Improved Prototypical Network Model for Classification of Farmland Shelterbelt Using Sentinel-2 Imagery

Abstract

Farmland shelterbelt plays an important role in protecting farmland and ensuring stable crop yields, and it is mainly distributed in the form of bands and patches; different forms of distribution have different impacts on farmland, which is an important factor affecting crop yields. Therefore, high-precision classification of banded and patch farmland shelterbelt is a prerequisite for analyzing its impact on crop yield. In this study, we explored the effectiveness and transferability of an improved Prototypical Network model incorporating data augmentation and a convolutional block attention module for extracting banded and patch farmland shelterbelt in Northeast China, and we analyzed the potential of applying it to the production of large-scale farmland shelterbelt products. Firstly, we classified banded and patch farmland shelterbelt under different sample window sizes using the improved Prototypical Network in the source domain study area to obtain the optimal sample window size and the optimal classification model. Secondly, fine-tuning transfer learning and learning from scratch directly were used to classify the banded and patch farmland shelterbelt in the target domain study area, respectively, to evaluate the extraction model’s migratability. The results showed that classification of farmland shelterbelt using the improved Prototypical Network is very effective, with the highest extraction accuracy under the 5 × 5 sample window; the accuracies of the banded and patch farmland shelterbelt are 92.16% and 90.91%, respectively. Using the fine-tuning transfer learning method in the target domain can classify the banded and patch farmland shelterbelt with high accuracy, above 95% and 89%, respectively. The proposed approach can provide new insight into farmland shelterbelt classification and farmland shelterbelt products obtained from freely accessible Sentinel-2 multispectral images.

1. Introduction

A farmland shelterbelt is banded or patch forest stand created in or along the edges of farmland and is a type of planted forest ecosystem that can provide a variety of utilities [1,2]. Farmland shelterbelts can not only improve the microclimate of farmland and reduce the adverse effects of natural disasters on farmland through the influence of forest belts on airflow, temperature, moisture, soil, and other environmental factors, but also create an environment conducive to the growth and development of crops and ensure stable and high crop yields [3,4,5,6]. In the process of agricultural modernization, countries around the world regard the farmland shelterbelt system as a key component, especially in the development of ecological agriculture. The construction of farmland shelterbelts plays a crucial and long-term role in improving ecological security and food security [7,8]. Farmland shelterbelt is usually distributed in the form of bands or patches, and their distribution form, area, and structure have different effects and roles on farmland, which are the main factors affecting crop yield [9,10,11,12]. Therefore, rapid and accurate access to the distribution of different forms of farmland shelterbelt is a prerequisite for exploring and analyzing their impact on crop yields and is of great significance for evaluating the protective role of farmland shelterbelts and their effect on improving crop yields.

Field survey is the most accurate method for differentiating farmland shelterbelt, but it consumes huge human, material, and financial resources, especially for differentiating farmland shelterbelt on a large scale. The advantages of remote sensing technology in observing a wide area over a short time and obtaining a large amount of information have caused it gradually to replace the traditional field survey as the main way to identify farmland shelterbelt [13,14,15,16]. In the early stages, farmland shelterbelt was identified by visual interpretation of medium-resolution Landsat imagery [17,18]. Later, with the emergence of high-resolution remote sensing data, this has become the mainstream method to classify farmland shelterbelt by the comprehensive use of spectral, spatial texture, and shape features [19,20,21,22]. At present, the extraction of farmland shelterbelt mainly adopts classification methods, including pixel-based methods such as maximum likelihood, decision tree, and support vector machine, as well as object-oriented classification methods. Wiseman et al. [23] accurately extracted farmland shelterbelt with a 95.8% accuracy by using an object-oriented approach and combining spectral features, shape, texture, and other features of high-resolution aerial images. Liknes et al. [24] utilized 1-m resolution remote sensing imagery combined with an integrated method of image segmentation and random forest to achieve automatic extraction of information on farmland shelterbelt with an accuracy of 84.8%. Based on ZY-3 and Landsat 8 data, combined with spectral and spatial geometric features, Xing et al. [25] comprehensively utilized mathematical morphology and object-oriented methods to realize the rapid and automatic extraction of farmland shelterbelt, and the extraction accuracy was 89.89%. However, the current methods for extracting farmland shelterbelt have certain limitations. Object-oriented approaches have not yet achieved fully automated extraction, while commonly used pixel-based classification methods fail to adequately consider the spatial characteristics of shelterbelt and surrounding features. Both methods lack sufficient transferability and are ineffective in distinguishing between banded and patch farmland shelterbelt. In addition, research on the extraction of farmland shelterbelt is still relatively limited and concentrated on the small scale, and there are few methods applicable to large-scale extraction.

Few-shot learning (FSL) refers to a deep learning task that learns to make accurate predictions by training with a very small number of labeled examples [9,26]. Unlike deep learning networks that are based on a large number of training samples, FSL requires a smaller number of samples. A Prototypical Network is a few-shot algorithm based on metrics, which is mainly used for feature extraction by training a convolutional neural network (CNN), and then the Prototypical Network classifier is utilized to complete the work of few-shot image classification [27]. In the task of classification of farmland shelterbelt, the farmland area is extremely large, which will limit the acquisition of uniformly distributed farmland shelterbelt samples; therefore, the Prototypical Network with its simple structure, fast fitting speed, and transferable learning is suitable for the classification of farmland shelterbelt [28,29]. In the task of classification of farmland shelterbelt, it is difficult to obtain a large number of uniformly distributed samples of farmland shelterbelt due to the very large area of the farmland; therefore, the Prototypical Network with its simple structure, fast training speed, and transferable learning, which is an FSL network, is suitable for the classification of farmland shelterbelt. However, models trained through FSL are particularly susceptible to overfitting [30], which poses a challenge in achieving accurate classification of farmland shelterbelt. Therefore, this study starts from expanding the sample richness and adding feature enhancement to the network model in order to minimize the overfitting phenomenon in the process of extracting farmland shelterbelt.

Here, by comprehensively analyzing the distribution characteristics of two forms of farmland shelterbelt, the advantages and disadvantages of FSL, the issue of model overfitting, and the demand for farmland shelterbelt products, we propose an improved Prototypical Network model, which incorporates data augmentation and a convolutional block attention module, to explore the feasibility of efficiently identifying banded and patch farmland shelterbelt, as well as the potential for producing large-area farmland shelterbelt products. The main objectives of this study are as follows: (a) extract banded and patch farmland shelterbelt with high precision using the proposed improved Prototypical Network; (b) analyze the optimal sample window size for extracting banded and patch farmland shelterbelt; (c) explore the transferability of the extraction model and analyze the feasibility of classification of farmland shelterbelt under different farmland environments to produce a large-scale farmland shelterbelt distribution map.

2. Materials

2.1. Study Area

In this study, Fujin, Youyi, Yi’an, and Hailun in Heilongjiang Province and Fuyu in Jilin Province were used as the study area, as shown in Figure 1. Heilongjiang Province and Jilin Province are located in the northeast of China, with rich arable land resources and fertile land, and are the main distribution areas of black soil in China. In the northeast of Heilongjiang Province is the Sanjiang Plain, and in the west is the Songnen Plain, with a cultivated area of about 17,195,400 hectares. The central part of Jilin Province is the Songliao Plain, with a cultivated area of 7.03 million hectares [31]. Both provinces have a temperate monsoon climate with four distinct seasons and simultaneous rain and heat. Heilongjiang Province and Jilin Province, with corn, rice, and soybeans as the main food crops, are important grain production bases in China [32]. Farmland shelterbelt is widely distributed in the cultivated areas, in the form of banded or patch farmland shelterbelt, and the tree species of the farmland shelterbelt are mainly poplars (Populus L.).

Fujin is used as the study area for the source domain for model training, and Youyi, Yi’an, Hailun, and Fuyu are used as the study areas for the target domain for transfer learning. The complexity of farmland scenarios in different study areas varies; the farmland environment is highly differentiated, and the distribution of farmland shelterbelt has some similarity but is characterized by its own features.

2.2. Satellite Data and Pre-Processing

Satellite images from Sentinel-2 were used to identify banded farmland shelterbelt and patch farmland shelterbelt in complex farmland environments. The Sentinel-2 data has 13 bands covering the visible and near-infrared regions, with resolutions ranging from 10 m to 60 m. In this study, we used bands with spatial resolutions of 10 m (blue, green, red, and near infrared) and 20 m (red-edge 1, red-edge 2, red-edge 3, narrow NIR, short-wave infrared 1, and short-wave infrared 2).

Cloud-free Sentinel-2 Level-1C images were acquired from the Copernicus Open Access Hub (CSDB, https://scihub.copernicus.eu/ (accessed on 11 October 2023)) over the five study areas. The image acquisition dates covering the study areas of Fujin, Youyi, Fuyu, Hailun, and Yi’an are as follows: 29 May 2020; 17 May 2021; 28 May 2023; 9 June 2022; and 28 May 2023. The Level-1C products provided Top-of-Atmosphere reflectance after ortho-rectification and sub-pixel multispectral registration. The Level-2A products giving Bottom-of-Atmosphere reflectance were obtained by atmospheric correction, and the bands with 20-m resolution were resampled to 10-m resolution. Then, the image covering the study area was obtained by splicing and clipping.

2.3. Sample Point Collection

The field survey was conducted in April 2022 in the study area. Samples of banded farmland shelterbelt, patch farmland shelterbelt, road, water, building, crop, and other land-use types were collected, among which all categories except banded and patch farmland shelterbelt were classified into the others category. Combined with the field survey sample points and visual interpretation, sample points of banded farmland shelterbelt, patch farmland shelterbelt, and others were selected in each of the five study areas. The specific number of sample points is shown in the

Table 1. Acquisition of sample points from the five study areas.

Region	Class Name	Training	Testing
Fujin	Banded farmland shelterbelt	150	50
	Patch farmland shelterbelt	100	35
	Others	150	50
Youyi	Banded farmland shelterbelt	35	80
	Patch farmland shelterbelt	30	65
	Others	40	95
Hailun	Banded farmland shelterbelt	35	90
	Patch farmland shelterbelt	35	85
	Others	45	100
Yi’an	Banded farmland shelterbelt	45	85
	Patch farmland shelterbelt	35	80
	Others	45	95
Fuyu	Banded farmland shelterbelt	45	135
	Patch farmland shelterbelt	30	75
	Others	40	120

. Details of the sample data construction method based on the sample points are described in Section 3.2 and Section 3.4.

3. Method

3.1. Overview

In this study, we carried out the classification of farmland shelterbelt based on the improved Prototypical Network constructed; the main research process was as follows: First, in the source domain, the improved Prototypical Network was used to classify banded and patch farmland shelterbelt under different sample windows, aiming to identify the optimal sample window size and the best model. Then, the optimal model was transferred to the target domain to explore the transferability of the constructed network. The specific technical flowchart is shown in Figure 2.

3.2. Architecture of the Improved Prototypical Network

We constructed a classification framework suitable for classifying farmland shelterbelt, which was based on the classification principles of the Prototypical Network and incorporated methods such as data augmentation and a convolutional block attention module (CBAM). Our overall framework is shown in Figure 3.

3.2.1. Prototypical Network

The Prototypical Network is a simple and efficient metric-based classification method for few-shot learning based on the idea of embedding, where the points of each class are clustered around a prototype. The Prototypical Network is modeled by subsampling classes and data points to simulate multiple one-shot tasks; that is, it can be thought of as an N-way K-shot classification problem, where N denotes the class and K denotes the number of samples in each class [27]. In the Prototypical Network, the sample data are divided into a support set and a query set. The support set is used to update the parameters of the prototype, that is, to calculate the prototype. The query set is used to verify the strengths and weaknesses of the algorithm and determine the best algorithm, that is, the optimization prototype.

The basic principle of the Prototypical Network is to map the data to a representation space by training an encoder, and then classify them using a parameter-free decoder. Specifically, each image in the support set is processed through an encoder to extract information, resulting in an embedding vector representation for each image. Then, the embedding of each class in the support set is averaged to obtain a prototype representation for each category, referred to as the class prototype [30]. The same encoder is applied to the query set images, producing an embedding vector for each query image. The similarity between the query image’s embedding and the class prototype is then calculated, which constitutes the parameter-free decoding process. The Softmax function is used to convert the similarity into a probability distribution, and the class with the highest similarity is predicted as the class of the query set.

3.2.2. Data Augmentation

Models trained for FSL are particularly prone to overfitting, and addressing overfitting to improve model generalization potential is a central theme in small-sample modeling. Data augmentation directly exploits important invariants of the data to improve model performance [33,34]; commonly used data augmentation methods are SamplePairing, CutOut, RandomErasing, GridMask, MixUp, and CutMix. Analyzing the distribution characteristics of farmland shelterbelt, this study adopted MixUp and the improved CutMix method for data augmentation, aiming at enriching the farmland scene around the farmland shelterbelt target, improving the representativeness and richness of the sample data of the banded farmland shelterbelt and patch farmland shelterbelt, and obtaining a model with good generalization performance.

MixUp is a data-independent and simple data enhancement method, which randomly selects two vectors of training samples and their corresponding labels and uses linear interpolation to generate a new vector and the corresponding label as the enhanced data, that is, to construct a virtual sample [35]. The specific formula is as follows:

$$\tilde{x}=\lambda x_i+(1-\lambda )x_j$$

(1)

$$\tilde{y}=\lambda y_i+(1-\lambda )y_j$$

(2)

where $x_i$, $x_j$ are raw input vectors, $y_i$, $y_j$ are one-hot label encodings $(x_i,y_i)$, and $(x_j,y_j)$ are two samples randomly selected from the training data and λ ∈ [0, 1].

MixUp can be used both between samples of the same class and can operate between samples of a different class. In this study, MixUp was only used between samples of the same class to ensure that the new samples generated retained the original labels. In addition, the value of λ was set to 0.5.

The CutMix method involves mixing two samples directly at the pixel level of the image, randomly selecting two images, and randomly cropping a rectangular region from each image, exchanging the two cropped regions to synthesize a new image, and the new image labels are computed using the weighted average of the area of the two original image-labeled cropped regions [36]. The CutMix is a randomly cropped rectangular region for exchange synthesis. However, the samples in this study were all labeled at the center point, so we used an improved CutMix method [37]. This method retains the center pixel and sets the cropped rectangular region as the center, as shown in Figure 4. This increases the richness of the samples and helps to improve the stability and generalization of the model. In this study, the size of the rectangular region is set to be (z − 2) × (z − 2), and z is the size of the sample window.

3.2.3. Convolutional Block Attention Module

A convolutional block attention module (CBAM) can be applied to any CNN. The CBAM is able to pay attention to both the channel and spatial dimensions of the deep feature map, and achieves adaptive feature refinement by multiplying the input features with the attention maps obtained by sequential inference in the two dimensions [38]. A CBAM can be seamlessly integrated into other CNNs for efficient feature extraction and end-to-end model training [39], and, due to its lightweight nature, the model inference overhead from module integration is negligible [40,41]. The application of a CBAM to the classification of farmland shelterbelt aims to spatially realize fine and efficient feature extraction, improve the efficiency of feature extraction on images, and enhance the classification performance. A CBAM divides the attentional process into two separate parts, in the order of the channel attention module followed by the spatial attention module. In this study, the CBAM is placed directly in the convolutional block and is used to enhance feature extraction.

3.2.4. Improvement of the Prototypical Network Algorithm

The improved Prototypical Network model constructed in this study consists of three parts, that is, data augmentation, image encoder, and loss function. Among them, the data augmentation part uses MixUp and improved CutMix methods; the order of MixUp and improved CutMix is randomized, and all the data augmentation methods are only applicable to the query set and not to the support set. The image encoder part is composed of multiple convolutional blocks; the number of convolutional blocks (Layer 1… Layer N) depends on the window size of the input data. Each convolutional block consists of two convolutional layers (Conv2d) with 3 × 3 windows, two batch normalization layers (Batch_norm), two nonlinear activation functions (Relu), one maximal pooling layer (Max_pool2d), and one CBAM. The fully connected layer (Flatten), which takes the output of the last convolutional block (1 × 1 × F) as input, is transformed into F eigenvalues by the Flatten. The Euclidean distance from the query set to each class prototype in the projection space is computed, and the probability of belonging to each class is computed as a basis for categorization using the Softmax activation function. Taking a sample with a window size of 5 × 5 as an example, the network structure is shown in

Table 2. The network structure of our image encoder backbone.

Improved Prototypical Network				CBAM Layer
Convolution Block	Layer	Input Shape	Output Shape	Layer	Input Shape	Output Shape
Convolution Block 1	Conv2d	[10, 5, 5]	[64, 5, 5]	Channel attention module
	BatchNorm2d	[64, 5, 5]	[64, 5, 5]	AdaptiveAvgPool2d	[64, 3, 3]	[64, 1, 1]
	ReLU	[64, 5, 5]	[64, 5, 5]	Flatten	[64, 1, 1]	[64]
	Conv2d	[64, 5, 5]	[64, 5, 5]	Linear	[64]	[16]
	BatchNorm2d	[64, 5, 5]	[64, 5, 5]	ReLU	[16]	[16]
	ReLU	[64, 5, 5]	[64, 5, 5]	Linear/avg_fc	[16]	[64]
	MaxPool2d	[64, 5, 5]	[64, 3, 3]	AdaptiveMaxPool2d	[64, 3, 3]	[64, 1, 1]
	CBAM layer	[64, 3, 3]	[64, 3, 3]	Flatten	[64, 1, 1]	[64]
Convolution Block 2	Conv2d	[64, 3, 3]	[64, 3, 3]	Linear	[64]	[16]
	BatchNorm2d	[64, 3, 3]	[64, 3, 3]	ReLU	[16]	[16]
	ReLU	[64, 3, 3]	[64, 3, 3]	Linear/max_fc	[16]	[64]
	Conv2d	[64, 3, 3]	[64, 3, 3]	avg_fc + max_fc	-	[64]
	BatchNorm2d	[64, 3, 3]	[64, 3, 3]	Sigmoid	[64]	[64]
	ReLU	[64, 3, 3]	[64, 3, 3]	Spatial attention module
	MaxPool2d	[64, 3, 3]	[64, 1, 1]	−/avg_out	[64, 3, 3]	[1, 3, 3]
				−/max_out	[64, 3, 3]	[1, 3, 3]
				Concatenate	-	[2, 3, 3]
	CBAM layer	[64, 1, 1]	[64, 1, 1]	Conv2d	[2, 3, 3]	[1, 3, 3]
Flatten	Flatten	[64, 1, 1]	[64]	Sigmoid	[64, 3, 3]	[1, 3, 3]

. The loss function is shown in Equations (3) and (4) [37]:

$$L_{loss}=-log{\displaystyle \frac{exp\left(-d\left(f_{\varnothing }\left(x_i\right),c_{x_i}\right)\right)}{{\sum }_{k=1}^{N_c}exp\left(-d\left(f_{\varnothing }\left(x_i\right),c_k\right)\right)}}$$

(3)

$$c_k={\displaystyle \frac{1}{N_s}}{\sum }_{x_i\in S_k}{f}_{\varnothing }\left(x_i\right)$$

(4)

where $c_k$ represents the prototype of each class; $S_k$ is the support set of the class k; $f_{\varnothing }$ represents the embedding function; $c_{x_i}$ is the prototype to which class sample $x_i$ belongs; $N_c$, $N_s$ represent samples and support samples of each class, respectively; and d is the squared Euclidean distance.

In this study, the improved Prototypical Network was trained using Fujin as the source domain (Experiment A) to evaluate the effectiveness of the constructed Prototypical Network for the classification of farmland shelterbelt. At the same time, to better demonstrate the effectiveness of the used architecture incorporating data augmentation and the CBAM (Experiment B), we also conducted experiments comparing different Prototypical Network architectures (

Table 3. Experiment plan.

Experiment	Description
A	Performance of the improved Prototypical Network in the classification of farmland shelterbelt
B	Comparison of different Prototypical Network architectures	B1: Prototypical Network without data augmentation, without CBAM B2: Prototypical Network integrating CBAM, without data augmentation B3: Prototypical Network integrating data augmentation, without CBAM
C	Classification of farmland shelterbelt using RF and SVM
D	Transfer learning in the target domain	D1: Transfer learning using a fine-tuning method D2: Learning from scratch in the target domain

).

Considering the shape (width, length, etc.) of the banded and patch farmland shelterbelts and their location in relation to the surrounding features, window size starts from 3 × 3 and increases sequentially in 20-m steps to 15 × 15, comparing the effects of different window samples on extraction accuracy and finding the optimal window applicable to the classification of banded and patch farmland shelterbelt.

The sample dataset is based on the remote sensing image used, centered on the screen coordinate of the longitude and latitude of the sample point and clipped with different window sizes from 3 × 3 to 15 × 15. Then, a sample data set consistent with the number of sample points was obtained in different window sizes.

The sample data of Fujin were divided into training samples and test samples in a ratio of 3:1. The initial learning rate was set to 5 × 10⁻⁴, and at every 2000 training iterations the learning rate was halved.

3.3. Random Forest and Support Vector Machine

Random forest (RF) is a commonly used non-parametric machine learning algorithm. When RF deals with classification problems, each decision tree (DT) in RF is a classification tree, and for each test sample, each DT in the forest will give the final category; then, the output categories of each DT within the forest are considered together to decide the category of the sample by voting.

A support vector machine (SVM) is a supervised learning algorithm commonly used for classification and regression, with unique advantages in handling few-shot sample, non-linear, and high-dimensional pattern recognition. The core idea is to find an optimal hyperplane in the feature space that separates samples of different classes. An SVM uses a kernel function to map input data into a high-dimensional feature space, making data that were originally linearly inseparable become linearly separable in the new space. Commonly used kernel functions include the polynomial function, neural network kernel function, Gaussian radial basis function, and sigmoid function.

In this study, the same training and validation sets were used in the study area of Fujin in the source domain, and the RF algorithm and SVM algorithm were used for the classification of banded and patch farmland shelterbelt (Experiment C). The two parameters for the RF algorithm, the number of variables preselected at the tree node (mtry) and the number of trees in a forest (ntree), were set to 2 and 2000, respectively.

For the SVM, we used the Gaussian radial basis function and set gamma in the kernel function and the penalty parameter to 0.025 and 100, respectively.

3.4. Transfer Learning

There are two ways of model training, learning from scratch and transfer learning. Transfer learning, as the name suggests, is transferring the parameters of an already learned and trained model to a new model to help train the new model [42]. The benefit is that instead of starting over to design and train a completely new network, the training is based on an already trained network model, on which parameter and knowledge migration is performed, requiring only a small amount of computational resource overhead and training time to achieve support for new tasks [43,44]. The core methods of transfer learning mainly include feature reuse, fine-tuning, and domain adaptation. Fine-tuning refers to adjusting some parameters of a pre-trained model to adapt it to a new task. This is the transfer learning of the model, typically faster and more accurate.

In this study, a fine-tuning transfer learning approach was used to migrate the model for classification of farmland shelterbelt obtained from training in Fujin City, which was the source domain, to the target domain including Hailun, Yi’an, Youyi, and Fuyu (Experiment D). The specific implementation method was as follows:

(1) Construct the sample set of the target domain. Based on the optimal window size s obtained from the training in the source domain (Fujin), the images of the four target regions were cropped according to the s × s window to construct the sample set, and the sample set was divided into a training set and a validation set in the ratio of 1:2.

(2) Model fine-tuning. The optimal model obtained based on source domain training was fine-tuned using the training set of the target region. All parameters of Convolution Block 1 were frozen and the parameters of Convolution Block 2 were fine-tuned.

(3) Accuracy evaluation. The accuracy of the fine-tuned model was evaluated using the validation set of the target domain, and the distribution map of the farmland shelterbelt in the target domain was obtained.

In addition, in order to evaluate the effect of transfer learning, this study conducted model training by learning from scratch in each of the four target regions, and evaluated the classification accuracies of banded and patch farmland shelterbelt in each target region.

3.5. Validation

In order to evaluate the accuracy of the model in classifying the banded and patch farmland shelterbelt, OA, kappa coefficient (Kappa), producer accuracy (PA), and average accuracy (AA) were utilized as the evaluation indexes.

$$OA={\displaystyle \frac{\sum _{i=1}^n{N}_{ii}}{M}}$$

(5)

$$Kappa={\displaystyle \frac{M\sum _{i=1}^n{N}_{ii}-\sum _{i=1}^n{N}_{i+}\times \sum _{i=1}^n{N}_{+i}}{M^2-\sum _{i=1}^n{N}_{i+}\times \sum _{i=1}^n{N}_{+i}}}$$

(6)

$$PA={\displaystyle \frac{N_{ii}}{N_{i+}}}$$

(7)

$$AA={\displaystyle \frac{\sum _{i=1}^{N}PA}{n}}$$

(8)

where M and n represent the total number of samples and categories in the dataset, respectively. For the i-th category, N_ii, N_+i, and N_i+ represent the number of correctly classified samples, reference samples, and total classified samples (including correctly classified samples and misclassified samples), respectively.

4. Results

4.1. Performance of the Farmland Shelterbelt Extraction

The results for farmland shelterbelt using the improved Prototypical Network are shown in

Table 4. Classification results of farmland shelterbelt under different sample windows in Fujin (Experiment A).

Evaluation Index		3 × 3	5 × 5	7 × 7	9 × 9	11 × 11	13 × 13	15 × 15
OA		90.37	93.33	93.33	92.59	93.33	92.59	92.59
Kappa		0.8547	0.8985	0.8991	0.8876	0.8987	0.8871	0.8871
AA		89.86	93.05	92.65	91.98	92.84	92.49	92.49
PA	Other	94.23	96.08	98	96.08	96.08	97.96	97.96
	Banded farmland shelterbelt	95.35	92.16	95.74	93.75	93.88	88.89	88.89
	Patch farmland shelterbelt	80	90.91	84.21	86.11	88.57	90.62	90.62

. The classification accuracy is good regardless of the window size, where OA and kappa are greater than 90% and 0.85, respectively. In the classification results, the accuracy of banded farmland shelterbelt, patch farmland shelterbelt, and others varies with the change of window size. The overall trend of other and patch farmland shelterbelt is that with the increase of window size, the precision (PA) first increases, then decreases, and then increases to a stable state. The precision of banded and patch farmland shelterbelt first increases and then decreases, and then tends to be stable with the increase of the window size. Among them, the classification accuracy of others is the highest, and for farmland shelterbelt, with 3 × 3 to 11 × 11 window size, the accuracy of banded farmland shelterbelt is higher than that of patch farmland shelterbelt. The result is reversed with windows size greater than 11 × 11 (13 × 13, 15 × 15). Considering that the objective of this study is to achieve the classification of farmland shelterbelt, that is, the classification of different forms of farmland shelterbelt, we take the accuracy of banded and patch farmland shelterbelt as the main criterion, and also take into account the accuracy of other land types; therefore, 5 × 5 is determined as the best sample window size for farmland shelterbelt classification, and the model obtained under this sample window size is the optimal classification model, with banded and patch farmland shelterbelt having an accuracy of 92.16% and 90.91%, respectively.

Table 4,

Table 5. Classification results of farmland shelterbelt using the Prototypical Network without data augmentation and CBAM in Fujin (Experiment B₁).

Evaluation Index		3 × 3	5 × 5	7 × 7	9 × 9	11 × 11	13 × 13	15 × 15
OA		88.89	91.11	90.37	87.41	87.41	88.89	87.41
Kappa		0.8321	0.8647	0.8536	0.8073	0.8078	0.8302	0.8071
AA		88.32	90.72	89.80	87.23	86.93	88.65	87.22
PA	Other	92.30	92.45	94.12	90.57	92.31	94.12	92.30
	Banded farmland shelterbelt	93.18	91.84	90	84.91	84.62	85.16	83.64
	Patch farmland shelterbelt	79.49	87.88	85.29	86.21	83.87	86.67	85.71

,

Table 6. Classification results of farmland shelterbelt using the Prototypical Network integrating CBAM without data augmentation in Fujin (Experiment B₂).

Evaluation Index		3 × 3	5 × 5	7 × 7	9 × 9	11 × 11	13 × 13	15 × 15
OA		88.15	89.63	88.89	89.63	89.63	88.15	89.62
Kappa		0.8204	0.8421	0.8310	0.8419	0.8417	0.8193	0.8424
AA		87.58	89.62	88.56	89.17	89.43	87.65	89.3
PA	Other	90.74	87.50	89.09	97.96	96	92.31	95.92
	Banded farmland shelterbelt	90.91	93.48	91.30	85.19	85.19	86.27	86.27
	Patch farmland shelterbelt	81.08	87.88	85.29	89.17	87.10	84.38	85.71

and

Table 7. Classification results of farmland shelterbelt using the Prototypical Network integrating data augmentation, without CBAM in Fujin (Experiment B₃).

Evaluation Index		3 × 3	5 × 5	7 × 7	9 × 9	11 × 11	13 × 13	15 × 15
OA		89.63	92.59	93.33	91.11	92.59	91.18	89.71
Kappa		0.8427	0.8876	0.8987	0.8652	0.8874	0.8662	0.8441
AA		89.13	91.98	92.84	90.36	92.10	90.70	89.28
PA	Other	90.74	96.08	96.08	96.08	96.08	96.00	95.83
	Banded farmland shelterbelt	93.33	93.75	93.88	91.67	92	86.11	88.24
	Patch farmland shelterbelt	83.33	86.11	88.57	83.33	88.24	90.00	83.78

show the classification results of farmland shelterbelt using different Prototypical Network architectures. The improved Prototypical Network achieves the highest accuracy in extracting banded and patch farmland shelterbelt, while the Prototypical Network without the integration of data augmentation and CBAM performs the worst. On the whole, whether integrating data augmentation alone or the CBAM alone, the classification accuracies for banded and patch farmland shelterbelt are all significantly improved, especially the integration of data augmentation. All of these demonstrate the effectiveness of the improved Prototypical Network in the classification of farmland shelterbelt.

Based on the obtained optimal classification model, the farmland shelterbelt of the whole of Fujin was classified, and the distribution map is shown in Figure 5. The classification results of banded and patch farmland shelterbelt are basically consistent with the actual situation, indicating that banded and patch farmland shelterbelt are effectively distinguished. As can be seen by observing the detailed magnified map, the method used in this study is able to identify banded farmland shelterbelt with different lengths such as a single tree, several trees, and a row of trees, as well as patch farmland shelterbelt with different area sizes.

4.2. Extraction Results for RF and SVM

We used the same sample training set and testing set to classify banded and patch farmland shelterbelt using RF and the SVM in the source domain, Fujin study area, as shown in the

Table 8. Classification results of farmland shelterbelt using RF and SVM in Fujin (Experiment C).

Region	Algorithm	OA	Kappa	AA	PA
					Other	Banded Farmland Shelterbelt	Patch Farmland Shelterbelt
Fujin	RF	80.74	0.7057	80.92	79.03	83.72	80.00
Fujin	SVM	82.22	0.7320	81.79	89.58	82.61	73.17

. The results show that not only the overall classification accuracies of the two algorithms are lower (OA, Kappa), but also the accuracies of banded farmland shelterbelt, patch farmland shelterbelt, and others are all low. For RF, the accuracies for banded and patch farmland shelterbelt were 83.72% and 80%, respectively; for the SVM, the accuracies for banded and patch farmland shelterbelt were 82.61% and 73.17%, respectively. These figures indicate that banded and patch farmland shelterbelt cannot be effectively distinguished and classified using the RF and SVM algorithms.

4.3. Effectiveness of Transfer Learning in Different Regions

Based on the analysis that the optimal sample window size is 5 × 5, the sample datasets for each of the four target regions were constructed. On this basis, the optimal classification model obtained from the source domain training was utilized for fine-tuning transfer learning, as well as training from scratch directly; the results of the two methods for classifying the farmland shelterbelt are shown in

Table 9. Classification results of farmland shelterbelt in the target domain.

Method	Region	OA	Kappa	AA	PA
					Other	Banded Farmland Shelterbelt	Patch Farmland Shelterbelt
Transfer learning	Fuyu	97.20	0.9569	97.29	98.20	96.32	97.33
	Youyi	98.75	0.9810	98.80	100	96.39	100
	Hailun	98.91	0.9836	98.90	99.01	98.89	98.81
	Yi’an	96.27	0.9433	95.74	98.95	98.73	89.55
Learning from scratch	Fuyu	96.58	0.9474	96.50	99.07	95.62	94.81
	Youyi	97.50	0.9621	97.18	100	97.44	94.12
	Hailun	97.45	0.9617	97.34	100	96.63	95.40
	Yi’an	95.02	0.9246	94.45	98.92	98.72	85.71

.

Using a small number of samples to fine-tune the obtained model in the four target regions for the classification of farmland shelterbelt, the overall classification accuracy remained high, with high accuracy for both banded and patch farmland shelterbelt. The accuracy of banded and patch farmland shelterbelt using the transfer learning method was significantly higher than the accuracy of learning from scratch in the target area, except for the banded farmland shelterbelt in the Youyi study area, which indicates that the transfer learning method through fine-tuning can effectively identify the banded and patch farmland shelterbelt, and that the transfer learning method through fine-tuning can learn more, which can help to improve the classification of farmland shelterbelt. The classification accuracy of the four target study areas was over 96% for banded farmland shelterbelt and over 89% for patch farmland shelterbelt. This indicates that the improved Prototypical Network is a practical tool for large-scale monitoring of farmland shelterbelt, enabling rapid, high-accuracy monitoring at the regional scale and providing data support for the management and configuration of farmland shelterbelt.

The distribution maps of the four target areas are shown in Figure 6, Figure 7, Figure 8 and Figure 9. By comparing the classification results with the Sentinel-2 image data, the extracted banded farmland shelterbelt and patch farmland shelterbelt are basically consistent with the actual distribution, which further illustrates that the transfer capability of our proposed method is strong, and that this method is fast and suitable for the production of fine products related to large-scale farmland shelterbelt.

5. Discussion

The improved Prototypical Network constructed in this study belongs to FSL, which relies on only a small number of samples to complete the training of a model, and is applicable to the situation where it is difficult to obtain a large number of farmland shelterbelt samples in a regional-scale farmland scenario; it reduces the workload of sample data collection, and helps the classification model to be applied to practical production work. In the task of classification of farmland shelterbelt, due to the complexity of the farmland scene and the large differences in distribution patterns of different forms of farmland shelterbelt, the representativeness and richness of the small amount of sample data obtained may not be good enough, which leads to overfitting of the classification model. In this study, improvements were made in terms of data augmentation and feature enhancement to alleviate the overfitting problem of the model, especially the inclusion of data augmentation methods. The MixUp and improved CutMix methods used in this study, compared with the commonly used methods such as rotation and flipping, fully consider the distribution characteristics of banded and patch farmland shelterbelt, which helps to increase the scenarios of farmland shelterbelt samples and enrich the representativeness and diversity of the sample data in order to enhance the generalizability of the model.

For the improved Prototypical Network constructed, the size of the input sample window is variable, and the number of layers of the network increases or decreases with the size of the input sample window. The sample window not only considers the spectral information of the sample point itself, but also fully considers the spatial information of the sample point and the surrounding features. The size of the sample window has a great influence on classification accuracy [30]. Different features have their own unique spatial distribution characteristics, and an appropriate sample window size can effectively reflect the features and improve classification accuracy. In this study, by comparing and analyzing the extraction accuracy of different sample window sizes, it was determined that the 5 × 5 sample window is the optimal window size for extracting the banded farmland shelterbelt and patch farmland shelterbelt, which also indicates that the 5 × 5 window range, corresponding to a ground range of 50 m × 50 m, can effectively reflect the spatial distribution information of banded and patch farmland shelterbelt and the surrounding features.

Other researchers have also conducted studies on the classification of farmland shelterbelt. Deng et al. [45] utilized the decision tree method to extract the farmland shelterbelt through spectral curve analysis, principal component transform analysis, and shape index analysis based on the TM image in Dehui city of Jilin Province, and the results showed that the accuracy was 85.52%. Wang et al. [46] used the multispectral data of a UAV to extract farmland shelterbelt based on the Deeplabv3+ semantic segmentation model in the 3rd Division of Xinjiang Production and Construction Crops; the MIoU of farmland shelterbelt derived from the Deeplabv3+ was 85.54%. Lei et al. [47] used four quarters of Landsat 8 OLI time series data as the data source and mixed vegetation phonological characteristics, water, and vegetation index together, gradually extracting and masking other land features and combining the idea of hierarchical classification in the study area, and ultimately achieved the extraction of farmland shelterbelt; the accuracy of farmland shelter forest reached 85.93%. The performance of banded and patch farmland shelterbelt extracted by our proposed method is obviously better. This shows the superiority of our method in distinguishing and extracting the banded and patch farmland shelterbelt, and realizes for the first time the classification of farmland shelterbelt with end-to-end automation, which is helpful for the subsequent detailed analysis of the distribution, morphology, and other factors of the farmland shelterbelt on the impact of yield at a large scale. However, there are still some limitations in our method, and it was found from the obtained distribution map of the farmland shelterbelt that the edge of the patch farmland shelterbelt was more likely to be incorrectly classified into the banded farmland shelterbelt because its distribution characteristics were similar to those of the strip shelterbelt. Therefore, further optimization should be carried out in relation to this problem in subsequent research.

Based on the optimal model obtained by training in the source domain, this study carried out transfer learning in the target domain using a small number of samples for fine-tuning, and all obtained very good extraction results, with the accuracies of banded and patch farmland shelterbelt above 95% and 89%, respectively, and basically higher than that of the model learned from scratch directly in the target domain. This indicates that the improved Prototypical Network model that we constructed has strong transferability, and also further indicates that the improved Prototypical Network is almost free of the overfitting phenomenon. The workload of sample data collection as well as the training time of the model can be greatly reduced by transfer learning, which is beneficial to the production of farmland shelterbelt maps at a large scale, such as at provincial scale and above. The high-quality production of farmland shelterbelt maps is the basis for exploring the role of banded and patch farmland shelterbelt in regional crop yield, and the timely understanding of the distribution of farmland shelterbelt rationally adjusts the configuration structure of banded and patch farmland shelterbelt, improves the protective benefits of farmland shelterbelt, supports the continuous improvement of the comprehensive production capacity of food, and safeguards national food security.

6. Conclusions

In this work, we propose an improved Prototypical Network model incorporating data augmentation and a CBAM for the classification of farmland shelterbelt based on Sentinel-2 data. The improved Prototypical Network was utilized to classify banded and patch farmland shelterbelt under different sample window sizes, and the optimal sample window size of 5 × 5 was determined, under which the banded and patch farmland shelterbelt were classified with the highest accuracy of 92.16% and 90.91%, respectively. This indicates that the few-shot method used can effectively realize the classification of farmland shelterbelt and solve the problem of sample data collection limitation. On this basis, the best classification model was transferred to the target domain, and the fine-tuning method was used to optimize the model; the results showed that the transfer learning method of classifying the farmland shelterbelt was all very effective, and the accuracies of the banded and patch farmland shelterbelt in the four target study areas were above 95% and 89%, which were significantly higher than the results obtained by training from scratch directly in the target domain. This suggests that the proposed model has strong transferability, alleviates overfitting problems, and has the potential to produce fine products on farmland shelterbelt at a large scale. Overall, this study proposes a feasible method for classification of farmland shelterbelt, which can increase the richness of sample data and alleviate the overfitting problem by utilizing data augmentation and a CBAM. In the future, based on the large-scale banded and patch farmland shelterbelt products obtained by this method, our work will focus on exploring the specific influence of banded and patch farmland shelterbelt on crop yields, in order to further adjust the configuration structure of the farmland shelterbelt and to improve the protective effect.