Fusing multi-source data to map spatio-temporal dynamics of winter rape on the Jianghan Plain and Dongting Lake Plain, China

TAO Jian-bin, LIU Wen-bin, TAN Wen-xia, KONG Xiang-bing, XU Meng

1 Key Laboratory for Geographical Process Analysis & Simulation of Hubei Province/School of Urban and Environmental Sciences,Central China Normal University, Wuhan 430079, P.R.China

2 Key Laboratory of the Loess Plateau Soil Erosion and Water Loss Process and Control, Ministry of Water Resources/Yellow River Institute of Hydraulic Research, Zhengzhou 450003, P.R.China

Abstract Mapping crop distribution with remote sensing data is of great importance for agricultural production, food security and agricultural sustainability. Winter rape is an important oil crop, which plays an important role in the cooking oil market of China. The Jianghan Plain and Dongting Lake Plain (JPDLP) are major agricultural production areas in China. Essential changes in winter rape distribution have taken place in this area during the 21st century. However, the pattern of these changes remains unknown. In this study, the spatial and temporal dynamics of winter rape from 2000 to 2017 on the JPDLP were analyzed. An artificial neural network (ANN)-based classification method was proposed to map fractional winter rape distribution by fusing moderate resolution imaging spectrometer (MODIS) data and high-resolution imagery. The results are as follows: (1) The total winter rape acreages on the JPDLP dropped significantly, especially on the Jianghan Plain with a decline of about 45% during 2000 and 2017. (2) The winter rape abundance keeps changing with about 20-30% croplands changing their abundance drastically in every two consecutive observation years. (3) The winter rape has obvious regional differentiation for the trend of its change at the county level, and the decreasing trend was observed more strongly in the traditionally dominant agricultural counties.

Keywords: winter rape, spatio-temporal dynamics, time-series MODIS data, artificial neural network

1. Introduction

Rape, also known as rapeseed, is the third-largest source of vegetable oil in the world (USDA 2018). Rape is also an important oil crop and the fifth-largest crop in China following rice, wheat, corn and soybean. Winter rape accounts for over 90% of the total rapeseed in both planting area and production of China, and is mainly concentrated in the Yangtze River Valley. The Jianghan Plain and Dongting Lake Plain (JPDLP) are major agricultural production areas of grain, cotton and oil in the Yangtze River Valley.The planting area and yield of winter rape has fluctuated frequently in recent years due to rapid urbanization, the fast development of the market economy, and the impact from international markets. However, the spatio-temporal patterns of winter rape in this area remain unknown.Therefore, monitoring the spatio-temporal dynamics of winter rape is of great significance to the development of the vegetable oil market and food security of China (She et al. 2013).

Coarse resolution vegetation indices have been widely used in the remote sensing community and have achieved successful applications in crop mapping worldwide (Wardlow et al. 2007; Chen et al. 2011; Conrad et al.2011; Esquerdo et al. 2011; Pan et al. 2012; Vintrou et al.2012; Atzberger and Rembold 2013; Estel et al. 2015; Liu et al. 2015; Shrestha et al. 2017). Time-series normalized difference vegetation index (NDVI) data from the moderate resolution imaging spectrometer (MODIS) satellite sensor carry useful information about the seasonal vegetation development, and this information can facilitate the work of crop mapping. There are many published studies on crop mapping using MODIS data in Northeast or North China, however few such studies have focused on Central or South China. The mapping of grain crops, such as rice,wheat, maize, and others has been intensively investigated in previous studies, however there has been little research on winter rape. Crop rotation and intercropping are very common due to the fragmented cropland fields constrained by the local natural environment and at the same time, they are controlled by farmer’s planting behavior on the JPDLP.So the problem of mixed pixels is often observed in satellite imagery in this area. Therefore, mapping winter rape on the JPDLP is more challenging concerning data sources and classification methods.

The artificial neural network (ANN)-based classification method has unique advantages and it is heavily used by researchers all over the world. Neural networks can independently learn the nonlinear relationships between the time-series NDVI profiles and the end member abundance(Atzberger and Rembold 2013), and the ANN based subpixel resolution is better than that of traditional spectral linear resolution methods (Verbeiren et al. 2008). Generally speaking, it is relatively easy to extract winter crops in images because they have quite different phenological calendars from natural vegetation. However, it is quite difficult to distinguish winter rape from winter wheat, since they are both dominant winter crops and have similar phenologies. The potential of ANN methods for winter rape mapping at a large-scale remains unknown.

The overall objective of this study is to explore the spatial and temporal dynamics of winter rape on the JPDLP. The specific objectives include: (1) Developing an ANN-based method to map fractional winter rape distribution at the regional scale by fusing multi-source data. An ANN model can build a map between winter rape abundance and MODIS Enhanced Vegetation Indices (EVI) profiles. (2) Analyzing the spatio-temporal patterns of winter rape, in order to obtain the knowledge of the winter rape dynamics on the JPDLP, and to provide a basis for agricultural production. The novelty of this research is that it is the first work to explore the spatiotemporal dynamics of winter rape in South China, where it spans a vast territory with varied topography and complex planting structures. Fractional winter rape maps for 2000-2017 were obtained, and the spatial and temporal patterns over the 18 years were analyzed using statistical methods.

2. Research area and data

2.1. Research area

The research area is the JPDLP, which is located in the middle reaches of the Yangtze River Valley of China and covers a vast plain area including the Jianghan Plain in Hubei Province and the Dongting Lake Plain in Hunan Province (Fig. 1). This research focused on the main winter crop areas of the JPDLP, which include 37 counties(districts) covering an area of approximately 60 000 km2.The cropland takes up 65.35% of the total land resources.The JPDLP is an important national agricultural product base for grain, cotton and oil production and the main crops in this area are rice, winter wheat, winter rape, cotton and soybean. Winter rape occupied about one-seventh of the national total planting area in 2015 (HPBS 2016; NBSC 2016). The dominant land-cover types in the research area are forests, shrublands, grasslands, croplands, water bodies, and constructed surfaces. Forests, shrublands, and grasslands are combined to a more general type, natural vegetation, since our focus on croplands does not require precise classification of the natural vegetation.

2.2. Data

The datasets used in this research include: (1) MOD13Q1 EVI products for 2000-2017; and (2) Landsat Thematic Mapper (TM)/Enhanced Thematic Mapper Plus (ETM+)/Operational Land Imager (OLI), HJ-1A/B and Gaofen-1(GF-1) satellite images for 2000-2017. Other ancillary data include 1:4 000 000 vector datasets of China, Google Earth historical images, and other sources.

MODIS dataThe MODIS data used in this study were the MOD13Q1 v006 data, consisting of MODIS products recorded by the EOS/Terra Satellite covering years 2000 to 2017. The products include 250-m resolution NDVI and EVI data, reflectance data and quality control data, which were synthesized over 16 days based on the Maximum Value Composite (MVC) method. The products were corrected geometrically and atmospherically. The dataset was downloaded from the website of the U.S. NASA LP DAAC working group. The MODIS image sequence numbers of the tile covering the research area are h27v05, h27v06,h28v05, h28v06 (h, horizontal; v, vertical). In addition, a data pixel reliability layer was extracted from the MOD13Q1 v006 products, the spatial and temporal resolutions of which were consistent with the NDVI and EVI datasets. Since NDVI is chlorophyll sensitive and the EVI is more responsive to canopy structural variations, including leaf area index (LAI),canopy type, plant physiognomy, and canopy architecture(Huete et al. 2002), the EVI data layer was extracted as the vegetation index datasets in this research.

Fig. 1 The research area and its main land-cover types (from Globe Land30 2010 datasets). A, Jianghan Plain; B, Dongting Lake Plain; C, Honghu Lake; D, Dongting Lake.

High-resolution imagesHigh-resolution remote sensing images including Landsat TM/ETM/OLI, HJ-1A/B and GF-1 were obtained to generate samples. By referencing the agricultural phenological calendar, we collected all available high-resolution images covering the winter rape flowering period (mid-March to early April) on the JPDLP (Table 1).We have chosen the images with the highest quality when there were multiple satellite images in one year. The Landsat images were obtained from the U.S. Geological Survey (USGS) website (https://earthexplorer.usgs.gov/),and the HJ-1 and GF-1 images were obtained from the China Resources Satellite Application Center website (http://www.cresda.com/CN/). The GlobeLand30 land-cover data (http://www.globallandcover.com/GLC30Download/index.aspx)were analyzed along with Google Earth high-resolution images to generate winter rape samples.

Table 1 High-resolution images used in this research1)

2.3. Data preprocessing

Geographical geometric correction, image clipping, mosaic and resampling were performed on all the datasets. All the datasets were re-projected into the geographic coordinates of the WGS 1984 coordinate reference system with the UTM 49N projection coordinate system. The method adopted for sampling was the nearest neighbor.

The interference from sun illumination angle, clouds,snow and other sources of noise affected the reliability of the MODIS EVI time-series data. The TimeSat Software package was used to fit and smooth the time-series EVI data (Jönsson and Eklundh 2004). The fitting was optimized in an iterative process, and a smooth curve describing the time-series EVI data could therefore be reconstructed. The above process reduced the noise effects induced by clouds and clearly revealed the phenological pattern contained in the temporal EVI profiles.

3. Methods

The proposed method for mapping sub-pixel winter rape consists of three steps: (1) preparing MODIS-like samples from multi-source high-resolution remote sensing images;(2) developing an ANN-based method to estimate winter rape abundance from time-series MODIS data; and (3)analyzing the spatio-temporal distribution of winter rape using spatial analysis and statistical tests.

3.1. Samples at the sub-pixel level

Winter rape and winter wheat are the main winter crops on the JPDLP, having similar phenology (both are sown in late October and harvested in early June of the next year), so the key issue for extracting winter rape is to distinguish it from winter wheat. Through intensive study of the remote sensing images within the life cycle of winter crops, we found that the most appropriate period for distinguishing them visually in true-color composite satellite images is from middle March to early April. During this period, winter rape is in the flowering period and exhibits a bright yellow color,while the winter wheat is in the jointing stage, showing up as dark green in the images (Fig. 2-A).

The reflectance of red and green bands is relatively high for winter rape; however, the reflectance of all visible bands,especially the red band, is low for winter wheat (Fig. 3).According to this spectral difference, we constructed a winter rape index as follows:

The histogram statistics of the spectral bands and the winter rape index (winter rape vs. winter wheat) indicated that the ability to separate winter rape and winter wheat on the winter rape index image was significantly improved (Fig. 4). It is also easier to distinguish winter rape from other land-cover types on the winter rape index image. The index eliminated the regional differences in phenology and enhanced the contrast between winter rape and winter wheat.

We prepared samples for every observation year independently. Two sample areas were selected from the high-resolution images, respectively, one for training and the other for validation. To ensure the accuracy of the results, samples were chosen to cover all land-cover types in the research area, such as winter rape, winter wheat,natural vegetation, water bodies, constructed surfaces,and others. The spatial extent of samples kept changing due to the varied coverage of multi-source high-resolution images and the demand for minimizing the interference of clouds. The spatial extent of the samples is indicated in Fig. 5 (taking 2015 GF-1 data as an example). The samples accounted for approximately 5% of the total pixels in the research area.

A decision tree combining the self-organizing data analysis technique (ISODATA), along with visual interpretation using Google Earth satellite images, was used to obtain highresolution winter rape samples. The classification maps were then simply reclassified to a binary winter rape/other land-cover type classification scheme. To match the resolution of MODIS data, the sample data were aggregated to 250 m, so that each MODIS-like pixel provided the percentage of winter rape in that pixel. The fractional image has the data range of 0-1, in which ‘1’ indicates winter rape coverage of 100% and ‘0’ indicates the coverage of 0%. The sample data were used (1) as an input for ANN, and (2) for validation samples of the ANN result.

3.2. Winter rape extraction at the sub-pixel level using ANN

The proposed method is based on the theory that different land-cover type proportions within MODIS pixels determine variations in EVI profiles. There are significant variations in main land-cover types such as natural vegetation, double crop, single crop, water bodies and constructed surfaces(Tao et al. 2017a, b). Our previous research also found that winter crops have quite different phenological calendars than natural vegetation, showing distinct phenological characteristics which differ from those of the background land-cover types. The dominant winter crops on the JPDLP are winter rape and winter wheat. So the main obstacle for extracting winter rape is to distinguish it from winter wheat.The EVI curves of the two winter crops had opposite trends between middle March and early April (DOY: 065-097),during which the winter rape entered the flowering period with a decreasing EVI, while winter wheat continued to grow with an increasing EVI (Fig. 6). These curves from EVI profiles clearly show the potential for discriminating between the classes.

Fig. 2 A comparison of the true color composite of GF-1 image (A) and the winter rape index image (B) in the core area of the Jianghan Plain, China, in 2015.

Fig. 3 Spectral curves of winter rape and winter wheat.

ANN can learn the nonlinear relationships between time-series EVI signatures and the fractional winter rape information. The topology of the network used in this study is a simple three-layer feed-forward backpropagation neural network (BP neural network). The input layer has 18 input nodes, including 17 feature nodes indicating the time-series EVI images covering the life cycle of winter crops and one sample node. The output layer has only one node, which is the winter rape abundance modeled by the network. The number of neurons in the hidden layer was determined according to the formula: M=2n+1, where,n is the number of input layer nodes (Kavzoglu and Mather 2003). Therefore, the number of neurons in the hidden layer was set to 37, finally resulting in a compact 18-37-1 network topology. The hyperbolic tangent sigmoid(y(x)=2/(1+exp(-2x))-1) was used as the transfer function for the hidden layer and the log-sigmoid (y(x)=1/(1+exp(-x)))for the output layer. The ANN output automatically remained within the boundary of 0.0-1.0 which is consistent with the data range of winter rape abundance. Table 2 shows the training results from 2000 to 2017.

3.3. Spatial-temporal analysis of the pattern

Spatial patternHotspot analysis was used to express the spatial pattern of winter rape. Hotspot analysis uses vectors to identify the locations of statistically significant hot spots and cold spots in data. Moran’s index can indicate clustering or dispersion patterns in the data and Z scores and P-values for evaluating the significance of that index.A high Z score and small P-value for a feature indicates a significant hot spot, while a low negative Z score and small P-value indicates a significant cold spot. The higher (or lower) the Z score, the more intense the clustering. A Z score near zero means no spatial clustering. Getis-Ord Gi (Gi＊) in ArcGIS 10.2 was used to calculate the Moran’s I index value in addition to both a Z score and a P-value.The Getis-Ord local statistic is given as:

Fig. 4 Histograms comparing the features of winter rape vs. winter wheat. A, blue band. B, green band. C, red band. D, near infrared. E, short wave infrared. F, winter rape index. DN, digital number. About 5 000 samples were selected for winter wheat and winter rape, respectively, in the 2015 GF-1 image.

Fig. 5 Winter rape mainly covers Jiangling, Qianjiang and Xiantao from GF-1 in 2015.

Fig. 6 Enhanced Vegetation Indices (EVI) curves for main landcover classes on the Jianghan Plain and Dongting Lake Plain,China. DOY, day of a year. The mean values of about 300 samples for each land-cover type in 2015 MODIS data were used to draw this figure.

Table 2 Training results of the neural wetwork

where wi,jis the spatial weight between feature i and j, xjis the attribute value for feature j, and n is equal to the total number of features.

Temporal patternThe temporal pattern was evaluated from three aspects: the change of total planting areas, annual change analysis, and the change trend at the county level.

To measure the change of total planting areas, we simply summarized the winter rape abundance using the zonal statistics tool in ArcGIS at the county level and then summarized to get the total areas. Annual change analysis included yearly change maps, as well as the intensity and the proportion of changes of winter rape abundance. The yearly change maps were presented as winter gains and winter rape losses, which is the difference of abundance values between two consecutive years (positive and negative values correspond to gains and losses, respectively). The accumulated amount of changes in winter rape abundance in the time-series was used to evaluate the intensity of changes. At the pixel level, we obtained the difference of abundance values between two consecutive years, and summarized the absolute values of the differences within the time-series. At the county level, we first summarized the abundance value to the county level. Then the method similar to the one used at the pixel level was used to assess the intensity of changes. We assessed the proportion of changes by counting those pixels with drastic changes in winter rape abundance, then dividing by the total winter rape pixels. If the difference of abundance between two consecutive years is greater than 0.5, it can be seen as having a drastic change.

The Sen Trend method, a non-parametric trend test method proposed by Sen (1968), was used to measure the change trend at the county level. The advantages of Sen Trend analysis are that it does not require the sample to obey a certain distribution, and that it is not influenced by outlier data or measurement errors (Sun et al. 2015). The Sen Trend equation is as follows:

where β is the trend of winter rape areas, i and j indicate the interval of the time series, and xiand xjindicate the areas for years i and j. If β is greater than 0, there is an increasing trend in the time-series, and vice versa.

The Mann-Kendall (Mann 1945; Kendall 1948) test was performed to test the significance level of trend absence in the time-series data. The null hypothesis for these tests is that there is no trend in the series. The alpha was set to 0.05. If the P-value of a test is less than alpha, then the null hypothesis is rejected. If the P-value is greater than alpha, then there is an insufficient evidence to reject the null hypothesis.

4. Results

4.1. Accuracy validation

Calibration of the samplesSeveral studies have pointed out that the main limitation of times-series MODIS data for crop mapping is the availability of high-quality training data calibrated by the model (Chang et al. 2007). We first compared the winter rape samples with high-resolution images through visual interpretation. We found that the winter rape samples (Fig. 7-B) were very consistent with the winter rape distributions in true-color composite remote sensing images (Fig. 7-A).

We also evaluated the accuracy of winter rape samples using Google Earth images (1 m) as ground reference.Since the Google Earth historical images covering the flowering period of winter rape in the core area of the research area were available only for March 25, 2015, and fortunately Landsat, HJ-1 and GF-1 data were also available for March 2015, all the sample data in the overlapped windows were selected to evaluate the accuracy of winter rape samples quantitatively.

We delineated the winter rape in the Google Earth image(coverage approximately 50 km2) to collect the ground reference. The vector map was then converted to a raster image with resolution matching the high-resolution images.The accuracies for the three data sources are given in Table 3.

This calibration indicates that the samples we derived from high-resolution images are reliable and can be applied to regional-scale mapping.

Validation of the resultsA comparison of winter rape samples derived from the GF-1 image and the ANN modeled abundance image is shown in Fig. 8 (taking 2015 data as an example). From a visual comparison, the distributions of winter rape in these two images have good spatial consistency.

Correlation analysis was conducted to validate the accuracies of the ANN modeled results. The samples derived from high-resolution images at five-year intervals(2000-Landsat TM, 2005-Landsat ETM, 2010-HJ-1A, and 2015-GF-1) were used as the reference data (see Fig. 9).The results showed that the correlation was very high,with correlation coefficients of 0.85, 0.86, 0.93 and 0.88,respectively. The deviation was very small proving the reliability of the proposed method in sub-pixel winter rape mapping.

4.2. Spatial pattern

The proposed method was applied to all the observation years to obtain the time-series winter rape abundance maps. We computed the mean abundance of all the observation years to get an average abundance map for the JPDLP (Fig. 10-A). The areas with high-level abundance were concentrated on the central and western Jianghan Plain, as well as the northern and western Dongting Lake Plain.

Local spatial autocorrelation analysis of the average winter rape abundance showed that the hot spots included most of the counties on the Jianghan Plain and the northern counties on the Dongting Lake Plain (Fig. 10-B).

Fig. 7 A comparison of high-resolution images and their classification results. A, time series high-resolution images (true color composites). B, winter rape samples.

4.3. Temporal pattern

The change of total planting areasThe total winter rape planting areas on the JPDLP dropped significantly,from 1 262 973 ha in 2000 to 894 339 ha in 2017 (Fig. 11).Furthermore, there are opposite trends on the Jianghan Plain and the Dongting Lake Plain. The winter rape planting areas dropped sharply on the Jianghan Plain, from 999 680 ha in 2000 to 543 831 ha in 2017, with a total decline of about 45% and an average annual decline of 3.52%. However, the winter rape planting areas on the Dongting Lake Plain increased significantly from 263 293 ha in 2000 to 350 508 ha in 2017, with an average annual increase of 1.7%.

Annual change analysisAnnual change analysis included generating yearly change maps, and quantifying the intensity and the proportion of changes of winter rape abundance.

Yearly change maps of winter rape abundance during 2000-2017 are given in Fig. 12. Generally speaking,these changes included winter rape gains and winter rape losses, both covering all of the winter rape extent. Thehigh abundance areas were generally consistent with the frequently changed areas. The gains and losses appeared alternately in two consecutive observation periods. These indicate frequent fluctuations and an unstable state of winter rape production. Winter rape gains and losses were highly concentrated in most observation years, roughly consistent with the boundaries of administrative districts, demonstrating the relationship between the growers’ willingness to plant winter rape and government policy. Winter rape losses dominated the trend after 2012, which can be explained as the impact from international markets.

Table 3 Accuracies of the samples in 2015

We took the core area of Jianghan Plain, Qianjiang County and its adjacent areas, as an example to observe the details of the changes (Fig. 13). The winter rape abundance has dropped sharply since 2000. Those concentrated winter rape areas in the north of Qianjiang and Xiantao in 2000 almost disappeared by 2017.

Fig. 8 A comparison of winter rape distributions from different sources. A, samples derived from GF-1 image. B, artificial neural network (ANN) modeled abundance image.

Fig. 9 Validation of artificial neural network (ANN) modeled results. A, 2000. B, 2005. C, 2010. D, 2015.

We can observe the intensity of changes in winter rape abundance at both pixel and county levels. Generally speaking, the accumulated amount of changes in winter rape abundance was intensive, covering almost all the winter rape extent, especially the high abundance areas.The highest intensity of these changes is mostly observed in the dominant production areas of winter rape, especially the core areas of Jianghan Plain (Fig. 14).

We can also observe the proportion of changes in winter rape abundance (Fig. 15). Overall, these changes were frequent, with about 20-30% of croplands changing their abundance drastically every two consecutive observation years.

Fig. 10 The spatial pattern of winter rape. A, the average winter rape abundance during 2000-2017. B, hot spots analysis of the average winter rape abundance.

The change trend at the county levelWe explored the change trend of winter rape planting areas at the county level using the Sen Trend Test. At the county level, the trend exhibits a distinctive characteristic, and the change trend has obvious regional differentiation. Generally speaking,the counties with a significantly decreasing trend mainly cover the northwest of the JPDLP, and the counties with a significantly increasing trend mainly cover the southeast of the JPDLP (Fig. 16-A). It is worth noting that the counties with decreasing trends are concentrated in the traditionally dominant production areas. Statistical tests indicate that the significant changes (with 0.05 significance level at 95%confidence level) mainly exhibit decreasing trends (with the exception of only one county on the Dongting Lake Plain)(Fig. 16-B).

5. Discussion

We can observe significant changes in spatio-temporal patterns of winter rape and distinct regional differentiation of these changes at the county level on the JPDLP during 2000-2017. Firstly, the total winter rape planting areas on the JPDLP dropped significantly, especially on the Jianghan Plain with a decline of 45% during 2000-2017. Secondly,the changes of winter rape abundance were intensive with about 20-30% of croplands changing their abundance drastically in every two consecutive observation years.Thirdly, winter rape has obvious regional differentiation for the change trends at the county level, and the decrease was more pronounced in the traditionally dominant production counties. Finally, the high abundance areas were generally consistent with the frequently changing areas, indicating the vulnerability of winter rape production to economic and environmental factors. However, the intensively changed areas were not strictly consistent with the significantly decreasing areas. This analysis combined with our previous research (Tao et al. 2017a) that demonstrates that cropland abandonment is the main reason for losses in some counties, however, in some other counties, the transformation of winter rape into winter wheat can explain the decreases.

Fig. 12 Yearly change maps of winter rape abundance.

Fig. 13 The details of the changes in the core area of Jianghan Plain, China.

Fig. 14 The intensity of changes in winter rape abundance during 2000-2017. A, at the pixel level. B, at the county level.

Fig. 15 The proportion of changes in winter rape abundance during the observation period of 2000-2017.

Fig. 16 The change trend of winter rape planting areas at the county level during 2000-2017. A, Sen Trend map of the winter rape planting areas at the county level. B, the counties with significant changes at 95% confidence level. Red and green indicate the areas decreasing and increasing drastically.

The fragmented cropland fields and diversity of landscapes in South China posed challenges for largescale crop mapping. The existing research on winter rape mapping has focused mainly on the small or medium scale.Mapping winter rape involves dealing with weak information stored in EVI profiles and has high requirements for the model. ANN is essentially a model-based method and its performance is highly dependent on its inputs. We have made several efforts to optimize the input in order to improve the performance of ANN. Firstly, the EVI profiles of winter rape and winter wheat show a significant difference, and this characteristic offers great potential for distinguishing winter rape. Secondly, reliable winter rape samples were obtained through combining spectral bands and the winter rape index.The winter rape index enhanced the spectral difference between winter rape and winter wheat, suppressing the uncertainty caused by regional differences in phenology.Finally, taking advantage of the self-learning optimal solution of ANN, we built a non-linear mapping between winter rape abundance and the EVI profiles. This method of combining medium-resolution and high-resolution data can be applied for winter rape mapping at a large scale. The proposed method offered high accuracies, which were between 0.85 and 0.93 against the validation samples.

The uncertainties of the proposed method can be explained in two ways. Firstly, much noise and many biases remain in the reconstructed time-series EVI data.Secondly, high-quality winter rape samples are not readily available due to the revisitation cycle of the satellite and cloud cover. Acquiring high-resolution sample data is a key step for applying the proposed method successfully.Due to the scale effect of spatial resolutions spanning from 1 m to tens of meters, the accuracy of winter rape samples against Google Earth images is not very high. However,these samples are sufficient for regional scale mapping when aggregating to MODIS-like resolution.

6. Conclusion

Spatio-temporal dynamics of winter rape can provide relevant information for agricultural production and decisionmaking policies, as well as the sustainable development of the cooking oil market. This paper reports our work on exploring the spatio-temporal patterns of winter rape on the JPDLP of China. An ANN-based classification method fusing time-series MODIS and high-resolution remote sensing images was proposed to map fractional winter rape and the spatial and temporal patterns of winter rape from 2000 to 2017 on the JPDLP were analyzed. We found essential declines of winter rape planting areas on the JPDLP and distinct regional differentiation of these changes at the county level. We also found that an ANN method combining time-series coarse resolution images and multi-source high-resolution images (Landsat, HJ-1 and GF-1, etc.) has the ability to map large-scale winter rape at the sub-pixel level.

The novelty and contributions of our work are summarized as follows. Firstly, this is the first work to explore the spatio-temporal patterns of winter rape in the middle reaches of Yangtze River Valley of China. Secondly, the proposed method has the superiority of mapping winter rape accurately at the sub-pixel level and at the regional scale. This solution can take full advantage of the accuracy of the high-resolution image and the wide coverage and high frequency of the coarse resolution images. Finally,our method deals with the problem of mixed pixels on the Yangtze River Valley to some extent. The ANN output is the winter rape abundance maps, which can accurately describe the spatial distribution of winter rape at the subpixel level.

Acknowledgements

This work was supported by the Natural Science Foundation of Hubei Province, China (2017CFB434), the National Natural Science Foundation of China (41506208 and 61501200) and the Basic Research Funds for Yellow River Institute of Hydraulic Research, China (HKYJBYW-2016-06).