Research on a Classification Method of Crop Planting Structures Based on Indexes and an Improved U-Net Network

doi:10.21203/rs.3.rs-951543/v1

Download PDF

Research Article

Research on a Classification Method of Crop Planting Structures Based on Indexes and an Improved U-Net Network

https://doi.org/10.21203/rs.3.rs-951543/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

This paper proposes a method for extracting crop planting structures based on indexes, image fusions and an improved U-Net deep semantic segmentation network; this method can automatically and accurately extract multiple types of information regarding crop planting structures. Taking the Landsat 8 commercial multispectral satellite dataset as an example, three indexes are used to highlight the characteristics of vegetation, water and soil. The GS method is used to combine the indexes to fuse panchromatic bands and obtain 15-m high-resolution fused images, and the images are sliced to obtain training sets and test sets of images. When using different combinations of index-fused images and the improved U-Net network to classify the test set data, the overall accuracy rate reaches more than 92.4%, and the accuracy rate of the crop planting structure (wheat, cotton, and woodland) reaches up to 94.2%. The results showed that the participation of AWEI in the image synthesis process can significantly improve the classification effect of the crop planting structure by more than 1.2%, and the kappa coefficient of the NDVI-AWEI method reaches 0.886; thus, this method is the most stable and suitable for actual production. This method can assist relevant enterprises in adjusting and optimizing their regional planting structures, in the intelligent allocation of irrigation water and in other related decisions.

Scientific Communication

Information Retrieval and Management

Remote Sensing

Indices

G-S fusion

U-Net Network

Crop Planting Structure

The classification of crop planting structures based on remote sensing has the capabilities of large-area detections and rapid imaging. This type of classification is a high-tech method used to quickly determine the crop planting structure. In recent years, high-resolution remote sensing satellites have developed rapidly^[1]–[4]. High-resolution remote sensing images can be used as data sources to more accurately detect crop planting structures^[5]–[8]]. Traditional methods such as decision tree classifications and NDVI curves can extract only shallow features and cannot integrate multiple source images, making it impossible to effectively improve the accuracy when obtaining planting structure information^[9]–[11].

With the development of machine learning, neural networks (NNSs) ^[12], support vector machines (SVMs) ^[13]–[14] and other algorithms have been applied to classify high-resolution remote sensing images. Nevertheless, these are all shallow learning algorithms and cannot express complex functions well^[15]–[17]. Therefore, these models cannot adapt to the semantic segmentation problems with large sample sizes and high complexities^[18]. Convolutional neural networks (CNNs) have the advantages of automatic feature extraction and automatic classification, and they show significant benefits for remote sensing^[19]–[21]. As critical network models in semantic segmentation research, fully convolutional neural networks (FCNs) are widely used for high-resolution remote sensing images^[22]–[27]. Many scholars have transformed and developed FCNs, creating a series of convolution-based segmentation models, including Segnet^[28], unet^[29], deeplab^[30], ultra-scale fcn^[31], and reseg^[32], and U-Net, among which U-Net can achieve the highest number of samples with predicted results^[33]–[37].

The subtle features of some planting structures are only visible on high-resolution panchromatic images (PANs) of the corresponding spectral images. These features are challenging to recognize using low-spatial-resolution multispectral images (MSIs). When using traditional methods to classify spectral images, it is challenging to make full use of the high spatial resolutions of PANs and the rich spectral information of MSIs to accurately obtain crop planting structures. At present, the commonly used image fusion methods are HIS (hue, luminance, and saturation) transformation, GS transformation (Gram-Schmidt), principal component analysis (PCA), the NND fusion algorithm (nearest-neighbor diffusion), the UNB fusion algorithm (UNB), wavelet transform (WT), etc. The image spectral fidelity after G-S fusion is high high^[38]–[42], so this paper uses the G-S fusion method for the image fusions.

The above research mainly uses a single-spatial-resolution data source in the selection of training images. Previous studies have rarely considered the synthesis index or used image fusion algorithms to improve the spectral complexity and spatial resolutions of pictures. In terms of network selection, the depth and weight file size are often selected to be more in-depth. A more extensive semantic segmentation network is not conducive to improved results.

This article takes the Xiaokaihe Irrigation District in Binzhou City, Shandong Province as an example and uses Landsat 8 remote sensing image data acquired on May 18, 2018, to carry out a study on an index-based and improved U-Net network classification method.

A field survey was conducted in the crop planting region of the study area in mid-May, and the classification data of the planting structure were manually produced according to the survey results to ensure their accuracy. The main crop types in the study area in mid-May included three types of winter wheat at the filling stage, cotton at the seedling stage and jujube trees. According to the actual needs of the irrigated areas, the water and surrounding villages and towns were classified at the same time. These areas consisted of five significant categories of cotton and wheat. Among them, road wasteland and urban construction land was classified as city land; jujube trees and forestland were classified as forestland; and salt ponds and pools were classified as water. The survey results included the time and location of each sample and the type of land cover.

In this paper, we use the real label data to make a verification set of 128×128 pixel slices of the training images. As training data, 80% of the verification set photographs and one-to-one correspondence are randomly selected to make the training set for training the network. The other 20%, constituting the test set data, are used to evaluate the results.

To insert images in Word, position the cursor at the insertion point and either use Insert | Picture | From File or copy the image to the Windows clipboard and then Edit | Paste Special | Picture (with “float over text” unchecked).

IEEE will do the final formatting of your paper. If your paper is intended for a conference, please observe the conference page limits.

A. Index selection

It is difficult to accurately describe and distinguish the spectral characteristics of many types of different features based on original remote sensing image data. This paper selects NDVI (normalized difference vegetation index), AWEI (automatic water extraction index), and SAVI (soil-regulated vegetation index) to synthesize 1-7-band remote sensing images based on different index characteristics to increase the spectral complexity. Among them, NDVI and SAVI can effectively distinguish the spectral differences among vegetation types and the difference between plants and soil. AWEI can efficiently identify water. The selection of these three indexes has positive significance for the classification of planting structures in irrigated areas.

1) Normalized difference vegetation index

The normalized difference vegetation index (NDVI) is based on the characteristics of chlorophyll in vegetation. Chlorophyll has a lower reflectance in the visible light band, while its reflectances in the red band and green band are relatively high. The reflectance gradually decreases in the near-infrared band to the mid-infrared group. The formula used to construct the most representative normalized difference vegetation index is as follows:

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

3.1

where NIR is the near-infrared band reflectance and RED is the red band reflectance.

2) NND Image Fusion Method

The Automated Water Extraction Index (AWEI) improves factors such as low classification accuracies in water extractions and relatively small threshold selections^[43]. The expression is as follows:

$$\text{A}\text{W}\text{E}\text{I}\text{n}\text{s}\text{h}=\frac{4\text{*}(\text{G}\text{R}\text{E}\text{E}\text{N}-\text{S}\text{W}\text{I}\text{R}1)}{0.25\text{*}\text{N}\text{I}\text{R}+0.75\text{*}\text{S}\text{W}\text{I}\text{R}2}$$

3.2

$$\text{A}\text{W}\text{E}\text{I}\text{s}\text{h}=\text{B}\text{L}\text{U}\text{E}+2.5\text{*}\text{G}\text{R}\text{E}\text{E}\text{N}-1.5\text{*}(\text{N}\text{I}\text{R}+\text{S}\text{W}\text{I}\text{R}1)-0.25\text{*}\text{S}\text{W}\text{I}\text{R}2$$

3.3

where BLUE is the blue band, SWIR1 and SWIR2 are the shortwave infrared bands, and AWEI_sh is more suitable for scenes with shadows than AWEInsh. In this paper, to better distinguish the boundary pixels between the water and the ground features, AWEI_sh is selected as the synthetic element and is hereafter referred to as AWEI.

3) Soil regulation vegetation index

The soil adjusted vegetation index (SAVI) introduces soil adjustment factors to improve the identification ability of the normalized index under different soil backgrounds. The formula is as follows:

$$\text{S}\text{A}\text{V}\text{I}=\frac{(\text{N}\text{I}\text{R}-\text{R}\text{E}\text{D})\text{*}(\text{L}+1)}{\text{N}\text{I}\text{R}+\text{R}\text{E}\text{D}+\text{L}}$$

3.4

where L is the vegetation cover corresponding to the soil adjustment factor, 0 indicates no vegetation, and 1 indicates complete vegetation coverage. According to the study area's situation, this paper selects the best factor value in the range of 0.3-0.8 with L=0.5.

B. Image Fusion

A single remote sensing image usually cannot have both a high spectral resolution and a high spatial resolution. Remote sensing image fusion technology is an effective method used to solve this problem.

G-S fusion is not limited by band, and can better maintain texture information, especially the high fidelity of spectral features. The multispectral low spatial resolution images are used to simulate the high-resolution band images. As the first part of GS transform, the simulated high-resolution band images and low-resolution band images are transformed. By adjusting the statistical value of the high-resolution band images, the first component GS1 of GS transform is matched to obtain the improved high-resolution band images. The improved high-resolution band images are replaced by the first part of GS transform to generate a new data set, and the spatial resolution enhanced multispectral images are generated by reverse GS transform. The images fusion process is shown in Figure 1.

C. The Improved U-net Convolutional Neural Network

The traditional U-Net network model has a long training period, a significant weight file, and a high production cost. It is not conducive to the automatic or accurate classification of planting structures. This paper proposes the use of a deep, separable convolution module instead of the full convolution method in the original network model. The Mish activation function is used instead of the ReLU activation function to further improve the training stability and optimize the prediction accuracy, as shown in Figure 2.

1) Depth-separable convolution

Depth-separable convolution (depthwise separable convolution), which is composed of two parts, depthwise (DW) and pointwise (PW), is used in this paper to extract features. Compared with conventional convolution operations, the number of parameters and operational cost of this method are relatively low^[44]. Depth-separable convolution effectively reduces the number of required parameter calculations. The parameter values are obtained by adding the two parts, consisting of approximately one-third of the conventional convolution calculation burden. Depth-separable convolution considers both channel and regional changes, while the conventional convolution method only considers the region first and then considers the channel, separating the channel and region.

2) Mish activation function

In this paper, the ReLU (ReLU = max(0, x)) activation function is replaced in the U-Net network with a smoother Mish function with no boundaries (that is, the positive values can reach any height) to avoid saturation due to capping. This continuous activation function allows better information to penetrate deep into the neural network, resulting in better accuracy and generalization abilities.

Theoretically, slightly negative values allow better gradient flows compared to the hard-zero boundaries applied in ReLU; this characteristic further improves the efficiency and accuracy of the operation^[45]. The formula of the Mish function is Mish=x*tanh(ln(1+e^x)).

3) The improved u-net network model

In this paper, a deep, separable convolution module is used to replace the full convolution method. A Mish activation function is used to restore the ReLU activation function to design an improved U-Net network model. Subsequent research shows that the improved network model is more classified than the traditional U-Net model when predicting the planting structure. The accuracy is further optimized, and the model structure is shown in the figure below.

4) network training process

The entire set of pairs of images and labels is input into the U-Net model and fitted in a long iteration process that includes training the network model and predicting the classifications. The training dataset is input into the network model for training, and the trained weight file is used to predict the test set to obtain the classification results. The training and prediction processes are shown in Figure 3. This article uses the Keras Python deep-learning library^[46] and TensorFlow back-end driver^[47] to achieve the model building, training, and prediction processes, as shown in Figure 3.

5) method of Evaluation

In this paper, the accuracy, recall rate, mean of the two reconciled values, and kappa coefficient are used as indicators of the evaluation method. These indicators are calculated based on the confusion matrix, and the accuracy is calculated as follows:

$$\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}=\frac{1}{6}{\sum }_{i}{C}_{ii}/{\sum }_{j}{C}_{ij}$$

3.5

where, C_ii indicates the number of correctly classified samples and C_ij suggests type-I samples that are mistaken for type J.

The recall metric represents the correct average proportion of pixels classified into a given category. The recall is calculated as follows.

$$\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}=\frac{1}{6}{\sum }_{i}{C}_{ii}/{\sum }_{i}{C}_{ij}$$

3.6

The model can also be further evaluated by calculating the harmonic mean F1 of the F accuracy and recall rate. The F1 value is calculated as follows.

$$\text{F}1=2\text{*}\frac{\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}\text{*}\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}{\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}+\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}$$

3.7

The kappa coefficient measures the consistency between the predicted class and the artificial label. The kappa coefficient is calculated as follows.

$$\text{K}\text{a}\text{p}\text{p}\text{a}=\frac{\frac{\frac{1}{6}{\sum }_{i}{C}_{ii}}{\sum _{i=1}^{6}\sum _{j=1}^{6}{C}_{ij}}-\frac{\frac{1}{6}{\sum }_{i}{C}_{ii}{\sum }_{i}{C}_{ij}}{{\left(\sum _{i=1}^{6}\sum _{j=1}^{6}{C}_{ij}\right)}^{2}}}{1-\frac{\frac{1}{6}{\sum }_{i}{C}_{ii}{\sum }_{i}{C}_{ij}}{{\left(\sum _{i=1}^{6}\sum _{j=1}^{6}{C}_{ij}\right)}^{2}}}$$

3.8

Equations (3.5)–(3.8) use the definitions given in reference^[48] and are modified according to the actual situation.

In the research process, the classification effect of the U-Net deep semantic segmentation network on the planting structure is found to be better than those of the FCN network and the SegNet network model. Therefore, in this section, only U-Net and its improved models are compared and analyzed. In this paper, the conventional U-Net model, the U-Net model with an added deep, separable convolution module, and the activation function network model are trained in 200 batches on the same test dataset to obtain the accuracy curves shown in the following figure 4.

Figure 4. Results of U-Net and its improved models obtained from a comparison diagram: (a) U-Net training accuracy rate curve; (b) Separable-UNet training accuracy rate curve; (c) Mish-Separable-UNet training accuracy rate curve.

In this paper, the G-S image fusion method was used in combination with a deep semantic segmentation network to provide valuable spectral feature information. That is, this method was applied to improve the classification accuracy of the planting structure.

Compared with the traditional U-Net model, after adding the depth-separable convolution module, the model reaches an overfitting batch from 100 to 30, significantly reducing the calculation cost and increasing the fitting rate. After using the Mish activation function, the accuracy and calculation rate are slightly increased due to the smoothing characteristics of the function. The improved network model achieves lighter parameters, and the weight file size is reduced from

303 MB to 53 MB, a reduction of 82.5%, significantly improving the production cost.

The original remote sensing images of seven bands from 1 to 7 are synthesized with 1, 2, and 3 exponents. Then, the G-S method is used to fuse the panchromatic groups to further increase the spatial resolution and spectral richness of the training images and predict the results, as shown in the picture below.

The statistical test set prediction results are compared with the label data to generate the confusion matrices shown in Figure 6 below to judge the classification quality of the planting structure.

Figure 6. The classification result from confusion matrix (a) seven-band confusion matrix; (b) synthetic NDVI confusion matrix; (c) synthetic AWEI confusion matrix; (d) synthetic SAVI confusion matrix; (e) synthetic NDVI and AWEI confusion matrix; (f) Synthesis of NDVI and SAVI confusion matrix; (g) Synthesis of AWEI and SAVI confusion matrix; (h) Synthesis of three exponential confusion matrices.

Table 1 compares the prediction results of the various methods with the actual label data to calculate the accuracy rate. After the investigation, most of the woodland area is found to comprise economic crops such as jujube trees, so wheat, woodland, and cotton areas are classified as crop areas.

TABLE 1

Classification accuracy statistical table.

Index	Water	Wheat	Forest	Cities	Cotton	Crops	Accuracy
No Indices	80.6%	77.5%	71.0%	95.0%	86.9%	81.7%	90.3%
NDVI	81.4%	84.3%	73.8%	92.8%	86.6%	83.5%	92.4%
AWEI	81.4%	83.9%	74.2%	93.9%	88.2%	84.1%	92.7%
SAVI	80.3%	84.4%	69.7%	93.3%	85.6%	82.4%	91.6%
NDVI, AWEI	81.8%	85.2%	74.4%	93.5%	87.5%	84.2%	93.7%
NDVI, SAVI	81.0%	82.3%	73.8%	93.1%	87.2%	83.4%	92.6%
AWEI, SAVI	82.1%	81.8%	74.4%	93.5%	87.5%	83.9%	93.3%
Three Indices	81.5%	85.3%	74.1%	93.0%	87.8%	84.1%	93.8%

As shown in Table 1, the overall accuracy of the 7-band fusion image without the addition of any index reached 90.3%, and the accuracy of increasing the index image participation in the image fusion process was higher than that obtained without the index. Adding the index uses the improved U-Net network in the training process for identifying crop cultivation areas, the structural classification has a positive effect on the results.

With the addition of each of the three indexes individually, all showed an increase in the overall accuracy rate, among which AWEI had the most significant impact on the classification of the planting structure. The crop identification accuracy was increased by 2.4%, and the city and water pixels were better identified. This index is superior to the other two indexes, which reached accuracies of 81.4% and 93.9%, and the improvement effect is visible. The overall accuracy rate reaches 92.7%, which is the best among the three indexes. However, this index tends to judge pixels as cotton when recognizing mixed pixels of wheat and cotton, and its ability to identify crops in a large mixed area remains to be discussed.

Under the condition of two-index 9-band fusion images, AWEI still plays an important role. The classification effect of training images containing AWEI is better than that of the NDVI-SAVI 9-band fusion images. The addition of NDVI further improves the crop recognition ability of AWEI and the mutual recognition ability among crops. The accuracy rates of the three types of plants, wheat, cotton, and woodland, increased to 85.2%, 74.4%, and 87.5%, respectively. At the same time, this method also guarantees that the high recognition accuracy of water and cities reaches 81.8% and 93.5%, respectively. Only the water identification efficiency is lower than that of the AWEI-SAVI method, with a difference of 0.3%, and the overall accuracy rate of this method reaches 93.7%.

Compared with the NDVI-AWEI 9-band fusion image method, the overall accuracy of the 10-band fusion image is improved by only 0.1%, and the crop classification recognition is reduced. SAVI's improvement in the fusion image is limited. Considering the production costs, the priority is lower. In summary, the NDVI-AWEI 9-band fusion image method combined with the improved U-Net network model classification planting structure method has advantages.

This paper evaluates the reliability and accuracy of the planting structure classification results of each index method combined with the improved U-Net model by calculating the kappa coefficient, precision, recall and harmonic mean F1, as shown in Table 2 below. Compared with nonindexed images, the classification accuracy of the indexed fusion image planting structure is significantly improved. The kappa coefficient increases from 0.868 to more than 0.874 with the addition of the index, further improving the consistency between the predicted class and the artificial label. The accuracy, recall, and average value are enhanced by more than 0.02. In the new method, the NDVI- AWEI fusion method has the same accuracy as the three-index method, reaching 0.873. The kappa coefficient and F1 value are better than the three-index method, at 0.886 and 0.872, respectively; thus, this method can provide accurate and stable classification results of the crop planting structure. As shown in Figure 5, the classification confusion in the boundary areas of feature types is still the primary source of error. Figure 7 below shows the statistics of the confusion types and their probabilities under various methods.

TABLE 2

Statistics of each method

Band	Index	Precision	Recall		Kappa
7	No Indices	0.853	0.851	0.852	0.868
8	NDVI	0.866	0.865	0.865	0.880
	AWEI	0.871	0.869	0.870	0.885
	SAVI	0.861	0.856	0.858	0.874
9	NDVI, AWEI	0.873	0.871	0.872	0.886
	NDVI, SAVI	0.863	0.862	0.863	0.878
	AWEI, SAVI	0.868	0.867	0.867	0.882
10	Three Indices	0.873	0.869	0.871	0.885

Due to the increase in spectral richness and the increase in the information contained in the training images, the feature information extracted using the improved U-Net network is enhanced to a certain extent. The overall confusion rate of the prediction results declines, but the edge is not fundamentally resolved. Regarding misclassifications, forest pixels still maintain the highest probability of confusion due to their discrete nature. Using the improved U-Net network to train the NDVI-AWEI 9-band fusion image method compared to other means, the minimum confusion level in each misjudgment type can be maintained at 2%-9%, representing the improved U-Net model planting structure classification model.

This paper proposes a method for planting structure classification based on exponential and improved U-Net deep semantic segmentation network. High spatial resolution GS fusion images highlighting different features are obtained by combining multiple index data and Landsat8 original remote sensing images, combined with improved U-Net deep semantic segmentation network realizes the classification of crop planting structure.

This paper proposes a method for classifying planting structures based on an exponential and improved U-Net deep semantic segmentation network. High-spatial-resolution GS-fusion images highlighting different features are obtained by combining multiple index data and Landsat 8 original remote sensing images with an improved U-Net deep semantic segmentation network to classify crop planting structures.

This research shows that, compared with the unfused index data, the crop classification accuracy of the proposed method is more than 82.4%, and the overall efficiency of the categorization of features (including rivers, towns, etc.) is more than 91.6%. The NDVI-AWEI method has a kappa coefficient of 0.886, and its accuracy is above 0.873. Compared with the three-index method, less synthetic data is required, the training dataset is smaller, and the production cost is lower. Among the three indexes, AWEI has the highest classification accuracy for images. The method based on the combined NDVI-AWEI indexes and improved U-Net network is the most suitable for the automatic and accurate classification of planting structures.

Further improvements to this method are expected to continue to improve the classification accuracy of planting structures through the following two aspects: (1) we will attempt to use the CNN network to construct a fusion algorithm to optimize the image fusion process and retain more detailed spectral information; and (2) we will attempt to use other indexes joined in the fusion image process to further strengthen the different images and meta-weight distributions to reduce boundary confusion and obtain a better classification performance.

Xiong, Q. X. & Huang, J. F. Estimation of autumn harvest crop planting area based on NDVI sequential characteristics[J]. Transactions of the CSAE, 25 (1), 144–148 (2009).
Zhang, J. H., Feng, L. L. & Yao, F. M. Improved maize cultivated area estimation over a large scale combining MODIS-EVI time series data and crop phonological information[J]. Photogramm. Remote Sensing, 94, 102–113 (2014).
Wardlow, B. D. & Egbert, S. L. Large-area crop mapping using time-series MODIS 250m NDVI data: An assessment for the U. S. Central Great Plains[J]. Remote Sensing of Environment, 112, 1096–1116 (2008).
Xu, Q. Y. et al. Crop information identification based on MODIS NDVI time-series data[J]. Transactions of the Chinese Society of Agricultural Engineering. CSAE, 30, 134–144 (2014).
Zhao Shuang. Precise Mapping and Spatiotemporal Analysis of Paddy Rice Area in Complex Surface Landscape [D] (China University of Geosciences(Beijing), 2018).
Baojia, D. U. et al. Crop Mapping based on Sentinel-2A NDVI Time Series Using Object-Oriented Classification and Decision Tree Model[J]. Journal of Geo-information Science, 21 (05), 740–751 (2019).
Yanying, B. A. I., Julin, G. A. O. & Baolin, Z. H. A. N. G. Extraction of crop planting structure Based on time-series NDVI of Landsat8 images [J/OL]. Arid Land Geography:1-10[2019-07-17].http://kns.cnki.net/kcms/detail/65.1103.x.20190517.1346.024. html.
Xiong Yuankang, Z. Q. Cropping Structure Extraction with NDVI Time-series Images in the Northern Tianshan Economic Belt[J/OL]. Arid Land Geography:1-18[2019-07-17].http://kns.cnki.net/kcms/detail/65.1103.X. 20190711.1807.028.html.
Fan, Y. A. N. G., Jianhua, G. U. O. & Hai, T. A. N. WANG, Jingxue Image segmentation based on gray histogram and K-nearest neighbors algorithm [J].Science of Surveying and Mapping,2017, 42(03):7–11.
Shu-kui, B. O. JING Yong-ju. One-class Information Extraction from Remote Sensing Imagery Based on Nearest Neighbor Rule [J]. Computer Science,2016,43(S1):217-218+259.
Lin Yiyang, L., Shijin, M. & Zhaohui Scene classification of multi-feature fusion remote sensing image based on clustering analysis[J].Electronic Measurement Technology,2018,41 (22):82–88.
Pacifici, F., Chini, M. & Emery, W. J. A neural network approach using multi-scale textural metrics from very high-resolution panchromatic imagery for urban land-use classification[J]. Remote Sensing. Environ, 113, 1276–1292 (2009).
Huang, X. & Zhang, L. An SVM ensemble approach combining spectral, structural, and semantic features for the classification of high-resolution remotely sensed imagery[J]. IEEE Trans. Geosci. Remote Sensing, 51, 257–272 (2013).
Liu, C., Hong, L., Chen, J., Chun, S. S. & Deng, M. Fusion of pixel-based and multi-scale region-based features for the classification of high-resolution remote sensing image[J]. Remote Sensing, 19, 228–239 (2015).
Yuan, Y., Lin, J. & Wang, Q. Hyperspectral Image Classification via Multitask Joint Sparse Representation and Stepwise MRF Optimization[J]. IEEE Transactions on Cybernetics, 46, 2966–2977 (2016).
Xie, F. D., Li, F. F., Lei, C. K. & Ke, L. N. Representative Band Selection for Hyperspectral Image Classification [J]. Journal of Visual Communication and Image, 7, 338 (2018).
BALL, J. E. & ANDERSON D T, CHAN C, S. A comprehensive survey of deep learning in remote sensing: theories, tools and challenges for the community[J]. Journal of Applied Remote Sensing, 11 (4), 042609 (2017).
ZHANG, L., ZHANG, L. & DU, B. Deep learning for remote sensing data: a technical tutorial on the state of the art [J]. IEEE Geoscie nce and Remote Sensing Magazine, 4 (2), 22–40 (2016).
ZHU X X, T. U. I. A. D. et al. Deep learning in remote sensing: a comprehensive review and list of resources[J]. IEEE Geoscience and Remote Sensing Magazine, 5 (4), 8–36 (2017).
MOU, L., ZHU X X & RiFCN Recurrent network in fully convolutional network for semantic segmentation of high-resolution remote sensing images[EB/OL]. (2018-05-16). https:// arxiv.org/abs/1805.02091.
ZHAO W Z, DU S H & EMERY W J Object-based convolutional neural network for high-resolution imagery classification [J].IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2017:1–11.
Xu, F. A. N. G. et al. High-resolution remote sensing image classification combining with mean -shift segmentation and fully convolution neural network[J].Laser & Optoelectronics Progress, 2018(55):446–454.)
Guangsheng, C. H. E. N. et al. Remote sensing image pixel -level classification with genetic algorithm and fully convolu-tional network[J]. Journal of Chinese Computer Systems, 39 (7), 1590–1595 (2018).
SHRESTHA, S. VANNESCHI L. Improved fully convolutional network with conditional random fields for building extraction[J]. Remote Sensing, 10 (7), 1135 (2018).
FU, G. et al. Classification for high-resolution remote sensing imagery using a fully convolutional network[J]. Remote Sensing, 9 (5), 498 (2017).
MOU, L. & ZHU X, X. Vehicle instance segmentation from aerial image and video using a multitask learning residual fully convolutional network[J]. IEEE Transactions on Geoscience and Remote Sensing, 56 (11), 6699–6711 (2018).
LONG, J., SHELHAMER, E. & DARRELL, T. LONG J, DARRELL T. Fully convolutional networks for semantic segmentation[J].IEEE
Badrinarayanan, V., Kendall, A., Cipolla, R. & SegNet A deep convolutional encoderdecoderarchitecture for image segmentation[EB/OL]. (2015-11-02). https://arxiv.org/abs/ 1511.00561.
Ronneberger, O., Fischer, P., Brox, T. & U-Net Convolutional Networks for Biomedical Image Segmentation[J].Medical Image Computing and Computer-Assisted Intervention (MICCAI) 2015, 93 51: 234-241.
Chen, L. C. et al. Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs[J].arXiv, 2016; arXiv:1606.00915.
Lin, H. N., Shi, Z. W. & Zou, Z. X. Maritime Semantic Labeling of Optical Remote Sensing Images with Multi-Scale Fully Convolutional Network[J]. Remote Sensing, 9, 480–501 (2017).
Visin, F. et al. A recurrent neural network-based model for semantic segmentation[J].arXiv, 2016; arXiv:1511.07053.
Neil Flood,Fiona Watson,Lisa Collett. Using a U-net convolutional neural network to map woody vegetation extent from high-resolution satellite imagery across Queensland, Australia[J].International Journal of Applied Earth Observations and Geoinformation, 2019,82.
Hirotsugu, N. Mizuho Nishio,Rikiya Yamashita,Ayako Ono,Kyoko Kameyama Nakao,Koji Fujimoto,Kaori Togashi. Quantitative and Qualitative Evaluation of Convolutional Neural Networks with a Deeper U-Net for Sparse-View Computed Tomography Reconstruction[J]. Academic Radiology,2019.
Yue, X. U., Mengru, F. E. N. G., Jiatian, P. I. & Yong, C. H. E. N. Remote Sensing Image Segmentation Method Based on DeepLearning Model [J/OL]. Journal of Computer Applications:1-12[2019-07-18].http://kns.cnki.net/kcms/detail/51.1307.TP.20190621.1037.002.html.
Xu Hongwei, Y., Peixin, W., Min, X., Zhenyu, S. & Yubao Automated segmentation of cystic kidney in CT images using residual double attention motivated U-Net model [J/OL]. Application Research of Computers:1-5[2019-07-18].https://doi.org/10.19734/j.issn.1001-3695.2019.03.0092.
Wang Haiou, L. et al. Design of Superpiexl U-Net Network for Medical Image Segmentation [J]. Journal of Computer-Aided Design & Computer Graphics, 31 (06), 1007–1017 (2019).
Minqiang, C. H. E. N., Puyang, A. N. & Junchao, Z. H. O. U. GF-1 Remote Sensing Image Fusion and Quality Evaluation Based on Mountain Area[J].Beijing Surveying and Mapping,2019,33(03):249–254.
Pengfei, Y. A. N. G., Xiuying, L. I. A. O., Hui, C. H. E. N. G., Murong, S. H. U. A. I. & Yiwen, X. I. E. A comparative study of remote sensing image fusion methods based on spectral gradient angle and spectral information divergence index[J].Engineering of Surveying and Mapping,2018,27(11):51–55.
Sun, W., Chen, B. & Messinger, D. W. Nearest-neighbor diffusion-based pan-sharpening algorithm for spectral images[J]. Opt. Eng, 53 (1), 013107 (2014).
Jiwei, C. H. E. N., Qian, S. H. E. N., Li, Z. H. U. & Shuangting, W. A. N. G. Influence of Different Fusion Methods Based on GF Images on the Spectral Fidelity of Urban Water[J].Journal of Beijing University of Technology,2017,43(05):677–682.
Yalan, Z. & Leiguang, W. Lu Xiang. Comparison of Image Fusion Methods for Gaofen-2 Panchromatic-Multispectral [J]. Journal of Southwest Forestry University(Natural Sciences),2018,38(02):103–110.
Feyisa, G. L. et al. Automated Water Extraction lndex: A new technique for surface water mapping using Landsat imagery[J].Remote Sensing of Environment,2014, 140(1):23-35.
Guo, J. et al. Network Decoupling: From Regular to Depthwise Separable Convolutions[C]. 29the British Machine Vision Conference (BMVC 2018). 2018.
Misra, D. & Mish A Self Regularized Non-Monotonic Neural Activation Function[J/OL]. https://arxiv.org/abs/1908.08681. 2019.
Chollet, F. et al. 2015. Keras. https://keras.io.
Abadi, M. et al. 2015. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems.. Software available from tensorflow. org. https://www.tensorflow.org/.
Yue, X. U., Mengru, F. E. N. G., Jiatian, P. I. & Yong, C. H. E. N. Remote sensing image segmentation method based on deep learning model [J].Journal of Computer Applications,2019,39(10):2905–2914.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Research on a Classification Method of Crop Planting Structures Based on Indexes and an Improved U-Net Network

Status:

Version 1

Abstract

Figures

Introduction

Date

Math

A. Index selection

1) Normalized difference vegetation index

2) NND Image Fusion Method

3) Soil regulation vegetation index

B. Image Fusion

C. The Improved U-net Convolutional Neural Network

1) Depth-separable convolution

2) Mish activation function

3) The improved u-net network model

4) network training process

5) method of Evaluation

Results And Analysis

Conclusion

References

Additional Declarations

Status:

Version 1