Mapping the Distribution and Dynamics of Coniferous Forests in Large Areas from 1985 to 2020 Combining Deep Learning and Google Earth Engine

Abstract

Mapping the distribution of coniferous forests is of great importance to the sustainable management of forests and government decision-making. The development of remote sensing, cloud computing and deep learning has provided the support of data, computing power and algorithms for obtaining large-scale forest parameters. However, few studies have used deep learning algorithms combined with Google Earth Engine (GEE) to extract coniferous forests in large areas and the performance remains unknown. In this study, we thus propose a cloud-enabled deep-learning approach using long-time series Landsat remote sensing images to map the distribution and obtain information on the dynamics of coniferous forests over 35 years (1985–2020) in the northwest of Liaoning, China, through the combination of GEE and U²-Net. Firstly, to assess the reliability of the proposed method, the U²-Net model was compared with three Unet variants (i.e., Resnet50-Unet, Mobile-Unet and U-Net) in coniferous forest extraction. Secondly, we evaluated U²-Net’s temporal transferability of remote sensing images from Landsat-5 TM, Landsat-7 ETM+ and Landsat-8 OLI. Finally, we compared the results obtained by the proposed approach with three publicly available datasets, namely GlobeLand30-2010, GLC_FCS30-2010 and FROM_GLC30-2010. The results show that (1) the cloud-enabled deep-learning approach proposed in this paper that combines GEE and U²-Net achieves a high performance in coniferous forest extraction with an F1 score, overall accuracy (OA), precision, recall and kappa of 95.4%, 94.2%, 96.6%, 95.5% and 94.0%, respectively, outperforming the other three Unet variants; (2) the proposed model trained by the sample blocks collected from a specific time can be applied to predict the coniferous forests in different years with satisfactory precision; (3) Compared with three global land-cover products, the distribution of coniferous forests extracted by U²-Net was most similar to that of actual coniferous forests; (4) The area of coniferous forests in Northwestern Liaoning showed an upward trend in the past 35 years. The area of coniferous forests has grown from 945.64 km² in 1985 to 6084.55 km² in 2020 with a growth rate of 543.43%. This study indicates that the proposed approach combining GEE and U²-Net can extract coniferous forests quickly and accurately, which helps obtain dynamic information and assists scientists in developing sustainable strategies for forest management.

1. Introduction

In recent years, extreme weather and climate events have occurred frequently around the world, posing a huge threat to the sustainable development of mankind. As a part of the terrestrial ecosystems, forests cover 31 percent of the global land area [1], and they play an essential role in maintaining the balance of hydrological cycles, promoting the climate regulation and protecting the biodiversity [2,3,4]. As an integral part of the forests, coniferous forests can grow well in harsh environments with a strong stress resistance, and they also have ecological, industrial and medicinal values [5,6,7,8]. Therefore, coniferous forests are widely used for afforestation, which can be of great significance for absorbing the greenhouse gases, soil and water conservation, wind prevention and sand fixation [9,10,11]. The Northwestern Liaoning of China is adjacent to the Horqin Sandy Land in Inner Mongolia, China [12]. The harsh ecological environment (soil erosion, high accumulated temperature, low precipitation, etc.) seriously affects the lives of local residents. The coniferous forests distributed in Northwestern Liaoning include natural forests and artificial forests [13]. The artificial coniferous forests are formed by artificial afforestation and aerial seeding afforestation [14]. The plant density is uneven and the forest age structure is complex, including young forests, middle-aged forests, mature forests and over-mature forests [15]. Coniferous forests are often disturbed by natural factors (fires, pests and diseases, drought, etc.) and human factors (deforestation and reclamation, illegal occupation of forest land, etc.) [16]. It is critical to calculate the area of coniferous forests quickly and accurately for formulating the forest management strategies and government decision-making.

Remote sensing technology has been widely used to measure forest structural and compositional attributes (tree species, tree height, diameter at breast height, crown width, etc.) due to its wide observation range, high temporal resolution and fewer restrictions on the ground environment. Until now, the images from satellites, such as Landsat, Sentinel-1/2 and Gaofen-1 have been publicly available, and they have been widely used in the field of tree species classification and identification. For example, Qi et al. [17] successfully mapped the distribution of bamboo forest in China using Landsat-8 images. Xie et al. [18] mapped the distribution of six tree species in an area of 90,000 km² on Sentinel-1 and 2 satellite remote sensing images; Luo et al. [19] used different algorithms to achieve the extraction of mango trees in Hainan Province, China, using Gaofen-1 images. However, as Sentinel-1/2 and Gaofen-1 were launched post 2013, it is difficult to obtain long-term dynamic information of coniferous forests by using them. Since the first Landsat series satellites were launched in 1972, they have collected Earth-observing data covering the world for 40 years so far, making it possible to monitor and analyze the long-term spatiotemporal dynamic of the forest resources [20].

In recent years, many studies have focused on the impact of forest change on ecological issues [21,22,23,24,25]. The methods of obtaining the forest disturbance and degradation information using remote sensing technology can be divided into two categories at present: image-to-image change detection and time series analysis-based change detection. Various time series analysis methods have been developed over the last few decades (e.g., Landsat-based detection of trends in disturbance and recovery (LandTrendr); continuous change detection and classification (CCDC); continuous degradation detection (CODED); multi-variate time-series disturbance detection (MTDD); and breaks for additive seasonal and trend (BFAST)) [26]. For example, Hua et al. [27] assessed the forest cover dynamics in subtropical China during 1986–2019 by integrating time-series Landsat images and LandTrendr. Chen et al. [28] presented an approach, CCDC-spectral mixture analysis (CCDC-SMA), that combined time series analysis and spectral mixture analysis for monitoring abrupt and gradual forest degradation in temperate regions. Wu et al. [29] used BFAST and monthly Landsat time series for monitoring spatiotemporal dynamics of forests in a subtropical wetland. Reygadaset al. [30] compared the performance of three algorithms–CODED, LandTrendr and MTDD to detect and characterize the forest disturbances in the Southwestern Amazon (Ucayali, Peru and Acre, Brazil) during the 2000–2020 period. Compared with the time series analysis-based methods, image-to-image change detection is much more commonly used for mapping forest dynamics. So far, the scholars have tried to apply many pattern recognition algorithms to map the distribution of tree species. In early studies, maximum likelihood classifiers (MLC), K-means and ISODATA were the most widely used classification techniques. Non-parametric machine learning classification methods, such as support vector machine (SVM), random forest (RF), gradient boosting decision tree (GBDT) and extreme gradient boosting (XGBoost) were subsequently proposed and applied to tree species classification. However, these algorithms rely on image features (spectral features, texture features, etc.) extracted manually with low efficiency and strong subjectivity. Neural networks can automatically extract image features. Sumsion et al. [31] classified tree species and genus at the pixel level using multilayer perceptron, whose performance was better than SVM and RF. Raczko et al. [32] evaluated the three classification algorithms (SVM, RF and ANN) in an attempt to classify the five common tree species, and ANN achieved the highest median overall classification accuracy (77%) followed by SVM with 68% and RF with 62%. Liu et al. [33] used deep neural network (DNN), SVM and RF to extract Chinese pine at large scales, and found that DNN has the highest accuracy for Chinese pine extraction, followed by SVM and RF. Although the neural networks mentioned above achieved better classification results than traditional machine learning methods, they only classify images at the pixel level, without fully considering the relationship between pixels.

Deep convolutional neural networks (CNNs) are commonly used to identify targets and they can extract features at the object level. Zhang et al. [34] adopted an improved three-dimensional convolutional neural network model in tree species classification, and achieved a classification accuracy of 93.14%. Guo et al. [35] developed a deep fusion U-Net for mapping forests at tree species levels. The overall classification accuracy (OA) was 93.30%. Based on the backpack laser scanning data, Liu et al. [36] applied PointNet++ to classify tree species and identify individual tree point clouds. U²-Net is a two-level nested U-structure, which was proposed in the year 2020 [37], and it remains unknown how it performs in coniferous forest extraction at large scales.

Downloading and processing the remote sensing images over large areas using personal computers is time-consuming and laborious. As one of the most excellent cloud-based platforms for planetary-scale geospatial analysis, Google Earth Engine (GEE) has powerful computing power, many excellent image-processing algorithms and massive data, such as satellite imagery, meteorological data, elevation data and atmospheric data [38]. It means that researchers in poorer countries can obtain the same computing power and data as those in the advanced countries. With the easily accessible and user-friendly front-end, GEE provides a convenient environment for data interaction and algorithm development [39]. Users can, not only acquire a large amounts of data, but also can upload their own data and collections. These data can be processed through online programming efficiently and displayed interactively [40]. It has obvious advantages to conduct studies at large scales, as well as in long time series combining remote sensing technology and GEE [41]. The scholars have carried out experiments in fields, such as agriculture [42], wetland [43] and hydrology [40] using GEE.

In this paper, we thus propose a deep learning-based (i.e., U²-Net) approach using Landsat images to map the distribution and dynamics of the coniferous forests in GEE. To achieve this goal, we first used Sentinel-2 A/B, which has a higher resolution than Landsat images, to make sample blocks of coniferous forests in the northwest of Liaoning Province, China. Then, the U²-Net was deployed in GEE to extract the coniferous forests. The results were compared with three deep learning methods (i.e., Resnet50-Unet, Mobile-Unet, and U-Net). To further verify the effectiveness of this method, the extraction results of coniferous forests were compared with other products, such as GlobeLand30, GLC_FCS30 and FROM_GLC30. Following the evaluation of the temporal transferability, the proposed method was employed to obtain the distribution of coniferous forests at 5-year intervals from 1985 to 2020. Finally, the dynamics of coniferous forests in Northwestern Liaoning over the past 35 years were obtained by the image to image change detection method.

2. Materials

2.1. Study Area

The study area (Figure 1) is located in the northwest of Liaoning Province, China, with a total area of about 62,885 km², accounting for 42.49% of Liaoning Province [44,45]. As it is adjacent to Horqin sand, the weather of the study area is windy with a lot of sand. This area belongs to the semi-arid continental monsoon climate in the middle temperate zone, with a dry climate, as well as low and uneven distribution rainfall. The yearly precipitation is about 500 mm. The terrain is complex, with mountains and plains, hills accounting for 30% and 70%, respectively. This study area is rich in mineral resources. Mining and agriculture are the two local pillar industries [46,47,48].

The Northwestern Liaoning is an important ecological barrier area to prevent the Horqin sandy land of Inner Mongolia from invading North China. Due to natural and man-made reasons, the ecological environment in this area is fragile, with serious soil erosion and low forest cover. The large-scale afforestation project in China started in 1978, aiming at preventing the invasion of wind and sand into North China [49]. The afforestation consists of different tree species, such as Chinese pine, Pinus sylvestris, poplar, black locust and apricot [50]. The coniferous forest is mainly composed of Chinese pine and Pinus sylvestris, which are the main afforestation tree species in this area, widely distributed in North China [51]. According to the statistics, the area of Chinese pine in the study area is 433,800 ha², accounting for 62.63% of the whole Chinese pine area of 692,700 ha² in Liaoning Province [52]. However, due to reasons of logging, drought, pathogens and unreasonable forest management measures, forest dieback has occurred frequently [53].

2.2. Landsat Images and Pre-Processing

To obtain the distribution and dynamic change information of coniferous forests in the study area, we selected Landsat-5/7/8 atmospherically corrected surface reflection (SR) at 5-year intervals from 1985 to 2020, as the data source, which was derived from GEE. In the LANDSAT/LT05/C01/T1_SR collection, 253 scenes of Landsat-5 TM remote sensing images were collected, including 60 scenes in 1985, 98 scenes in 1990and 95 scenes in 1995. We collected 324 scenes of Landsat-7 ETM+ images from LANDSAT/LE07/C01/T1_SR, including 129 scenes in 2000, 99 scenes in 2005 and 96 scenes in 2010. The Landsat-8 OLI images obtained in LANDSAT/LC08/C01/T1_SR are 230 scenes, including 117 scenes in 2015 and 113 scenes in 2020 scene. Prior to using these images for coniferous forest extraction, preprocessing operations were required, including cloud cover screening, band selection, spectral index calculation, date selection, synthesis and cropping.

In this study, we selected six bands from the Landsat series images with less than 20% cloud coverage, namely blue (0.45–0.52 μm), green (0.52–0.60 μm), red (0.63–0.69 m), near-infrared (0.77–0.90 μm), shortwave infrared-1 (1.55–1.75 μm) and shortwave infrared-2 (2.08–2.35 μm). Their specific information is shown in

Table 1. The information of Landsat images used in this study.

Satellites	Sensors	Bands	Time	Spatial Resolution
Landsat 5	TM	B1, B2, B3, B4, B5, B7	1985, 1990, 1995	30 m
Landsat 7	ETM+	B1, B2, B3, B4, B5, B7	2000, 2005, 2010	30 m
Landsat 8	OLI	B2, B3, B4, B5, B6, B7	2015, 2020	30 m

. Spectral indices are effective tools to measure objects’ status on the surface of the Earth [54], which can help enhance the separability between ground features further. Therefore, four spectral indices, including the normalized difference vegetation index (NDVI), normalized difference water index (NDWI), enhanced vegetation index (EVI) and modified soil-adjusted vegetation index (MSAVI), were calculated and added to the images mentioned above. The formulas are shown in

Table 2. The spectral indices used in the study.

Spectral Indices	Calculation Formula
NDVI	(BNIR − BRed)/(BNIR + BRed)
NDWI	(BGreen − BNIR)/(BGreen + BNIR)
EVI	2.5 × (BNIR − BRed)/(BNIR + 6 × BRed − 7.5 × BBlue + 1)
MSAVI	(2 × BNIR + 1 − sqrt((2 × BNIR + 1)2 – 8 × (BNIR − BRed)))/2

. Farmland, building, water, coniferous forest, other vegetation (grassland and broadleaf forest) and bare land are the six land use and land cover (LULC) types whose spectral reflectance varies with the seasons throughout the year. To select the most appropriate time for the extraction of coniferous forest with high precision, we divide the year into four time periods (i.e., February~April, May~July, August~October, November~January) according to the phenological properties. It can be seen from Figure 2 that the spectral curves of coniferous forests from February to April and November to January are significantly different from those of other objects. Therefore, to obtain the final cloud-free, high-quality images suitable for coniferous forest extraction, we finally selected Landsat remote sensing images from November to April for compositing and clipping in GEE.

2.3. Training and Validation Sample Blocks for Deep Learning Models

The distribution of coniferous forests is scattered in large areas with unclear boundaries. Manual interpretation is a cost-intensive, time-consuming and subjective process, and accuracy cannot be guaranteed. Liu et al. [1] explored the ability of three different machine learning algorithms (i.e., DNN, SVM, RF) for Chinese pine extraction based on point samples, it is shown that the SVM classifier can achieve better results on Sentinel-2 multispectral imagery, with an extraction accuracy of as high as 91.1% after adding spectral indices (i.e., NDVI, NDWI, EVI, MSAVI). Therefore, we selected 20 subregions in the study area (Figure 3), 14 of them were used to train the deep learning models, and four of them were used as the verification data. The size of each subregion is 18 km × 18 km. We mapped coniferous forests in each of the 20 subregions using SVM on Sentinel-2 images, whose spatial resolution is much higher than Landsat, then the extracted results were post-processed, so as to eliminate salt and pepper noise and misclassification. The 20 coniferous forest maps were resampled from 10 m to 30 m, and the Landsat images of the corresponding area were obtained to form sample blocks with the resampled maps. In order to increase the number of sample blocks, the 20 sample blocks were cropped to 256 pixels × 256 pixels size using a random sliding window, and 640 sample blocks with smaller sizes were generated finally, with 512 training sample blocks and 128 validation sample blocks. Some of the sample blocks are shown in Figure 4.

2.4. Test Sample Points

To verify the performance of the deep learning methods in coniferous forest extraction tasks, we collected sample points from three years (i.e., 1990, 2010, 2020) in the study area. Due to the slow growth, the distribution of the coniferous forests will not change significantly in a short time. The field surveys were conducted in Northwest Liaoning Province in August 2020 and 2021. By using Z-survey GNSS geodetic receiver manufactured in China, 698 sample points were collected. In addition, we collected another 508 points by visual interpretation in Google Earth Pro. A total of 1206 sample points are distributed evenly in the study area (Figure 5). Combined with historical LULC data, the sample points of 1990 and 2010 were collected by visual interpretation in Google Earth Pro. In total, we collected 1199 verification points from 1990 images, and 1230 from 2010 images. Following the merging of the sample points of each year, the attributes of other types (i.e., farmland, building, water, grassland, broadleaf forest and bare land) and coniferous forest were set to 0 and 1, respectively, and then the merged sample points data were uploaded to GEE in the shapefile format. The information of the sample points is shown in

Table 3. The detailed information of the test sample points. Sample points in three periods (1990, 2010 and 2020) were collected.

Periods	Coniferous Forests	Other Types			Total
	Chinese Pine, Pinus sylvestris var. mongolica	Water	Building	Farmland, Grassland, Bare Land, Other Vegetation
1990	531	200	212	256	1199
2010	520	200	232	278	1230
2020	501	200	207	298	1206

.

3. Methods

To realize the coniferous forest extraction at a large extent using deep learning methods in GEE, three products of Google, namely GEE, Google Cloud Platform (GCP) and Google Colaboratory (Colab) were used.

3.1. Workflow Description

Figure 6 shows visually the workflow of the proposed method based on the Landsat series of remote sensing images. Firstly, we uploaded the external data (i.e., sample points and boundaries) to GEE in shapefile format, and the sample blocks were generated using the external data combined with SVM (Section 2.4). All of these data were stored in Google Assets. Then, the sample blocks were transferred from Google Assets in GeoTIFF format to Cloud Storage Bucket in TFRecord format. Cloud Storage Bucket is one of the modules in GCP. The converted sample blocks in TFRecord format were then used to train the deep learning models, and the computing power was provided by GPUs in Colab. We stored the trained parameters in Cloud Storage Bucket; before the trained models can be used in GEE, they need to be hosted by an AI platform by converting the format, all of these operations were carried out using Tensorflow in Google Colaboratory notebook [55]. In GEE, we collected Landsat images every five years from 1985 to 2020. By calling the trained deep learning models in the AI platform, the coniferous forest probability maps were generated, and pixels with a probability higher than 0.5 were labeled as coniferous forest and the remaining pixels were labeled as other objects in the map results. Additionally, we conducted the accuracy evaluation, area calculation and then revealed the long-term dynamic changes of coniferous forest in Northwestern Liaoning Province.

In this article, we employed a novel deep learning method named U²-Net to extract the coniferous forests of the study area. Another three deep learning methods (Resnet50-Unet, Mobile-Unet, and U-Net) were selected for comparison with U²-net. The same hyperparameters were set for all of these deep learning methods. The batch size and epochs were set to 16 and 50, respectively. Cross entropy was selected as the loss function. Adam was chosen as the optimizer with an initial learning rate = 0.001, and the parameters, such as momentum and decay were set to 0.9 and 0.000001, respectively.

3.2. U²-Net

U²-Net is a novel deep-learning model which was first proposed by Qin in 2020 for salient object detection (SOD). It has been proved by experiments that U²-Net performs much better than the other state-of-the-art methods on six public datasets [37]. In recent years, this model has already been used by scholars to extract olive crown information [56], building outlines [57], coastal aquaculture ponds [58] and to segment the median nerve [59]. The results show that U²-Net can achieve satisfactory results with high precision and outperform the mainstream deep learning models, such as HRNet, U-Net, DeepLabv3+, Segnet, U-Ne, and FCN.

3.2.1. The Architecture of U²-Net

U²-Net is a two-level nested U-structure without using any pre-trained backbones from image classification [37], and it was developed based on Unet. As shown in Figure 7, the U²-Net model consists of 11 blocks independent from each other, including six encoder blocks and five decoder blocks. In each block, a Unet-like structure called ReSidual U-block (RSU) is built. In the first four encoder stages (i.e., En_1, En_2, En_3 and En_4), the contextual information of different scales is obtained by using the convolution and down sampling operations, both of which are helpful for expanding the receptive fields. En_5 and En_6 are slightly different from En_1~4 in structure. As the feature maps’ resolution in the last two encoder stages (En_5, En_6) are relatively low, they are not suitable for further down sampling operations, which will result in the loss of useful contextual information. To solve this problem, dilated convolution is employed to replace the max-pooling operation. De_5, one of the decoder blocks, has the same structure as En_5 and En_6. De_1~4 have the same structure as En_1~4, respectively. In each decoder stage (i.e., De_1~5), the up sampled feature maps from its previous stage are concated with its symmetrical encoder stage, and the feature maps generated will be taken as the input of the next decoder stage. In the last part of the U²-Net, six saliency probability maps (i.e., S⁽¹⁾side, S⁽²⁾side, S⁽³⁾side, S⁽⁴⁾_side, S⁽⁵⁾_side and S⁽⁶⁾_side) are generated by using 3 × 3 convolution layer, sigmoid function and up sampling operations on the decoder stages De_1, De_2, De_3, De_4, De_5 and En_6. These saliency probability maps with a single channel have the same size as the input image, and they are then fused with a concatenation operation followed by a convolution layer and an activation function to generate the final coniferous forest probability map. Finally, pixels in the probability map with a probability higher than 0.5 were labeled as coniferous forests and the remaining pixels were labeled as non-coniferous forests in the map results [60].

3.2.2. Residual U-Blocks

Deep learning has been proven to be an excellent technique in numerous scientific fields [61]. CNNs are a branch of deep learning algorithms and are specifically used for image recognition and tasks that involve the processing of pixel data. Features are essential elements for deep learning, which can be obtained automatically by computer. Although they have been frequently used in CNNs, the convolutional filters with the size of 1 × 1 or 3 × 3 can only extract local contextual information. Under the circumstances, a light structure called residual U-blocks (RSU) was proposed in U²-Net, which can extract contextual information locally and globally. As is shown in Figure 8, the Unet-like structure of RSU-L(C_in, M, C_out) can obtain multi-scale contextual information through the encoder and decoder operations, where L is the number of layers in the encoder, C_in, C_out denotes input and output channels, and M denotes the number of channels in the internal layers of RSU. Firstly, a convolution layer is used to transform the input feature map (H × W × C_in) to an intermediate map (H × W × C_out) containing local features. Secondly, a Unet-like module with the height of L takes the feature map F1(x) as input, which is the core part of RSU. The large L is the richer local and global information that can be extracted from the feature maps. Operations, such as convolution, batch normalization, activation, max-pooling, bilinear interpolation and concatenation integrated into the encoder-decoder structure are used to obtain multi-scale features. Finally, the local features and the multiscale features are fused by a residual connection.

3.3. Unet Architecture Variants

U²-Net belongs to one of the Unet architecture variants, whose excellent performance has been proven by experiments [56,57,58,59]. To evaluate the feasibility of using U²-Net in coniferous forest extraction objectively, three U-Net architecture variants, namely U-Net, Resnet50-Unet and Mobile-Unet are were employed. Details of the three variants are illustrated below.

3.3.1. U-Net

The original U-Net [62] was first proposed by Ronneberger in 2015 with five stages from 572 × 572 pixels up to 28 × 28 pixels. The input image has only one single channel. In our adapted variants of U-Net, the stages are increased from five to six compared with the original architecture. The input image measures 224 × 224 × 10 pixels. As is shown in Figure 9, each of the six stages has two convolutional layers composed of operations, such as convolution, batch normalization and relu. The size of the feature map in each encoder stage is halved (i.e., 224, 112, 56, 28, 14, 7) and the channel number becomes twice as much as the previous stage (i.e., 16, 32, 64, 128, 256, 512). These are achieved by using max-pooling and convolution. The size of the last encoder stage is 7 × 7 pixels. The other indispensable branch, namely decoder stages, has similar structures to encoder stages. The first feature maps in the decoder stages are generated by concating up sampled feature maps from its previous stage and those from its symmetrical encoder stage. To restore the size equal to the input image, the last feature maps of each decoder stage are up sampled to a given size to its double, stage by stage from 7 × 7 to 224 × 224.

3.3.2. Resnet50-Unet

He et al. [63] introduced a deep residual learning framework called residual blocks in the year 2016. Resnet50 is one of the typical resnets, which has as many as 50 layers, including a convolution layer (size of 1 × 1 and 3 × 3), batch normalization, activation, max-pooling and shortcut connection. Resnet50 can learn more features from input images than models with shallow depth. In order to combine the advantages of resnet50 and Unet, a variant of Unet named Resnet50-Unet was proposed [64,65], in which the encoder stages of Unet were replaced with a resnet50, and the decoder of the Unet architecture remained the same. The input image measures 224 × 224 × 10 pixels and the code size is 7 × 7 pixels. Six stages are constructed in the model. The component diagram of the adapted U-Net architecture is shown in Figure 10.

3.3.3. Mobile-Unet

In 2018, a lightweight network named MobileNetV2 was proposed by Sandler et al. [66] from Google. A novel layer module, the inverted residual with linear bottleneck is introduced in MobileNetV2 which is suitable for devices with constrained computational power. Mobile-Unet is a variant based on MobileNetV2 and Unet, which has been used in areas, such as fabric defect detection [67], mapping sugarcane [68], hair segmentation [69] and so on. In order to balance fewer parameters and better accuracy, the Unet encoder stages are replaced with MobileNetV2. As is shown in Figure 11, the Mobile-Unet architecture proposed in this paper has six stages from 224 × 224 pixels up to 7 × 7 pixels. The 17 inverted residuals and linear bottlenecks are used, which can extract features with little consumption of computing resources and storage.

3.4. Accuracy Assessment

Deep learning methods, such as U²-Net, Resnet50-Unet, Mobile-Unet and U-Net are employed in this paper to map the distribution of coniferous forests. To measure their performance objectively and quantitatively, five commonly used accuracy evaluation metrics, including overall accuracy (OA), recall, precision, kappa and F1 score are calculated based on the obtained data from the field survey and visual interpretation, which have been mentioned in detail in Section 2.4.

True positive (TP), false positive (FP), true negative (TN) and false negative (FN) are four essential parameters for accuracy evaluation, which can be obtained for the error matrix. TP is an outcome where the model correctly predicts the positive class. TN is an outcome where the model correctly predicts the negative class. FP is an outcome where the model incorrectly predicts the positive class. FN is an outcome where the model incorrectly predicts the negative class. OA represents the proportion of correctly classified samples to all samples, see Equation (1). The recall represents the proportion of positive samples that are predicted correctly to the total positive samples, see Equation (2). Precision represents the proportion of positive samples predicted correctly to all predicted positive samples, see Equation (3). As a comprehensive indicator, F1 score can balance the conflict between precision and recall, which is the harmonic mean of precision and recall, see Equation (4). The kappa coefficient is an indicator for consistency testing.

$$\mathrm{OA}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}$$

(1)

$$\mathrm{Recall}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

(2)

$$\mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$

(3)

$$\mathrm{F}1=2\times \frac{\left(\mathrm{Recall}+\mathrm{Precision}\right)}{\left(\left(\mathrm{Recall}-\mathrm{Precision}\right)\right)}$$

(4)

$$\mathrm{kappa}=\frac{N{\displaystyle {\displaystyle \sum }_{i=1}^n}{x}_{ii}-{\displaystyle {\displaystyle \sum }_{i=1}^n}\left(x_i+x_{i+1}\right)}{N^2-{\displaystyle {\displaystyle \sum }_{i=1}^n}\left(x_i+x_{i+1}\right)}$$

(5)

4. Results and Analysis

4.1. Validation of the Models’ Performance

To assess the applicability of using GEE and U²-Net for large-area coniferous forest extraction, the Landsat 8 OLI images covering Northwest Liaoning from November 2020 to April 2021 were obtained, and then we composited them into a better-quality images through the functions in GEE. Based on this, four deep learning algorithms, namely U²-Net, Resnet50-Unet, Mobile-Unet and U-Net, were used to extract the coniferous forest in Northwest Liaoning, and the results were assessed quantitatively and qualitatively.

Five metrics were used in this paper, including F1 score, precision, recall, OA and kappa. It can be seen from

Table 4. The coniferous forest extraction accuracy of the deep learning models.

Models	F1-Score	Precision	Recall	OA	Kappa
U2-Net	0.954	0.942	0.966	0.955	0.94
Resnet50-Unet	0.943	0.923	0.964	0.946	0.921
Mobile-Unet	0.932	0.904	0.952	0.932	0.893
U-Net	0.928	0.896	0.899	0.933	0.897

that U²-Net performed better than the other three deep learning models under the same conditions, achieving an F1 score of 95.4%, as well as precision and recall of 94.2% and 96.6%, respectively. OA and kappa of the U²-Net were 95.5% and 94%, respectively. Among these deep learning models, the performance of U-Net was slightly inferior compared to the others in coniferous forest extraction, whose F1 score was 92.8%, precision was 89.6% and recall was 89.9%. The OA and kappa of U-Net were 93.3% and 89.7%, respectively. As the encoder stages were composed of Resnet50 and MobilenetV2, Resnet50-Unet and Mobile-Unet had stronger feature extraction capabilities and obtained higher accuracy than U-Net. Compared with Resnet50-Unet, Mobile-Unet and U-Net, U²-Net improved the F1 score by 1.1%, 2.2% and 2.6%, respectively.

Four typical areas were selected to display the classification results of coniferous forests. The extraction results of coniferous forests using the deep learning methods are shown in Figure 12. The first row and the second row contain low and medium-density coniferous forests, which are plantation forests and natural forests, respectively. Coniferous forests with a high density are distributed in the third and fourth rows, where the third row is a mixture of natural and plantation forests, and the fourth row is a natural forest. It can be seen that all of these four deep learning methods can achieve the purpose of extracting coniferous forests, but results varied depending on the methods. U²-Net performed better in the extraction tasks of different coniferous forests (i.e., low and medium density, high density, plantation and natural forest), and achieved a high consistency with labels. This might benefit from the strong feature extraction ability of the RSU structure of U²-Net. The result obtained by Resnet50-Unet was acceptable, but there was still a certain gap with U²-Net in terms of extraction accuracy through careful observation. Some scattered coniferous forests cannot be recognized completely by Resnet50-Unet, which reflected indirectly that the RSU performed better than the residual block in the feature extraction. The extraction results of Mobile-Unet and U-Net were significantly different from the labels. Only part of the coniferous forest can be identified correctly by Mobile-Unet, while U-Net misclassified the other objects as coniferous forests, resulting in the extracted coniferous forest area being much larger than the actual one.

4.2. Temporal Transferability Evaluation of U²-Net

Due to the wide and scattered distribution of coniferous forests, it is difficult to guarantee the accuracy of the sample blocks based on visual interpretation and manual extraction, which are cost-intensive and time-consuming, let alone for the extraction of coniferous forests in large areas. Therefore, we used Sentinel-2 A/B and Landsat 8 OLI remote sensing images combined with the SVM classifier and post-processing to obtain the sample blocks used to train deep learning models. The remote sensing images from November 2020 to April 2021 were selected. The U²-Net deep learning model was trained based on these sample blocks (Section 2.4), and then we obtained the coniferous forest distribution map of Northwest Liaoning in the year 2020. However, it remained unknown whether the trained model could be used to predict coniferous forests in remote sensing images collected from the years 1985–2020 and sensors of Landsat 7 ETM+ and Landsat 5 TM. Therefore, we used the U²-Net deep learning model trained by the 2020 data to extract the coniferous forests in images of Landsat-5 TM collected in 1990 and Landsat-7 ETM+ collected in 2010. We then assessed the results qualitatively and quantitatively.

First of all, we directly apply the U²-Net model fully trained by sample blocks generated from satellite images in the year 2020. Figure 13 shows the extraction results of coniferous forests in 1990, 2010 and 2020. Three typical areas located in the northeast, southwest and northwest of the study area were selected to display the extraction results. The land cover types in these areas including grassland, farmland, water, coniferous forest, broadleaf forest and bare land. The coniferous forests in the remote sensing images of different years and sensors can be extracted finely, which can be seen in Figure 14.

The sample points of the coniferous forests in 1990, 2010 and 2020 (Table 3) were collected and the same metrics as in Section 4.1 were used to evaluate the temporal transferability of the U²-Net quantitatively (Figure 13). F1 score, precision, recall, OA and kappa of the coniferous forests in 2020 were 95.4%, 94.2%, 96.6%, 95.5% and 94.0%, respectively. The values of these five metrics in 2010 were very similar to those in 2020, with an F1 score of 95.5%, precision of 94.2%, recall of 96.8%, OA of 95.4% and kappa of 94.0%, respectively. Compared with 2020, the coniferous forest extraction accuracy in 1990 had decreased significantly. Among the five metrics, the F1 score was 94.3%, precision was 93.5%, recall was 96.2%, OA was 93.1% and kappa was 0.925, decreased by 0.1%, 0.7%, 0.4%, 2.4%, and 1.5%, respectively. During the 30 years from 1990 to 2020, the changes of climate, ground humidity, atmospheric conditions and other factors influenced the reflectance of coniferous forests, which were the main reason for the decline in extraction accuracy. Through analysis, it can be found that although there is a long time interval (30 years), the highest decline rate is only 2.4% among the five evaluation metrics of three years (1990, 2010, 2020), and all of these accuracy metric values are above 92.5%, which is a satisfactory result. Therefore, it can be judged that the U²-Net deep learning model has good temporal transferability characteristics.

4.3. Comparison with Other Datasets

A great many global land-cover products with spatial resolutions ranging from 10 m to 1000 m have been released since 1990. In this study, three publicly available global land-cover products, namely GlobeLand30 (available online: www.globallandcover.com (accessed on 11 October 2022) [70], GLC_FCS30 (global 30 m land-cover classification with a fine classification system) (available online: https://data.casearth.cn/ (accessed on 12 October 2022) [71] and FROM-GLC (fine resolution observation and monitoring of global land cover) (available online: http://www.geodata.cn/ (accessed on 15 October 2022) [72] were employed to compare with the results obtained by U²-Net. We obtained the coniferous forest distribution in Northwest Liaoning for the year 2010 from the four datasets (i.e., GlobeLand30, GLC_FCS30, FROM-GLC and maps generated by U²-Net) that have the same spatial resolutions. The GlobeLand30 used an approach based on the integration of pixel- and object-based methods with knowledge and achieved an overall accuracy of 75.9%. By using high-quality training data from the global spatial temporal spectra library, the GLC_FCS30 was produced with an overall accuracy of 82.5%. The FROM-GLC is the first 30 m resolution global land-cover map using Landsat-5 TM and Landsat-7 ETM+ of the green season and achieved an overall accuracy of 64.9% by the SVM classifier.

Three representative coniferous forest distribution areas in Northwest Liaoning were selected and the corresponding results come from the four datasets mentioned above are shown in Figure 15. The first column shows the Landsat-7 ETM+ images, and columns 2 to column 5 are coniferous forest maps corresponding to the first column. Through visual interpretation, U²-Net could map the distribution of coniferous forests more completely compared to GlobeLand30, FROM-GLC and GLC_FCS30, which ranked from second to fourth in terms of their performance in coniferous forests extraction, respectively.

Additionally, we calculated the coniferous forest area in Northwest Liaoning in 2010. As shown in Figure 16, the area of U²-Net is 5120.12 km², which is very close to the 5172.44 km² of GlobeLand30. The area of FROM-GLC is 5339.16 km², which is 219.04 km² greater than U²-Net, and 166.72 km² greater than GlobeLand30. However, the area of GLC_FCS30 is 934.18 km², which is quite different from U²-Net, GlobeLand30 and FROM-GLC. This conclusion is consistent with that one based on a visual interpretation. Many factors can affect the extraction accuracy of coniferous forests. Although the remote sensing data sources used in the four datasets were exactly the same (Landsat-7 ETM+), the distribution of training samples, seasonal factors and algorithms can all cause differences in the extraction results of coniferous forests. GlobeLand30, FROM-GLC and GLC_FCS30 are global LULC products. To achieve better classification results, the remote sensing images selected were all from the green season. However, this period was not conducive to the extraction of coniferous forests, and it is easy to confuse coniferous forests with other tree species (poplar, willow, elm, etc.). It had been proved through experiments in Section 2.2 that the best seasons for extracting coniferous forests are winter and spring. The sample points with a higher density used in this study were obtained through field surveys and visual interpretation, which has a higher accuracy than those used in existing LULC data products. In terms of algorithms, GlobeLand30, FROM-GLC and GLC_FCS30 all used traditional machine learning methods, however, the deep learning methods have certain advantages in feature extraction and classification. In summary, the coniferous forests extraction method based on U²-Net in this paper is feasible.

4.4. The Distribution and Dynamics of Coniferous Forests

Through the comparative analysis of the above experiments, the full pipeline (Figure 6) designed using the deep learning model of U²-Net combined with GEE can achieve high-precision and fast extraction of coniferous forests in large areas. Firstly, the coniferous forest mapping deep learning model (U²-Net) was trained using the 2020 training patches. Based on the trained model, we predicted the coniferous forest distribution at 5-year intervals between 1985 and 2020. Figure 17 shows the extraction result of the coniferous forests in 2020. It can be seen from the figure that the distribution of coniferous forests is uneven, and most of the coniferous forests are distributed in the middle, west, southwest and northwest of the study area. It is because this area is dominated by mountains and hills, which are suitable for the growth of coniferous species, such as Pinus tabulaeformis and Pinus sylvestris. The coniferous forests in this area consist of planted and natural forests. The terrain in the eastern and northeastern parts of the study area is flat, and the land use type is mainly agricultural land. Only some planted coniferous forests are distributed in the eastern and northeastern parts of the study area.

We calculated the coniferous forest area in the study area in the years 1985, 1990, 1995, 2000, 2005, 2010, 2015 and 2020, respectively. The results are shown in Figure 18. The area of coniferous forests in Northwestern Liaoning showed an upward trend in the past 35 years. In 1985, the area of coniferous forests was 945.64 km², accounting for only 1.50% of the total area of the study area. By 2020, the area of coniferous forests had become 6084.55 km², accounting for 9.68% of the study area. The net increase of coniferous forest area was 5138.78 km² from 1985 to 2020, and the increased coniferous forest area accounted for 8.17% of the study area. During the 35 years, the coniferous forest area in the study area increased by 543.43%. In the two periods of 1985–1990 and 1990–1995, the number of coniferous forests increased rapidly. From 1985 to 1990, the area of coniferous forests increased by 942.40 km². From 1990 to 1995, the forest area increased by 2015.28 km², and by 1995, the total coniferous forest area was 3903.66 km², accounting for 6.21% of the study area. However, the year 2000 was a special time, and it was the only year in which the area of coniferous forests decreased over the 35 years. According to statistics, the area of coniferous forests in 2000 was 2310.29 km², and decrease of 1593.37 km² compared to 1995, the reduced coniferous forest area accounted for up to 2.53% of the administrative area.

By 2005, the area of coniferous forests had rebounded significantly. From 2000 to 2005, the area of coniferous forests increased by as much as 2507.98 km², and reached 4818.27 km². From 2005 to 2020, the area of coniferous forests increased steadily, among which the area of the coniferous forest increased by 301.91 km², 579.69 km² and 384.60 km² during the periods of 2005–2010, 2010–2015 and 2015–2020. The average annual growth rate of coniferous forests in the seven time periods (i.e., 1985–1990, 1990–1995, 1995–2000, 2000–2005, 2005–2010, 2010–2015 and 2015–2020) was 0.30%, 0.65%, −0.51%, 0.80%, 0.10%, 0.19% and 0.12%, respectively. More detailed information about coniferous forests from 1985 to 2020 is shown in

Table 5. The area’s change information concerning coniferous forests from 1985 to 2020.

Years	Increased Area (km2)	Reduced Area (km2)	Net Increased Area (km2)	Percentage of Net Increased Area (%)	Percentage of Disturbed Area (%)	Average Annual Growth Rate (%)
1985–1990	1179	236.6	942.4	1.50	2.27	0.30
1990–1995	2101.66	86.38	2015.28	3.20	3.50	0.65
1995–2000	263.45	1856.45	−1593.00	−2.53	3.39	−0.51
2000–2005	2662.53	154.55	2507.98	3.99	4.51	0.80
2005–2010	818.88	516.97	301.91	0.48	2.14	0.10
2010–2015	1359.43	779.74	579.69	0.92	3.43	0.19
2015–2020	1020.48	635.88	384.60	0.61	2.66	0.12

.

5. Discussion

This study shows the high feasibility of coniferous forest extraction using the proposed cloud-enabled deep-learning approach. We compared four deep learning models (i.e., U²-Net, Resnet50-Unet, MobilenetV2-Unet and U-Net), and the overall accuracy achieved by U²-Net reached over 95%, which was greater than the other three methods. The results were also compared with global land cover products. Thanks to the good temporal transferability of U²-Net, satisfactory coniferous forest maps can be generated from the remote sensing images of different sensors (i.e., TM, ETM+, OLI) and periods. Based on the excellent performance of the proposed approach, the dynamic information of coniferous forests from 1985 to 2020 in the northwest of Liaoning Province, China were obtained quickly and accurately.

5.1. The Combination of GEE with Deep Learning

To our knowledge, it is the first attempt to use GEE combined with deep learning in the area of tree species classification and identification. Although many scholars have tried to deploy deep learning algorithms on personal computers [34,35,36,73], they are often constrained by computing power and storage resources, making it difficult to conduct experiments at large scales. As is mentioned in Section 2.2, we collected a total of 807 Landsat remote sensing images in this study. By using GEE, we can accomplish the collection and preprocessing of the remote sensing images quickly. The total time taken was about 30 min, and it is unimaginable for personal computers. In recent years, deep learning has achieved remarkable performance in various artificial intelligence research [61]. It can automatically learn multi-level and multi-dimensional features in remote sensing images in an end-to-end way, and the process does not require human intervention. Compared with the methods combining GEE with machine learning [18,33], the proposed method of combining GEE with deep learning can achieve more accurate coniferous forests extraction results in a more convenient way, despite the complex deployment process of using GCP, Google Colab and GEE.

5.2. The Performance of U²-Net Model for Coniferous Forest Mapping

Some studies have verified the performance of this U²-Net. For example, Ye et al. [56] used U²-Net to extract olive crown information. Compared with three deep learning models (i.e., HRNet, U-Net and DeepLabv3+), he pointed out that U²-Net had a better performance with an overall accuracy of 95.19%. According to Wei et al. [57], U²-Net can yield better building recognition results than other semantic segmentation models (Segnet, U-Net and FCN). Even so, few scholars have used U²-Net for coniferous forest extraction, let alone deployed it in GEE. Our study indicates that the use of U²-Net can achieve better coniferous forest extraction than other deep learning models (Resnet50-Unet, Mobile-Unet and U-Net). Under the same training parameter settings (learning rate, epocs, etc.), the order of coniferous forest extraction accuracy from large to small is: U²-Net > Resnet50-Unet > Mobile-Unet >U-Net, which shows that the extraction accuracy is directly related to the models’ structure. U²-Net has a two-level nested U-structure, it can extract multi-scale features integrating local information and global information, which is the main reason for its excellent performance.

In addition, U²-Net has good temporal transferability. The sample blocks used to train the deep learning models in this study were obtained in 2020. Then, we used this trained model to extract the coniferous forests at 5-year intervals between 1985 and 2020. The Landsat-7 ETM+ remote sensing images in 2010 and Landsat-5 TM remote sensing images in 1990 were selected to verify the accuracy of the coniferous forests extraction results in the experimental area. From the perspective of visual interpretation, the extraction results are in good agreement with the real distribution of coniferous forests on remote sensing images. From a quantitative perspective, compared with the coniferous forests extraction accuracy in 2020, the accuracy in 2010 has slightly decreased. Among the five metrics of 1990, the F1 score decreased by 1.1%, precision decreased by 0.7%, recall decreased by 0.4%, OA decreased by 2.4%, and kappa decreased by 1.5% compared with those of 2020. Kappa has the lowest value in 1990 (see Figure 15), but it can also reach as high as 92.5%. Even after an interval of up to 30 years (1990–2020), a satisfactory extraction distribution map of coniferous forests can still be obtained. In this study, we selected the same number of bands with the same spectral range from three different types of sensors (OLI, ETM+, TM). Temperature, atmospheric composition and other factors can affect the spectral reflectance of each band in remote sensing images over time, which is the main reason for the change in the extraction accuracy of coniferous forests in different years. U²-Net can learn and extract multi-scale features in remote sensing images, thus can greatly improve the extraction accuracy in the past 35 years.

By comparing with three public global land-cover products with the same spatial resolution of 30 m generated from Landsat series datasets, we found that the distribution of coniferous forests extracted by U²-Net was most similar to that of actual coniferous forests. The distribution of coniferous forests in the other three datasets (GlobeLand30 [70], GLC_FCS30 [71] and FROM-GLC [72]) is different from the real situation to varying degrees (Figure 16). According to statistics, the area of coniferous forests in Northwestern Liaoning calculated using the proposed method is 5120.12 km², 5172.44 km² of GlobeLand30, 5339.16 km² of FROM-GLC and 934.18 km² of GLC_FCS30. The differences among the extraction results are caused by the algorithms, the size and distribution of the samples, extraction strategies and so on. GlobeLand30, FROM-GLC and GLC_FCS30 are all global land-cover products, and they need to accurately identify multiple types of land features in different regions all over the world, which have higher complexity than the extraction task in Northwest Liaoning. In addition, the algorithms used in these three datasets were all traditional machine learning methods with limited feature extraction ability compared with deep learning. However, traditional machine learning methods have small parameters and structures with lower complexity. At this stage, they are common algorithms to achieve global-scale information extraction. Using deep learning methods for LULC mapping on a global scale remains to be studied further.

5.3. The Distribution and Dynamics of Coniferous Forests

The distribution of coniferous forests is uneven and most of them are distributed in the central and western parts of Northwestern Liaoning (Figure 17). According to statistics, the area of coniferous forests in Northwestern Liaoning rose from 942.40 km² in 1985 to 6084.55 km² in 2020 and the net increase area in the past 35 years is 5138.78 km² with a growth rate of 543.43%. The rapid growth of coniferous forests area is mainly due to the Three-North Shelterbelt Project implemented by the Chinese government. To improve the ecological environment and slow down the process of desertification and soil erosion, the Chinese government has invested a lot of money and manpower since 1979 to plant a large number of artificial forests in the northwest, north, and northeast of China [74]. Northwest Liaoning is one of the key construction areas of the project. From 1985 to 1995, the area of coniferous forests increased from 945.64 km² in 1985 to 3903.66 km² in 1995 with a net increase area of 2961.26 km². However, the area in 2000 decreased sharply compared to that in 1995, and the area decreased by 1593.37 km². Deforestation, pests and diseases, urban development, mining and unsustainable forest management are the main reasons for the reduction of coniferous forests. Since 1998, the Chinese government has promulgated a series of laws to prevent coniferous forests from being destroyed by humans, and achieved remarkable results [75]. In 2005, the area of coniferous forests increased by 2507.98 km² to 4818.27 km² compared with 2000. In the following 15 years (2005–2020), the area of coniferous forests increased steadily, with an average annual growth rate of 1.75%.

Following decades of hard work, the ecological environment in Northwestern Liaoning has been significantly improved, effectively promoting the development of local society and economy. The extensive planting of coniferous forests has played an irreplaceable role in the process. To further promote the sustainable development of the ecological environment, the Chinese government has set goals of reaching a carbon peak before 2030 and carbon neutrality before 2050, in 2020 September [76]. Because of the strong carbon sink capacity, coniferous forests will play an important role in the achievement of the goal.

5.4. Limitations and Future Work

This study has some limitations. Firstly, only medium-resolution optical remote sensing images were used in this paper. It has been proven that synthetic aperture radar (SAR) [77] and high spatial resolution (HSR) remote sensing images [78] can work well for tree species classification and identification. However, the potential of combining different types of remote sensing images in coniferous forests extraction has not been explored, which is expected to improve the accuracy further. Secondly, scholars tried to use multi-classifier ensembled methods in recent years to classify the targets [79]. For example, Yang et al. [80] used an ensemble learning classifier integrating Alexnet and RF to recognize wind turbine blade damage, and the result showed that the ensembled method performed much better than Alexnet and RF. It is believed that the combination of deep learning algorithms and traditional machine learning algorithms can also achieve satisfactory results in coniferous forest extraction. Unfortunately, the performance has not been validated by studies, especially in cloud computing platforms. Thirdly, there are a large number of coniferous forests distributed in Northern China, which has made great contributions to the improvement of China’s ecological environment. However, due to experimental needs, this study only selected Northwestern Liaoning as the study area. Finally, we used the image-to-image change detection method to obtain dynamic change information from 1985 to 2020 in this study. However, time series analysis-based change detection methods can obtain continuous dynamic information, and they have been used in vegetation change detection [81,82,83].

Future work will be carried out in the following four aspects. Firstly, we will try to use the combination of the Landsat with other sensors, such as Sentinel-1, ALOS-2, PALSAR-2, Sentinel-2 and Planet to extract coniferous forests, and it is promising for providing much more information than Landsat alone. Secondly, more research is needed to integrate deep learning and traditional machine learning, and this algorithm will then be deployed in GEE to achieve results. Thirdly, in order to obtain more continuous information on the dynamic of coniferous forests, we will used time series analysis-based change detection methods, such as CCDC and LandTrendr instead of the image-to-image method in the future. Finally, the larger and more typical study areas will be selected for coniferous forest extraction.

6. Conclusions

We proposed a cloud-enabled deep-learning approach using long-time series Landsat remote sensing images to map the distribution and dynamics of coniferous forests for as long as 35 years (1985–2020) in the northwest of Liaoning, China. Mapping coniferous forests using U²-Net can achieve satisfactory results, with the F1 Score, OA, precision, recall and kappa as high as 95.4%, 94.2%, 96.6%, 95.5% and 94%, respectively. In addition, U²-Net showed excellent temporal transferability in mapping coniferous forests from 1985 to 2020. In the past 35 years, the area of coniferous forests in the study area has increased dramatically, with a growth rate as high as 543.43%. Although there was a temporary decrease in coniferous forests due to human disturbance, the area of coniferous forests has increased steadily year by year after the implementation of relevant forest protection policies. In general, we employed the U²-Net deep learning model combined with GEE cloud computing platform to extract coniferous forests at large scales, and on this basis, the change information of coniferous forests in the past 35 years has been carried out, which greatly saves manpower and time, and provides support for formulating reasonable forest management strategies and government decision-making.