quality of the GEDI data may be poor in such complex terrain areas, and the data quality may affect the accuracy of the forest canopy height retrieval. Owing to the 30 m resolution of Landsat-8 OLI remote sensing imagery, it is necessary to resample the discrete photon cloud data during the processing of GEDI data. This may result in deviations in forest canopy height, which can affect the accuracy of canopy height retrieval. Comparing the forest canopy height distribution map of the Henan section of the Yellow River Basin inferred by the CNN-SCA-BiLSTM model in 2022 (MAE = 5.2332 m, RMSE = 7.0426 m) with the global canopy height distribution map from 2020 (MAE = 4.0 m, RMSE = 6.0 m), the accuracy was slightly lower. However, achieving a level of accuracy close to the global canopy height map using only a single vegetation index dataset for regional forest canopy height inversion demonstrates the significant potential of the CNN-SCA-BiLSTM model in regional forest canopy height inversion. Some researchers have used MODIS data in combination with the BP-ANN method for canopy height retrieval, achieving an RMSE of 2.5 m and an R2 above 0.7 [24]. One could consider incorporating MODIS data into the forest canopy height retrieval model or designing a more optimized model to retrieve canopy height to further improve the prediction accuracy of canopy height. The integration of MODIS data or other datasets to enhance canopy height retrieval accuracy should be explored in future research. Factors such as forest species, terrain, and data collection time could potentially influence the accuracy of canopy height retrieval. Future work will focus on analyzing these factors that influence forest canopy height retrieval and attempt to mitigate their effects to improve the accuracy of canopy height retrieval for better results.

5. Conclusions

This study integrates Landsat-8 and GEDI data to construct a discrete CHM dataset, comparing the accuracy of the RF, SVM, BP-ANN, CNN-BiLSTM, and CNN-SCA-BiLSTM models. Additionally, it validates the capability of the CNN-SCA-BiLSTM model in generating continuous CHM maps. A comparative analysis of the forest canopy height estimation abilities of the four models in the Yellow River Basin, specifically in Luoyang City, yielded the following experimental results. (1) The RF, SVM, BP-ANN, and CNN-BiLSTM models all exhibited strong capabilities in the field of forest canopy height estimation. The MAE values were all below 2.5827 m, the RMSE values were below 3.3435 m, and the R2 values were all above 0.8027. (2) The CNN-BiLSTM deep learning model demonstrated superior performance compared to the machine learning (RF and SVM) and neural network (BP-ANN) models. The MAE and RMSE were reduced by approximately 0.56 m, and the R2 improved to around 0.90. CNN-BiLSTM showed strong feasibility for forest canopy height estimation. In a comparative analysis between the CNN-SCA-BiLSTM model and the base CNN-BiLSTM model, (1) introducing the attention mechanism improved the forest canopy height estimation performance of the CNN-BiLSTM model significantly. The MAE decreased by approximately 0.30474 m, the RMSE decreased by approximately 0.477956 m, and the R2 improved to approximately 0.94, an increase of approximately 0.05 compared with the base model, with similar model efficiency. Finally, applying the optimal forest canopy height estimation model, CNN-SCA-BiLSTM, to a larger scope, specifically the Henan section of the Yellow River Basin, yielded the following findings. (1) The CNN-SCA-BiLSTM model demonstrates strong capabilities in canopy height inversion, with a MAE of 5.2332 m and an RMSE of 7.0426 m in the forest canopy height inversion in the Henan section of the Yellow River Basin, achieving accuracy close to the global canopy height map. (2) As latitude decreases, the forest canopy height in the Henan section of the Yellow River Basin gradually increases, showing a trend of gradually increasing canopy height with lower latitudes. (3) The area possesses extensive forest ecosystems that hold significant ecological value for the restoration of the Yellow River Basin’s ecology. (4) The CNN-SCA-BiLSTM model reliably fitted the GEDI L2A data and Landsat-8 OLI remote sensing image data products, effectively estimating the forest canopy height in the Henan section of the Yellow River Basin.
This paper not only proposes a new model for forest canopy height inversion but also reveals the spatial variation and distribution of forest canopy height in the Henan section of the Yellow River Basin. Forest canopy height decreases gradually in the direction of increasing latitude in the Henan section of the Yellow River Basin, while it decreases gradually with increasing longitude. Canopy height is higher towards the upper reaches. In Luoyang City, forests are mainly distributed around Xin’an, Luanchuan, and Luoning Counties. In the Henan section of the Yellow River Basin, forests are primarily concentrated near Luoyang, Nanyang, and Sanmenxia. The distribution of discrete canopy heights and forest regions provides valuable insights into the layout and implementation of forest protection projects in the Henan section of the Yellow River Basin.

6. Patents

A patent titled “Forest Canopy Height Inversion Model with Dual Attention Mechanism Deep Network” is currently under review.

Author Contributions

Z.Z. and B.J. conceived and designed the study; B.J. completed the data analysis and wrote the manuscript; Z.Z., B.J., H.W. and C.W. contributed to investigation, data analyses, and writing. All authors have read and agreed to the published version of the manuscript.

Funding

This research was supported by the State Key Project of National Natural Science Foundation of China–Key projects of joint fund for regional innovation and development [grant number U22A20566].

Data Availability Statement

The raw/processed data required to reproduce these findings cannot be shared at this time as the data also form part of an ongoing study.

Acknowledgments

We appreciate the data provided by organizations such as the National Aeronautics and Space Administration (NASA) to support our experimental research.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Islam, M.R.; Jönsson, A.M.; Bergkvist, J.; Lagergren, F.; Lindeskog, M.; Mölder, M.; Scholze, M.; Kljun, N. Projected effects of climate change and forest management on carbon fluxes and biomass of a boreal forest. Agric. For. Meteorol. 2024, 349, 109959. [Google Scholar] [CrossRef]
  2. Chamberlain, C.P.; Sánchez Meador, A.J.; Thode, A.E. Airborne lidar provides reliable estimates of canopy base height and canopy bulk density in southwestern ponderosa pine forests. For. Ecol. Manag. 2021, 481, 118695. [Google Scholar] [CrossRef]
  3. Lu, Q.; Gu, J.; Ruan, H.; Shang, S.; Hu, X.; Tang, H.; Liu, N. Construction of classification system of natural resources under the concept of new nature protected area system. J. Nanjing For. Univ. (Nat. Sci. Ed.) 2023, 48, 125–134. [Google Scholar] [CrossRef]
  4. Gajendiran, K.; Kandasamy, S.; Narayanan, M. Influences of wildfire on the forest ecosystem and climate change: A comprehensive study. Environ. Res. 2024, 240, 117537. [Google Scholar] [CrossRef] [PubMed]
  5. Canadell, J.G.; Steffen, W.L.; White, P.S. IGBP/GCTE terrestrial transects: Dynamics of terrestrial ecosystems under environmental change. J. Veg. Sci. 2002, 13, 298–300. [Google Scholar] [CrossRef]
  6. Douss, R.; Farah, I.R. Extraction of individual trees based on Canopy Height Model to monitor the state of the forest. Trees For. People 2022, 8, 100257. [Google Scholar] [CrossRef]
  7. Qin, Z.; Xu, L.; Guo, C.; Yang, H.; Yang, Z.; Wang, M.; Wu, Z.; Xia, C. Application of ICESat-2/ATLAS in estimating forest structure parameters. For. Technol. Newsl. 2023, 1–8. [Google Scholar] [CrossRef]
  8. Ahmed, O.S.; Franklin, S.E.; Wulder, M.A.; White, J.C. Characterizing stand-level forest canopy cover and height using Landsat time series, samples of airborne LiDAR, and the Random Forest algorithm. ISPRS J. Photogramm. Remote Sens. 2015, 101, 89–101. [Google Scholar] [CrossRef]
  9. Liao, G.; He, P.; Gao, X.; Lin, Z.; Huang, C.; Zhou, W.; Deng, O.; Xu, C.; Deng, L. Land use optimization of rural production–living–ecological space at different scales based on the BP–ANN and CLUE–S models. Ecol. Indic. 2022, 137, 108710. [Google Scholar] [CrossRef]
  10. Potapov, P.; Li, X.; Hernandez-Serna, A.; Tyukavina, A.; Hansen, M.C.; Kommareddy, A.; Pickens, A.; Turubanova, S.; Tang, H.; Silva, C.E.; et al. Mapping global forest canopy height through integration of GEDI and Landsat data. Remote Sens. Environ. 2021, 253, 112165. [Google Scholar] [CrossRef]
  11. Sawada, Y.; Suwa, R.; Jindo, K.; Endo, T.; Oki, K.; Sawada, H.; Arai, E.; Shimabukuro, Y.E.; Celes, C.H.S.; Campos, M.A.A.; et al. A new 500-m resolution map of canopy height for Amazon forest using spaceborne LiDAR and cloud-free MODIS imagery. Int. J. Appl. Earth Obs. Geoinf. 2015, 43, 92–101. [Google Scholar] [CrossRef]
  12. Torresani, M.; Rocchini, D.; Alberti, A.; Moudrý, V.; Heym, M.; Thouverai, E.; Kacic, P.; Tomelleri, E. LiDAR GEDI derived tree canopy height heterogeneity reveals patterns of biodiversity in forest ecosystems. Ecol. Inform. 2023, 76, 102082. [Google Scholar] [CrossRef] [PubMed]
  13. Liu, X.; Su, Y.; Hu, T.; Yang, Q.; Liu, B.; Deng, Y.; Tang, H.; Tang, Z.; Fang, J.; Guo, Q. Neural network guided interpolation for mapping canopy height of China’s forests by integrating GEDI and ICESat-2 data. Remote Sens. Environ. 2022, 269, 112844. [Google Scholar] [CrossRef]
  14. Brüllhardt, M.; Rotach, P.; Schleppi, P.; Bugmann, H. Vertical light transmission profiles in structured mixed deciduous forest canopies assessed by UAV-based hemispherical photography and photogrammetric vegetation height models. Agric. For. Meteorol. 2020, 281, 107843. [Google Scholar] [CrossRef]
  15. Farhadur Rahman, M.; Onoda, Y.; Kitajima, K. Forest canopy height variation in relation to topography and forest types in central Japan with LiDAR. For. Ecol. Manag. 2022, 503, 119792. [Google Scholar] [CrossRef]
  16. García, M.; Saatchi, S.; Ustin, S.; Balzter, H. Modelling forest canopy height by integrating airborne LiDAR samples with satellite Radar and multispectral imagery. Int. J. Appl. Earth Obs. Geoinf. 2018, 66, 159–173. [Google Scholar] [CrossRef]
  17. Ni, W.; Zhang, D.; Wang, Y.; Pang, Y.; Zhang, Z.; Liu, J.; He, Y.; Guo, W. Forest height extraction from Gaofen-2 transorbital stereo data. J. Remote Sens. 2018, 22, 392–399. [Google Scholar]
  18. Yu, J.; Nie, S.; Liu, W.; Zhu, X.; Sun, Z.; Li, J.; Wang, C.; Xi, X.; Fan, H. Mapping global mangrove canopy height by integrating Ice, Cloud, and Land Elevation Satellite-2 photon-counting LiDAR data with multi-source images. Sci. Total Environ. 2024, 939, 173487. [Google Scholar] [CrossRef]
  19. Li, Z.; Xuan, F.; Dong, Y.; Huang, X.; Liu, H.; Zeng, Y.; Su, W.; Huang, J.; Li, X. Performance of GEDI data combined with Sentinel-2 images for automatic labelling of wall-to-wall corn mapping. Int. J. Appl. Earth Obs. Geoinf. 2024, 127, 103643. [Google Scholar] [CrossRef]
  20. Pang, S.; Li, G.; Jiang, X.; Chen, Y.; Lu, Y.; Lu, D. Retrieval of forest canopy height in a mountainous region with ICESat-2 ATLAS. For. Ecosyst. 2022, 9, 100046. [Google Scholar] [CrossRef]
  21. Mulverhill, C.; Coops, N.C.; Hermosilla, T.; White, J.C.; Wulder, M.A. Evaluating ICESat-2 for monitoring, modeling, and update of large area forest canopy height products. Remote Sens. Environ. 2022, 271, 112919. [Google Scholar] [CrossRef]
  22. Matasci, G.; Hermosilla, T.; Wulder, M.A.; White, J.C.; Coops, N.C.; Hobart, G.W.; Zald, H.S.J. Large-area mapping of Canadian boreal forest cover, height, biomass and other structural attributes using Landsat composites and lidar plots. Remote Sens. Environ. 2018, 209, 90–106. [Google Scholar] [CrossRef]
  23. Dong, H.; Yu, Y.; Fan, W. Verification of underforest terrain inversion performance of satellite-borne Lidar GEDI data. J. Nanjing For. Univ. (Nat. Sci. Ed.) 2023, 47, 141–149. [Google Scholar] [CrossRef]
  24. Lin, X.; Xu, M.; Cao, C.; Dang, Y.; Huang, Z. Estimates of Forest Canopy Height Using a Combination of ICESat-2/ATLAS Data and Stereo-Photogrammetry. Remote Sens. 2020, 12, 3649. [Google Scholar] [CrossRef]
  25. Wang, Z.; Cai, H.; Yang, X. A new method for mapping vegetation structure parameters in forested areas using GEDI data. Ecol. Indic. 2024, 164, 112157. [Google Scholar] [CrossRef]
  26. Guo, Q.; Du, S.; Jiang, J.; Guo, W.; Zhao, H.; Yan, X.; Zhao, Y.; Xiao, W. Combining GEDI and sentinel data to estimate forest canopy mean height and aboveground biomass. Ecol. Inform. 2023, 78, 102348. [Google Scholar] [CrossRef]
  27. Popescu, S.C. Estimating biomass of individual pine trees using airborne lidar. Biomass Bioenergy 2007, 31, 646–655. [Google Scholar] [CrossRef]
  28. Lefsky, A.M. A global forest canopy height map from the Moderate Resolution Imaging Spectroradiometer and the Geoscience Laser Altimeter System. Geophys. Res. Lett. 2010, 37, L15401. [Google Scholar] [CrossRef]
  29. Wang, S.; Liu, C.; Li, W.; Jia, S.; Yue, H. Hybrid model for estimating forest canopy heights using fused multimodal spaceborne LiDAR data and optical imagery. Int. J. Appl. Earth Obs. Geoinf. 2023, 122, 103431. [Google Scholar] [CrossRef]
  30. Tamiminia, H.; Salehi, B.; Mahdianpari, M.; Goulden, T. State-wide forest canopy height and aboveground biomass map for New York with 10 m resolution, integrating GEDI, Sentinel-1, and Sentinel-2 data. Ecol. Inform. 2024, 79, 102404. [Google Scholar] [CrossRef]
  31. Nandy, S.; Srinet, R.; Padalia, H. Mapping Forest Height and Aboveground Biomass by Integrating ICESat-2, Sentinel-1 and Sentinel-2 Data Using Random Forest Algorithm in Northwest Himalayan Foothills of India. Geophys. Res. Lett. 2021, 48, e2021GL093799. [Google Scholar] [CrossRef]
  32. Simard, M.; Pinto, N.; Fisher, J.B.; Baccini, A. Mapping forest canopy height globally with spaceborne lidar. J. Geophys. Res. Biogeosciences 2011, 116, G04021. [Google Scholar] [CrossRef]
  33. Gleason, C.J. Forest biomass estimation from airborne LiDAR data using machine learning approaches. Remote Sens. Environ. 2012, 125, 80–91. [Google Scholar] [CrossRef]
  34. Lin, X. Remote Sensing Diagnosis of Forest Canopy Height and Forest Aboveground Biomass Based on ICESat-2 and GEDI. Ph.D. Thesis, Chinese Academy of Sciences, Beijing, China, 2021. [Google Scholar]
  35. Gupta, R.; Sharma, L.K. Mixed tropical forests canopy height mapping from spaceborne LiDAR GEDI and multisensor imagery using machine learning models. Remote Sens. Appl. Soc. Environ. 2022, 27, 100817. [Google Scholar] [CrossRef]
  36. Li, P.; Rana, S.; Zhang, M.; Jin, C.; Tian, K.; Liu, Z.; Li, Z.; Cai, Q.; Geng, X.; Wang, Y. An investigation of the growth status of 19-year-old Idesia polycarpa ‘Yuji’ plantation forest in the mountainous region of Henan, China. Heliyon 2023, 9, e19716. [Google Scholar] [CrossRef] [PubMed]
  37. Zhao, P.; Li, S.; Ma, Y.; Liu, X.; Yang, J.; Yu, D. A new terrain matching method for estimating laser pointing and ranging systematic biases for spaceborne photon-counting laser altimeters. ISPRS J. Photogramm. Remote Sens. 2022, 188, 220–236. [Google Scholar] [CrossRef]
  38. Bhandari, K.; Srinet, R.; Nandy, S. Forest Height and Aboveground Biomass Mapping by synergistic use of GEDI and Sentinel Data using Random Forest Algorithm in the Indian Himalayan Region. J. Indian Soc. Remote Sens. 2024, 52, 857–869. [Google Scholar] [CrossRef]
  39. Zong, M.; Wang, G.Z.; Han, G.X.; Li, Y.Z.; Zhao, M. Spatial and Temporal Evolution and Driving Mechanism of Man-made Ditches in the Yellow River Delta from 1976 to 2015. J. Ludong Univ. (Nat. Sci. Ed.) 2017, 33, 68–75. [Google Scholar]
  40. Han, M.; Xing, Y.; Li, G.; Huang, J.; Cai, L. Comparison of Accuracy of Forest Maximum Canopy Height and Biomass Inversion Using GEDI Different Algorithm Groups Data. J. Cent. South Univ. For. Technol. 2022, 42, 11. [Google Scholar]
  41. Shruti, P.; Rekha, R. A Review of Convolutional Neural Networks, its Variants and Applications. In Proceedings of the 2023 International Conference on Intelligent Systems for Communication, IoT and Security (ICISCoIS), Coimbatore, India, 9–11 February 2023; pp. 31–36. [Google Scholar]
  42. Fayad, I.; Ienco, D.; Baghdadi, N.; Gaetano, R.; Alvares, C.A.; Stape, J.L.; Ferraço Scolforo, H.; Le Maire, G. A CNN-based approach for the estimation of canopy heights and wood volume from GEDI waveforms. Remote Sens. Environ. 2021, 265, 112652. [Google Scholar] [CrossRef]
  43. Chen, J.; Ying, Z.; Zhang, C.; Balezentis, T. Forecasting tourism demand with search engine data: A hybrid CNN-BiLSTM model based on Boruta feature selection. Inf. Process. Manag. 2024, 61, 103699. [Google Scholar] [CrossRef]
  44. Kumar, S.; Kumar, V. Multi-view Stacked CNN-BiLSTM (MvS CNN-BiLSTM) for urban PM2.5 concentration prediction of India’s polluted cities. J. Clean. Prod. 2024, 444, 141259. [Google Scholar] [CrossRef]
  45. Abdalla, A.; Wheeler, T.A.; Dever, J.; Lin, Z.; Arce, J.; Guo, W. Assessing fusarium oxysporum disease severity in cotton using unmanned aerial system images and a hybrid domain adaptation deep learning time series model. Biosyst. Eng. 2024, 237, 220–231. [Google Scholar] [CrossRef]
  46. Raj, N.; Prakash, R. Assessment and prediction of significant wave height using hybrid CNN-BiLSTM deep learning model for sustainable wave energy in Australia. Sustain. Horiz. 2024, 11, 100098. [Google Scholar] [CrossRef]
  47. Song, C.; Cao, J.; Zhao, Q.; Sun, S.; Xia, W.; Sun, L. A high-precision crown control strategy for hot-rolled electric steel using theoretical model-guided BO-CNN-BiLSTM framework. Appl. Soft Comput. 2024, 152, 111203. [Google Scholar] [CrossRef]
  48. Song, B.; Liu, Y.; Fang, J.; Liu, W.; Zhong, M.; Liu, X. An optimized CNN-BiLSTM network for bearing fault diagnosis under multiple working conditions with limited training samples. Neurocomputing 2024, 574, 127284. [Google Scholar] [CrossRef]
  49. Liu, M.; Lu, Y.; Long, S.; Bai, J.; Lian, W. An attention-based CNN-BiLSTM hybrid neural network enhanced with features of discrete wavelet transformation for fetal acidosis classification. Expert Syst. Appl. 2021, 186, 115714. [Google Scholar] [CrossRef]
  50. Zhu, X.J.; Li, H.L.; Lu, X.Q. An improved Attention-Based LSTM feature selection model. J. Beijing Inf. Sci. Technol. Univ. 2018, 33, 54–59. [Google Scholar]
  51. Lin, J.; Ma, J.; Zhu, J.; Cui, Y. Short-term load forecasting based on LSTM networks considering attention mechanism. Int. J. Electr. Power Energy Syst. 2022, 137, 107818. [Google Scholar] [CrossRef]
  52. Liu, G.; Ke, A.; Wu, X.; Zhang, H. GAN with opposition-based blocks and channel self-attention mechanism for image synthesis. Expert Syst. Appl. 2024, 246, 123242. [Google Scholar] [CrossRef]
  53. Wang, Y.; Pu, J.; Miao, D.; Zhang, L.; Zhang, L.; Du, X. SCGRFuse: An infrared and visible image fusion network based on spatial/channel attention mechanism and gradient aggregation residual dense blocks. Eng. Appl. Artif. Intell. 2024, 132, 107898. [Google Scholar] [CrossRef]
  54. Zhu, X.; Nie, S.; Wang, C.; Xi, X.; Lao, J.; Li, D. Consistency analysis of forest height retrievals between GEDI and ICESat-2. Remote Sens. Environ. 2022, 281, 113244. [Google Scholar] [CrossRef]
  55. Lang, N.; Jetz, W.; Schindler, K.; Wegner, J.D. A high-resolution canopy height model of the Earth. Nat. Ecol. Evol. 2023, 7, 1778–1789. [Google Scholar] [CrossRef] [PubMed]
Forests 15 01132 g001
Figure 1. Research area map of the Henan section of the Yellow River Basin showing (a) the Henan section of the Yellow River Basin and partial areas of Luoyang City, (b) the Luoyang City area within the Yellow River Basin and the distribution of GEDI point clouds trajectory, and (c) the Henan section of the Yellow River Basin and the distribution of GEDI point clouds trajectory.
Figure 1. Research area map of the Henan section of the Yellow River Basin showing (a) the Henan section of the Yellow River Basin and partial areas of Luoyang City, (b) the Luoyang City area within the Yellow River Basin and the distribution of GEDI point clouds trajectory, and (c) the Henan section of the Yellow River Basin and the distribution of GEDI point clouds trajectory.
Forests 15 01132 g001
Forests 15 01132 g002
Figure 2. Experimental flowchart.
Figure 2. Experimental flowchart.
Forests 15 01132 g002
Forests 15 01132 g003
Figure 3. CNN-BiLSTM canopy height prediction model (CNN layer and BiLSTM layer).
Figure 3. CNN-BiLSTM canopy height prediction model (CNN layer and BiLSTM layer).
Forests 15 01132 g003
Forests 15 01132 g004
Figure 4. Attention layer network structure.
Figure 4. Attention layer network structure.
Forests 15 01132 g004
Forests 15 01132 g005
Figure 5. Land cover classification results of (a) the Henan section of the Yellow River Basin and (b) the Luoyang region in the Yellow River Basin. (c) Enlarged view of land cover classification results in the forest area of the Luoyang region in the Yellow River Basin.
Figure 5. Land cover classification results of (a) the Henan section of the Yellow River Basin and (b) the Luoyang region in the Yellow River Basin. (c) Enlarged view of land cover classification results in the forest area of the Luoyang region in the Yellow River Basin.
Forests 15 01132 g005
Forests 15 01132 g006
Figure 6. Vegetation index extraction results ((a): original image. (b): WI. (c): MSR. (d): NDVI. (e): SLAVI. (f): EVI. (g): BI. (h): DVI. (i): GVI. (j): RVI. (k): SAVI).
Figure 6. Vegetation index extraction results ((a): original image. (b): WI. (c): MSR. (d): NDVI. (e): SLAVI. (f): EVI. (g): BI. (h): DVI. (i): GVI. (j): RVI. (k): SAVI).
Forests 15 01132 g006
Forests 15 01132 g007
Figure 7. Spatial distribution characteristics map of discrete CHM data in the Luoyang City area in the Yellow River Basin.
Figure 7. Spatial distribution characteristics map of discrete CHM data in the Luoyang City area in the Yellow River Basin.
Forests 15 01132 g007
Forests 15 01132 g008
Figure 8. Model accuracy comparison chart. (a–e): Comparison of true values and predicted values for the RF, SVM, BP-ANN, CNN-BiLSTM, and CNN-SCA-BiLSTM models.
Figure 8. Model accuracy comparison chart. (a–e): Comparison of true values and predicted values for the RF, SVM, BP-ANN, CNN-BiLSTM, and CNN-SCA-BiLSTM models.
Forests 15 01132 g008
Forests 15 01132 g009
Figure 9. Spatial distribution map of canopy height in the Henan Section of the Yellow River Basin.
Figure 9. Spatial distribution map of canopy height in the Henan Section of the Yellow River Basin.
Forests 15 01132 g009
Table 1. Common data sources and methods for forest canopy height retrieval.
Table 1. Common data sources and methods for forest canopy height retrieval.
Data SourceReferencesModel MethodsReferences
Low ResolutionMODIS(Sawada et al. [11])Linear RegressionPopescu et al. [27]
Medium ResolutionSentinel-2(Torresani et al. [12]; Liu et al. [13])XgbtreeRajit et al. [35]
Landsat-8(Ahmed et al. [8]; Liao et al. [9];
Potapov et al. [10])		MARS
High ResolutionAirborne Lidar(Brüllhardt et al. [14]; Ni et al. [17];
Yu et al. [18])		RF(Simard et al. [28]; Shufan et al. [29])
GEDI(Lin et al. [24] Wang et al. [25])SVM(Gleason et al. [33]; Rajit et al. [35])
ICEsat-2(Pang et al. [20]; Mulverhill et al. [21]; Matasci et al. [22]; Dong et al. [23])BP-ANN(Lin et al. [34]; LIN et al. [24])
Table 2. GEDI data acquisition information sheet.
Table 2. GEDI data acquisition information sheet.
Data TypeResearch AreaAcquisition PeriodNumber of Files
GEDI L2ALuoyang City1 January 2022–28 September 202258
Henan Section, Yellow River Basin1 June 2022–1 August 202226
Table 3. Landsat-8 OLI data acquisition information sheet.
Table 3. Landsat-8 OLI data acquisition information sheet.
Product LevelResearch AreaAcquisition PeriodPixel SizeCloud Cover (%)Swath Width (km)Number of Files
Level 1TLuoyang City26 December 202130 × 301185 × 1851
Henan Section, Yellow River Basin1 July 2022–1 September 20226
Table 4. GEDI L2A parameter extraction sheet.
Table 4. GEDI L2A parameter extraction sheet.
Parameter DescriptionSource
lat_lowestmode_a4Photon latitude/geolocation/lat_lowestmode_a4
lon_lowestmode_a4Photon longitude/geolocation/lon_lowestmode_a4
quality_flag_a4Photon mass (0 poor, 1 good)/geolocation/quality_flag_a4
degrade_flagSatellite lifting orbit (0 Lorbit, 1 descending orbit)/degrade_flag
sensitivity_a4Canopy coverage/geolocation/sensitivity_a4
elev_lowestmode_a4Ground elevation/geolocation/elev_lowestmode_a4
elev_highestreturn_a4Canopy elevation/geolocation/elev_highestreturn_a4
rh_a4Tree height (RH99)/geolocation/rh_a4
Table 5. Formulae for calculating the vegetation index.
Table 5. Formulae for calculating the vegetation index.
Vegetation IndexExplanationCalculation Formula
DVIAssess vegetation growth status and health statusb5−b4
BIVegetation index combined with brightness information3029 × b2 + 0.2786 × b3 + 0.4733 × b4 + 0.5599 × b5 + 0.508 × b6 + 0.1872 × b7
GVIAssess the amount, density, and health of green vegetation in a vegetated area−0.2941 × b2 − 0.243 × b3 − 0.5424 × b4 + 0.7276 × b5 + 0.0713 × b6 − 0.1608 × b7
RVIAssess vegetation characteristics in vegetated areasb5/b4
EVIAssess vegetation growth status and health status in vegetated areas2.5 × (b5 − b4)/(b5 + 6 × b4 − 7.5 × b2 + 1)
SAVIThe effects of soil background and atmospheric disturbances on the vegetation index were evaluated2 × (b5 − b4)/(b5 + b4 + 1)
NDVIAssess the growth status and health status of vegetated areas(b5 − b4)/(b5 + b4)
SLAVIModified soil vegetation indexb5/(b4 + b6)
MSRModified ratio vegetation indexRVI × (1 − (b6 − min(b6))/(max(b6) − min(b6))
WIEvaluation of land surface moisture in remote sensing images0.1511 × b2 + 0.1973 × b3 + 0.3283 × b4 + 0.3407 × b5 − 0.7117 × b6 − 0.4559 × b7
Table 6. Comparison sheet of the five models.
Table 6. Comparison sheet of the five models.
ModelTraining SetTest SetTraining Time
MAERMSER2MAERMSER2
RF2.5827493.3434860.8027292.470783.2103330.8087545′58.16″
SVM2.5160923.2585580.8126232.41663.1815260.8121716′22.37″
BP-ANN2.2219462.7945830.8621842.1641032.7436250.86031810′52.13″
CNN-BiLSTM1.6554182.2392550.9115151.6900932.3994980.8931621′44.21″
CNN-SCA-BiLSTM1.3506781.7612990.9452571.3538671.7574380.9426822′20.15″
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Zhao, Z.; Jiang, B.; Wang, H.; Wang, C. Forest Canopy Height Retrieval Model Based on a Dual Attention Mechanism Deep Network. Forests 2024, 15, 1132. https://doi.org/10.3390/f15071132

AMA Style

Zhao Z, Jiang B, Wang H, Wang C. Forest Canopy Height Retrieval Model Based on a Dual Attention Mechanism Deep Network. Forests. 2024; 15(7):1132. https://doi.org/10.3390/f15071132

Chicago/Turabian Style

Zhao, Zongze, Baogui Jiang, Hongtao Wang, and Cheng Wang. 2024. "Forest Canopy Height Retrieval Model Based on a Dual Attention Mechanism Deep Network" Forests 15, no. 7: 1132. https://doi.org/10.3390/f15071132

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop