Feature optimization combined with UPerNet-Twins model for eucalyptus extraction from Sentinel-2A image

To improve the extraction accuracy of eucalyptus from Sentinel-2A image, two key factors of feature construction and extraction model are considered. The original band spectrum, custom vegetation index, red edge spectral index, and texture features are obtained from the image. The Relief F-PSO-SVM model is used to screen out the best feature subset. A UPerNet-Twins combination model is used to realize the high-precision extraction of eucalyptus for the study area. The experiments show that the original spectrum plays a significant role in the extraction of eucalyptus. In addition, the texture features, vegetation index, and red edge spectral index are helpful to the extraction of eucalyptus, which are increased by about 5.84%, 3.81%, and 3.00%, respectively, compared with the IOU of the original spectrum. Moreover, the IOU of the optimal feature subset obtained by feature selection is increased by about 2.22% compared with the original feature set. The UPerNet-Twins combination model has the highest extraction accuracy of eucalyptus with an IOU of 0.8496 compared with DeeplabV3+, PspNet-UNet, FCN, UperNet, and UperNet-Vision Transformers models.


Introduction
Eucalyptus spp.(scientific name: Eucalyptus spp.) is the collective name of the genus Eucalyptus, the genus Eucalyptus and the genus Myrtaceae.It can be used in pulp and paper, wood-based panels, and construction industries [1] .Through the automatic or semi-automatic identification or extraction of remote sensing images, the acquisition and monitoring of key information such as eucalyptus distribution, area, and health status can be achieved.The extraction range, accuracy, timeliness, and human resource consumption are greatly superior to manual extraction [2] .Based on the GF-6 image, Wang et al.In [3] constructed a vegetation index and red edge index and used a decision tree as an extraction model to extract eucalyptus forest information in Liuzhou area.In [4], the GF-2 image was selected to extract eucalyptus by spectral and texture features and vegetation index.These studies have shown that it is not very effective to extract eucalyptus only by relying on a single original spectral feature.It is necessary to combine vegetation index, red edge index, texture, and other features on the basis of the original spectrum.Due to the 'dimension disaster', it is necessary to optimize the features to obtain features that are more helpful for extraction [5] .In recent years, feature optimization has promoted the improvement of remote sensing image classification and information extraction results [6] .

MCAI-2023
Journal of Physics: Conference Series 2724 (2024) 012023 Hao et al.In [7] used the Relief F feature selection method and decision tree classification to extract wetland information and obtained 86.9% classification accuracy.The Relief F-PSO-SVM feature selection method was proposed in [8], and the mixed features were extracted based on this method, which achieved the highest accuracy in land use classification.
In addition to feature selection, the extraction model is another key factor affecting remote sensing image classification and information extraction.In recent years, UPerNet [9] and Twins [10] models have been proposed in deep learning models.This paper combines the advantages of the UPerNet, and Twins models, recombines the UPerNet-Twins model, and considers the importance of feature construction to the performance of the model.The Relief F-PSO-SVM method is used to select the best feature subset, and the extraction of eucalyptus from Sentinel-2A image in the study area is discussed.

Overview of the study area
Due to the suitable geographical location, natural conditions, and climatic conditions, eucalyptus in Guangxi has formed the industrialization and scale of plantations.According to the data from 2020, the planting area of eucalyptus is more than 30 million acres, accounting for more than half of the total area of eucalyptus in China.The selected study area is located in Nanning, Guangxi, with a longitude range of 107°47′53′′ to 107°58′41′′ and a latitude range of 23°6′8′′ to 23°55′3′′, with a total area of about 2.5 square kilometers (Figure 1).After resampling of Sentinel-2A image data of the study area, the spectral resolution is set to 10 m, and only 10 bands (B2, B3, B4, B5, B6, B7, B8, B8a, B11, B12) are retained.

Multidimensional dataset construction
Because the single original spectral characteristics of other tall trees, such as pine and mahogany, are similar to the characteristics of eucalyptus, the original spectrum, vegetation index, red edge index, and extracted texture features are used to improve the extraction accuracy of eucalyptus.Nine vegetation indexes (NDVI, RVI, DVI, EVI, GSAVI, IPVI, NormR1, NormR2, NDeI) and six red edge indexes (NDVIre1, NDVIre2, NDVIre3, NDre1, NDre2, CIre) are obtained by using the original band weighting operation.Gray-Level Co-occurrence Matrix (GLCM) is used to extract texture features.After principal component analysis, the first two images were selected for texture feature extraction, and a total of 16 texture features were extracted.

Relief F-PSO-SVM feature selection method
Feature selection refers to the process of selecting the best feature subset in the feature set.Feature selection methods are generally divided into three types: Filter, Wrapper, and Embedded.The Relief F algorithm used in this paper belongs to the filtering type, and the PSO-SVM combination model belongs to the embedded type.The combination of the Filter-Wrapper algorithm can not only reduce the computational complexity and be very friendly to high-dimensional data sets but also take into account the interaction between features.This can result in the highest superiority of the selected feature subset.

Eucalyptus extraction feature combination scheme
In terms of extracting the feature set of eucalyptus, six feature combination schemes were designed based on the original feature set (Table 1).In order to more accurately explore the role of the red edge index, the red edge band (B5, B6, B7) is not included in the original band spectral feature scheme but is classified into the red edge index.

Best feature subset
Firstly, the Relief F algorithm was used to screen the 41 features (10 original spectral features, 9 vegetation indices, 6 red edge indices, and 16 texture features).The importance of the features was ranked according to their weight (Figure 3).The greater the weight is, the greater the correlation with the eucalyptus category is.Initially, 23 features with larger weights are retained, and the initial feature subset is obtained.Then, PSO-SVM is used to re-screen the initial feature subset, and the particle swarm achieves convergence and finally obtains the best feature subset with a feature number of 13 (Table 2).   1.The results are shown in Table 3. Table 3 shows that although scheme 1 has the lowest interaction using the original spectral features alone, it can still achieve a certain extraction accuracy.Hence, there will still be five original spectral features retained in the best feature subset scheme 6.The vegetation index, red edge index, and texture feature were added separately in schemes 2 ~ 4. The results showed that the extraction effect of scheme 4 with the texture feature was the best, and the interaction ratio reached 84.18%.Although scheme 5 contains all the features, it is not simple that the more features can show the better extraction effect.Scheme 6 has only 13 features, which is far less than scheme 5.
However, its extraction accuracy is the best among the 6 schemes.Displays the original images and labels of the study area and the prediction maps of the six schemes (Figure 4).By comparing different contrast schemes with the original image labels, it can be seen that using the best feature subset has the best effect.It can be seen from the prediction maps of schemes 1 to 3 that in areas where eucalyptus trees are sparsely distributed, there are a large number of missing and wrong mentions.The main reason is that the resolution is limited and there is a mixture of pixels in some small areas.Schemes 4 to 5 greatly reduced the phenomenon of missing extractions, but the phenomenon of unclear and ambiguous eucalyptus extraction still appeared in many places.Scheme 6 has been significantly improved, and it is very close to the original image label in all aspects.4), it can be seen that the UPerNet-Twins model is the best, followed by the PSPNet-UNet model, the UperNet model, the DeepLabV3 + model, the UperNet-Vision Transformers model, and the FCN model.Under the same optimal feature subset, the maximum difference in extraction accuracy IOU is about 15%, and the average IOU difference is about 8%.(1) The Relief F-PSO-SVM model is suitable for feature selection.13 feature subsets are selected from the original 41 features, with the highest extraction accuracy and effect can be obtained by using relatively few features.This shows the practicability of feature selection.

MCAI-2023
(2) The visual attention model shows superiority over the classical convolutional neural network in various visual tasks, but it still needs to be optimized and improved in efficiency.Compared with DeeplabV3+, PspNet-UNet, FCN, UperNet, and UperNet-Vision Transformers models, the IOU of the new combined UPerNet-Twins model is increased by at least about 3.8 % for eucalyptus extraction based on the optimal features.
Although eucalyptus extraction has achieved good accuracy in this paper, further research is still needed on multiple numbers of mixed species, a better feature optimization method, and a finer neural network model.

Figure 1 .
Figure 1.Location of Study Area.2.5.UPerNet-Twins combination model In this paper, UPerNet is combined with the Twins model to form the UPerNet-Twins model (Figure 2), which is mainly composed of a feature pyramid and pyramid pooling model.The left feature extraction module is replaced by the Twins model.The Twins model has two kinds of visual converter architectures.In this paper, the Twins-PCPVT model is used as the image semantic segmentation model.The process of image classification by the UPerNet-Twins model through feature extraction is as follows.(1) A remote sensing image with the size of H × W × C is input, where H, W, and C represent the length, width, and number of input channels of the input image, respectively; (2) Through the feature extraction of Twins module, four feature maps of different sizes are obtained respectively.The minimum feature map (a4) fuses the global average pooling layer of four feature maps of different sizes through the pyramid pooling model to obtain the feature map b4; (3) The feature map b4 is upsampled three times and fused with the corresponding size of the Twins model to obtain three feature maps (b3, b2, b1); (4) The feature maps b1, b2, b3, and b4 are magnified to the original feature size and fused by bilinear difference method; (5) The 3 × 3 convolution kernel is used to classify the pixels of the feature fusion image, and the final classification result is obtained by restoring the size to H × W by the bilinear difference method.

Figure 4 .
Figure 4. Prediction Map of Eucalyptus Extraction with Different Feature Combination Schemes.3.2.2.Comparison of different model results under the best feature subset.By comparing the extraction accuracy of eucalyptus in different models (Table4), it can be seen that the UPerNet-Twins model is the best, followed by the PSPNet-UNet model, the UperNet model, the DeepLabV3 + model, the UperNet-Vision Transformers model, and the FCN model.Under the same optimal feature subset, the maximum difference in extraction accuracy IOU is about 15%, and the average IOU difference is about 8%.
mean-1, mean-2, homo-23.2.Comparison of Eucalyptus extraction results3.2.1.Comparison of different feature combination results under the UPerNet-Twins model.Based onthe UPerNet-Twins model, eucalyptus was extracted according to the six feature combination schemes in Table

Table 3 .
Comparison of Extraction Accuracy of Eucalyptus with Different Feature Combinations.

Table 4 .
Comparison of Extraction Accuracy of Eucalyptus from Different Models.PSO-SVM model was used to optimize the features of the feature set obtained from the Sentinel-2A image of the study area, and the best feature subset was obtained.Eucalyptus extraction and comparison experiments were carried out under the combined UPerNet-Twins model in this paper.The main conclusions are as follows.