Extraction and spatio-temporal analysis of county-level crop planting patterns based on HJ-1 CCD

Zhang Xiaochun, Cao Zequn, Yang Dan, Wang Qiuhao, Wang Xiugui, Xiong Qinxue

Extraction and spatio-temporal analysis of county-level crop planting patterns based on HJ-1 CCD

Zhang Xiaochun1, Cao Zequn1, Yang Dan3, Wang Qiuhao1, Wang Xiugui1, Xiong Qinxue2

(1.,,430072,; 2.,,434025,;3..,.,323000,)

Crop-type classification and spatio-temporal change detection support rational agricultural management; perennial crop maps reflect changes in crop planting patterns and are useful for economic and social analysis. The satellite images used for mapping crops, however, are not at high spatial and high temporal resolution. They do not provide sufficient data for all stages of growth when mapping spatial distributions of crops in areas with a great variety of agricultural practices and products. A cost-effective solution using short repeat cycle Huan Jing (HJ)-1 Charge-Coupled Device (CCD) imagery and the freely available Landsat-8 imagery was proposed to produce annual crop maps reflecting spatio-temporal changes in planting areas in Jianli County, China. Phenological metrics such as maximum Normalized Difference Vegetation Index (NDVI) values, dates, and the number of days in the growth stages of the different crops were defined from time-series NDVI curves and used for crop classification. Typical planting areas were extracted from the 15 m pan-sharpened Landsat-8 images. The NDVI and time thresholds for phenological metrics were obtained from the NDVI time series curves during the crop growth stage of typical planting areas. Classification rules were established to create crop maps from 2009 to 2016 and applied land-use changes for different crops based on multi-year crop classification maps reflecting the distribution of dominant crops. The high-spatial-resolution China satellite images and crop area data from Jianli statistical yearbook were used to perform an accurate assessment. The average classification accuracy rate was 84% when compared with the high-spatial-resolution imagery, and the classification area matched up to 81.60% with the statistical crop area data. These results indicated that this method provided a possible means for classification permitting regular mapping of crop distributions in complex areas like Jianli County. By the spatio-temporal analysis of summer-harvest crops, it could be found that fewer farmers were willing to plant oilseed rape because of the high labor cost caused by this crop’s low agricultural mechanization level. For autumn-harvest crops, the government set the standard of the lowest purchase price for the middle-season rice, which greatly reduced the risk of planting the middle-season rice for farmers. This might guide farmers’ decisions and lead to small changes in the middle-season rice area. Meteorological data indicated that it had continuous rainfall and waterlogging in the summer of 2016, which led to the reduction of the middle-season rice area, and especially the big reduction of cotton area.

crops; remote sensing; decision trees; NDVI; HJ-1 CCD data; spatio-temporal analysis

0 Introduction

One objective set by the United Nations’ Sustainable Development Goals is to “end hunger, achieve food security and improved nutrition and promote sustainable agriculture” by 2030[1]. To reach this goal, government managers must know in detail where and when crops are grown[2]. Accurate and timely information about crop areas provides reliable data for crop yield forecasting, and optimal management of water resources in the growing areas[3-4]. Users can quickly extract crop area and distribution information by applying remote sensing technologies, and remote sensing images have become a critical data source for thematic maps of regional crop distributions[5-7]. In China, however, the crop planting structures are complicated by year-to-year variation in the area and types of crops. Therefore, it is difficult to map periodically cropland by the method of remote sensing classification.

The attributes of multi-scale and multi-temporal remote sensing data have been widely used for creating crop classification maps. There have been many studies on crop area extraction based on Moderate-resolution Imaging Spectroradiometer (MODIS) data[8]. Brown et al.[9]used MODIS data to carry out research on multi-year agricultural land classification in the Mato Grosso region of Brazil. Usman et al.[10]differentiated wheat and rice by identifying NDVI temporal profiles from MODIS data. Chen et al.[11]used MODIS NDVI time-series data and field survey data to identify soy, cotton, and maize cropping patterns. The spatial resolution of MODIS data is low, however, and only applicable to the extraction of crop information in large-scale regions; they are not suitable for extracting cropping patterns in small fragmented regions.

Remote sensed satellite sensors cannot have both high spatial and high temporal resolution, due to sensor material and satellite orbit parameters. Landsat-8, GaoFen-1 satellite (GF-1) and Systeme Probatoire d’Observation de la Terre (SPOT) provide mid-to-high spatial resolution images for detailed crop classification of small areas[12-13]. Yang et al.[14]completed a field-based rice classification in Wuhua County of China by integrating multi-temporal Sentinel-1A and Landsat-8 data. The repeat period for many common satellites providing mid-to-high resolution remote sensing images, however, is relatively long, making these less useful for crop classification on the county scale in China. For example, the Landsat-8 satellite is an Operational Land Image (OLI) sensor with a revisiting cycle of 16 d. Moreover, higher resolution satellite images have generally larger repeat cycles than the Landsat-8 images, so they are still insufficient for the growing season of crops. These images cannot fully meet the requirement of crop identification in time information[15]. We need sufficient data sources in the whole growth stage of different crops for exact crop spatial distribution, but for China, there is an alternative source for such data.

The Chinese HJ satellite has a wide of 700 km, and a short two days revisit period due to the HJ-1A/1B two-camera network. The HJ images have a 30 m resolution at the pixel level and are available free online, which makes them suitable for exactly extracting crop area information[16]. Wei et al.[17]extracted rice coverage area based on the HJ satellite data. Ge et al.[18]extracted the area of winter wheat by NDVI density slicing using the HJ-1A CCD image. Liu et al.[16]extracted the planting area in Wudi County using a Support Vector Machine (SVM) to classify Enhanced Vegetation Index (EVI) data. These studies have demonstrated that the HJ-1 CCD data was feasible and effective for extracting crop area information in the small and fragmented area, but in this study, only one specific type was extracted from remote sensing images. A more appropriate classification method identifying multiple kinds of crops based on remote sensing techniques needed to be found.

The decision-tree classification method is appropriate for extracting crop areas, and does not require a large amount of sample data when constructing decision tree rules. Decision tree classification makes full use of the advantages of long-term sequence images by multi-threshold limitation. Hermosilla et al.[19]used decision tree completing automatic building detection approaches combining high-resolution images and Light Detection and Ranging (LIDAR) data. Souza et al.[20]mapped and assessed the ten-year Landsat classification of deforestation and forest degradation in the Brazilian Amazon combing decision tree classification and normalized difference fraction index. The decision tree classification based on multi-temporal and multi-feature data is more efficient than traditional classification methods[21-22], however, how to set up the decision-tree classification rule is another problem.

The classification rules for a decision tree could be derived from phenological metrics that include some specific values, such as maximum NDVI values, dates and number of days in the growth stages of different crops. Zhong et al.[23]extracted phenology metrics from NDVI profiles and identified crop types based on phenological metrics using decision trees. Zhang et al.[24]identified one year’s crop type by extracting phenological metrics from NDVI curves and setting up classification rules. Wardlow et al[25]. completed crop classification maps, combining the NDVI time-series data and the phenological characteristics of different crops. Fan et al.[26]built a time-series curve based on the filtered smoothed NDVI and analyzed the spatial pattern of the crop planting system in China. These researchers obtained crop classification results at higher accuracy but lacked crop area information for successive years. Perennial crop classification maps support analysis of the causes of change in plant patterns and are applied to regional crop structure adjustment. However, the method of extracting perennial crop maps by using phenological metrics requires a large ground reference dataset for detecting thresholds in the decision tree[27]. Therefore, it is difficult to obtain accurate, timely, and periodical information about crop areas in successive years without labor-intensive and time-consuming field surveys.

This study aimed to develop a new method to identify crop types within small farming parcels and monitor spatio-temporal changes in these planting patterns. HJ-1 CCD time-series NDVI curves were used to derive the seasonal change rules that reflected the temporal characteristics of NDVI value in crop growth cycles. Pan-sharpened 15 m Landsat-8 images were used for determining typical planting areas. The thresholds for phenological metrics were obtained from the time-series NDVI curves in typical planting areas. A decision tree was built to establish classification rules using the thresholds for producing crop maps from 2009 to 2016. The spatio-temporal changes were analyzed by comparing perennial crop classification maps.

1 Materials and methods

1.1 Overview of the study area

The study area is Jianli County (112°35’E-113°19’E, 29°26’N-30°12’N) in the middle reach of the Yangtze River (Fig.1). Hubei province is located in southern China. As could be seen from Fig.1, Jianli County is in the southern part of Hubei Province. Jianli County as the representative of China’s southern counties is a typical agriculture-based region. The County is characterized by the flat terrain and sufficient sunlight. Annual precipitation is 1100 to 1300 mm, and over 80% of the precipitation occurs between April and October. Moreover, the amount of solar radiation from April to October is usually 75% of the radiation for the whole year in Jianli County, with temperatures equal or greater than 10°C for about 80% of the year. The terrain and climatic conditions are suitable for crop growth and development. The main land-use type in Jianli County was the cropland planted with oil-seed rape, wheat, cotton and rice. Given its agricultural natural environment, we selected Jianli County as the study area to explore a reliable method for regular crop mapping using HJ-1 CCD data.

1.2 Data acquisition and processing

1.2.1 Acquisition of HJ-1 CCD data

HJ-1A/B/C is a small satellite constellation designed specifically for environmental and disaster monitoring and forecasting. The A/B/C system includes two optical satellites HJ-1A/B and one radar satellite HJ-1C. The HJ-1A and HJ-1B satellites are equipped with CCD scanners. The HJ-1 CCD scanners execute scanning and imaging for earth observations with a swath width of 700 km, a ground pixel resolution of 30 m, with 4 spectrum bands. Jianli County is included in the HJ-1A and HJ-1B path entirely, while the repeat cycle is short enough to provide images of the study area for all twelve months of the year.

The cloud-free HJ-1 CCD images downloaded from China Center for Resources Satellite Data and Application (CRESDA) (http://www.cresda.com) covered the whole growth period of autumn-harvest and summer-harvest crops. The dates for the selected images used in crop classification from 2009 to 2016 were listed in Table 1.

Table 1 Aquisition time of HJ-1 CCD images

The cells in Table 1 showed the specific days of each month and year, each row expressed the month, and each column was the year. Table 1 showed that the number of selected dates in 2010 and 2015 was less than in other years, because frequent rainfall led to frequent thick cloud cover during these two years. The images from January to June were selected for the summer-harvest crop identification and the images from June to December were used for producing autumn-harvest crop maps.

1.2.2 Acquisition of Landsat-8 OLI data

Landsat-8 OLI images as well as HJ-1 CCD images were also used for crop classification in this study. The Landsat-8 images downloaded from the website (https://earthexplorer.usgs.gov/) have a spatial resolution of 30 m for multi-spectral bands and 15 m for the panchromatic band 8 with a revisit time of 16 d. In this study, the Landsat-8 OLI images on April 18th, 2016 (path 123 and row 39) and October 2nd, 2016 (path 124 and row 39) were selected for geometric correction with the HJ-1 CCD data. Moreover, these two Landsat-8 OLI images were also chosen for extracting typical planting area of Jianli County, and used for the confirmation of thresholds in the decision-tree classification.

1.2.3 Acquisition of validation images

The high-spatial-resolution validation data used to assess the accuracy of crop classification results were selected from three Chinese Earth observation satellites, the China-Brazil Earth Resources Satellite-1 02C (ZY-1 02C), GF-1 and China-Brazil Earth Resources Satellite-04 (CBERS-04). The ZY-1 02C is equipped with a panchromatic multispectral scanner and a panchromatic high-resolution scanner, which are used to acquire panchromatic and multispectral image data. GF-1 is equipped with a 2 m spatial resolution panchromatic multispectral scanner, 8 m resolution multispectral scanner and four 16 m resolution multispectral scanners. The CBERS-04 is equipped with a multispectral scanner with 5 m/10 m resolution panchromatic (Table 2). The validation images from ZY-1 02C, GF-1 and CBERS-04 satellites were downloaded from the CRESDA website (http://www.cresda.com).

Table 2 Parameters of three satellites with high-spatial-resolution validation images

The joint repeat cycle of three satellites reached about 5 d, and thus the satellites could provide sufficient-date images for validation. Table 3 showed the dates of cloud-free high-spatial-resolution images used for validating crop classification maps. However, the high-spatial- resolution images were only used to validate the classification maps in some areas of Jianli County, as the coverage area of three satellites with the high-spatial- resolution scanners does not include the whole of Jianli County.

Table 3 List of high-spatial-resolution images

As shown in Table 3, it was easy to acquire steadily these high-spatial-resolution images from three satellites after 2013. The images in April or May from 2013 to 2016 were selected for summer-harvest crop mapping, and the images in August or September from 2013 to 2016 were for autumn-harvest crop mapping.

1.2.4 Data processing

After selecting and downloading these satellite images, these images were preprocessed by using the Environment for Visualizing Images (ENVI) software. The preprocessing steps included radiometric calibration, geometric and atmospheric correction, and geo-referencing. They were also pan-sharpened and clipped to the study area. Landsat-8 imagery was used as a reference image for the geometric correction of the HJ-1 CCD, ZY-1 02C, CBERS-04 and GF-1 data. The HJ-1 and validation data were corrected in the image-to-image registration mode and all images were projected to the Universal Transverse Mercator Grid System (UTM) map projection, World Geodetic System (WGS) 84, and Zone 49 North.

1.3 Methods

After downloading and preprocessing the remote sensing images, typical planting areas were selected from Landsat-8 images. The thresholds of phenological metrics were extracted from typical planting areas. Finally, crop types were identified by using the decision-tree classification. Fig.2 showed the overall flow chart of the study.

Time-series NDVI curves of summer-harvest and autumn-harvest crops were established from the cloud-free HJ-1 images and were used to extract the phenological metrics of summer-harvest and autumn-harvest crops. Meanwhile, the typical planting areas of summer-harvest and autumn-harvest crops were acquired from two Landsat-8 images on the dates of April 18th, 2016, and October 2nd, 2016. Thresholds for these phenological metrics were extracted from the time-series NDVI curves of typical planting areas. The classification rules were established in the new decision tree using the thresholds and the classification maps were outputted.

1.3.1 Defining phenological metrics from the time-series NDVI curve

NDVI was an indicator of vegetation growth and vegetation coverage, while phenological metrics were temporal markers that indicate vegetation seasonality[22]. Phenological metrics could be directly derived from NDVI curves. The time-series NDVI curves for summer-harvest and autumn-harvest crops were extracted from cloud-free HJ-1 CCD images, and then phenological metrics for summer-harvest and autumn-harvest crops were defined from the time-series NDVI curves.

1) Two NDVI values were defined as the phenological metrics for summer-harvest crops including wheat and oilseed rape: i) NDVInpis defined as the NDVI value when the first-order derivative is the maximum value in the time-series NDVI curve during the Nutritive growth Period; ii）NDVIrpis defined as the NDVI value when the first-order derivative is the minimum value in the time-series NDVI curve during the Reproductive growth Period. The wheat NDVI curve in 2016 was taken as an example of the summer-harvest phenological metric. As shown in Fig.3, two periods including the nutritive growth period and the reproductive period were chosen for defining summer-harvest phenological metrics. Two metric values including NDVInpand NDVIrpmight represent the growth characteristic of summer-harvest crops.

2)Three phenological metrics were defined for autumn-harvest crop including middle-season rice and cotton.i) DURation (DUR) was defined as the number of days when NDVI was greater than 0.6, and the largest number of days was selected if the duration occurred more than once; ii) The beginning DaTe (DTbeg) was defined as the start date when NDVI was greater than 0.6, corresponding to the jointing stage of crops; iii) The ending DaTe (DTend) was defined as the date of the final stage when NDVI was greater than 0.6, corresponding to the end of the pustulation period.

The autumn-harvest phenological metrics were related to dates and time scales. According to phenological data, the beginning date of middle-season rice was late-June, and cotton was mid-July, while the grass and trees were in early June; the ending date of middle-season rice was in late August while the ending date of grass and trees was early-October. Therefore, these phenological metrics could be used for identifying autumn-harvest crops. The middle-season rice NDVI curve in 2016 was taken as an example of the autumn-harvest phenological metrics (Fig.4).

As shown in Fig.4, the NDVI values of middle-season rice increased from June and achieved the maximum in August. The summer-harvest and autumn-harvest phenological metrics that represented the physiological metrics of different crops were used for the decision-tree classification method.

1.3.2 Extracting the thresholds of phenological metrics from the typical planting area

The thresholds of phenological metrics were used for determining the classification rules in decision trees. These thresholds were extracted from the time-series NDVI curves of typical planting areas. The typical planting areas of summer-harvest and autumn-harvest crops were identified from the 15 m pan-sharpened Landsat-8 images in 2016 by the visual interpretation method.

The summer-harvest typical areas were identified from the false-color image composited from bands 6, 5, and 2 of the Landsat-8 image on April 18th, 2016. Given the spectral characteristics of wheat and rape, the oilseed rape in the flowering phase showed a yellow color, but the wheat still showed a green color during this period. The autumn-harvest typical planting areas were extracted from the false-color image composited from bands 6, 5, and 2 of the Landsat-8 image on October 2nd, 2016. The cotton in the opening phase showed a dark green color, but the middle-season rice still showed a green color during this period.

These typical planting areas were only identified from Landsat-8 images in 2016. The shape of yearly HJ-1 CCD time-series NDVI curves in 2009, 2010, 2011, 2012, 2013, 2014, 2015, 2017, and 2018 were compared with the curve shape in 2016 in the selected typical crop areas. The shape of the NDVI curves was similar to the curve of 2016. This comparison result indicated that the crop type in the typical planting areas for extracting the thresholds remained unchanged from 2009 to 2018. Therefore, the typical planting area identified from the curves in 2016 could still be used for extracting the thresholds from 2009 to 2015.

1.3.3 Crop classification

Crop classification maps were extracted using the decision-tree method. The decision-tree classification rules were established by using the thresholds of phenological metrics from the typical planting area. The crops which were distinguished for the summer-harvest crop map included oil-seed rape and cotton, and the crops which were distinguished for the autumn-harvest crop map included cotton and middle-season rice.

1) Summer-harvest crop classification

In the summer-harvest crop classification, two thresholds of the NDVInpphenological metric and one threshold of the NDVIrpphenological were used for establishing the decision-tree rule. The thresholds of the phenological metric in 2016 were taken as examples for the introduction of summer-harvest crop classification, and the thresholds for other years were listed in Table 4. NDVInp1phenological metric value differentiated between vegetation and non-vegetation covered areas through the value 0.2. NDVInp2metric value distinguished water bodies from non-vegetation using the value 0 and remained unchanged in all years from 2009 to 2016. The maturity period of wheat in Jianli County was in late May, while the maturity period of oilseed rape was in early May, which came from the phenological statistical data in the Jianli statistical yearbook[28]. Therefore, the NDVIrpphenological metric value of oilseed rape was less than wheat from the Heading period to the Maturity period. Wheat and oilseed rape could be differentiated according to the NDVIrpphenological metric by using the threshold 0.4. Based on these thresholds of phenological metrics in 2016, the summer-harvest crop classification rules in the decision tree were designed as following:

If NDVInp1> 0.2 and NDVIrp> 0.4, then wheat; If NDVInp1> 0.2 and NDVIrp< 0.4, then oil-seed rape; If NDVInp1< 0.2 and NDVInp2< 0, then water; If NDVInp1< 0.2 and NDVInp2> 0, then others.

2) Autumn-harvest crop classification

In the autumn-harvest crop classification, the thresholds of the DUR, DTbeg, and DTendphenological metrics were used for establishing the decision-tree rule. The thresholds of the phenological metric in 2016 were taken as examples for the introduction of autumn-harvest crop classification, and the thresholds for other years were listed in Table 5. The thresholds 120 (the early DTbeg), 160 (the late DTbeg), 240 (DTendof cotton), 213 (DTendof middle-season rice), 100 (DUR for cotton), and 70 (DUR for middle-season rice) of phenological metrics in 2016 were taken as examples. If the beginning date was from the 120thday to 160thday in 2016, the ending date was after the 213thday in 2016, and the number of duration days was more than 70, then the type of crop in the corresponding grid point was middle-season rice. If the beginning date was after the 160thday in 2016, the ending date was after the 240thday, and the number of duration days was less than 100, then the type of crop in the corresponding grid point was cotton. These thresholds for the decision-tree classification every year were extracted from the annual time-series NDVI curve of typical planting areas. Based on these thresholds of phenological metrics in 2016, the autumn-harvest crop classification rules in the decision tree were designed as following:

If 120 213 and DUR > 70, then middle-season rice; If DTbeg> 160, DTend> 240 and DUR < 100, then cotton.

1.4 Validation method

Two methods were used for validating the classification results, a comparison with statistical data and with high-spatial-resolution images. The first validation method was that the classification planting area results were compared with the statistical area in the Jianli County yearbook from 2009 to 2016. The delta value percentage between the classification planting area and the statistical area was used to evaluate the classification results against the statistical data. The second method was that high-spatial-resolution images were used to assess the accuracy of the classification results from HJ-1 CCD images.

The high spatial satellite images for validating the HJ-1 CCD classification maps were classified by using a supervised classification based on the SVM tool in the ENVI software. The maps were extracted from the sensors of ZY-1 02C, GF-1, and CBERS-4 images, which had high-spatial-resolution and multispectral images for greater legibility and thus sharper recognition of agricultural areas. The dates of these images were the optimal phenological period for crop planting structure identification when crop spectral characteristics varied greatly. Therefore, by using the SVM supervised classification, the classification results were extracted from 10 m ZY-1 02C images, 8 m GF-1 images, and 10 m CBERS-04 images, which were reliable and suitable for validating the classification maps from 30 m HJ-1 CCD images.

The visual interpretation was used for the supervised classification of the high spatial satellite images. For the summer-harvest crop map validation, given the spectral characteristics of wheat and rape, the oilseed rape in the flowering phase showed a yellow color, but the wheat still showed a green color from the beginning of March to the beginning of April. Yellow areas were selected as the oilseed rape training samples and green areas as the wheat training samples from the true color images with high spatial resolution. Then the spatial distribution of oilseed rape and wheat was extracted for map validation. For the autumn-harvest crop map validation, the high-spatial-resolution images were selected in September when the cotton crop was in the picking period and the middle-season rice crop was in the maturity period. The cotton crop in this period showed dark green and the middle-season rice showed yellowish in the true-color images with high spatial resolution. Although trees and grass might show as green in this period, the fields of cotton and middle-season rice had distinct boundaries, besides, the boundary of tree and grass area was not clear. Therefore, dark green areas were selected as the training samples of cotton and the yellowish area as the training samples of middle-season rice.

The ENVI software was used to calculate similarity metrics between the high-resolution image inversion results and the HJ satellite inversion results by two digital evaluation indicators including the kappa coefficient and overall accuracy. The kappa coefficient measured the agreement between the classification and truth values (the results from the high spatial satellite were considered the ‘truth’ values in this study[29]. The overall accuracy was about what proportion was mapped correctly out of all of the reference sites.

2 Results and analysis

2.1 Thresholds of phenological metrics for mapping crops

The typical planting areas of the 15 m pan-sharpened Landsat-8 images and the annual time-series NDVI curves were used to determine the classification thresholds for mapping yearly summer-harvest and autumn-harvest crops from 2009 to 2016. The thresholds for summer-harvest crops from 2009 to 2016 were shown in Table 4. The thresholds for autumn-harvest crops from 2009 to 2016 were shown in Table 5.

Table 4 Thresholds of phenological metrics for mapping summer-harvest crops from 2009 to 2016

Note: NDVInp1was defined as the NDVI value when the first-order derivative was the maximum value; NDVIrpwas defined as the NDVI value when the first-order derivative was the minimum value.

Table 5 Thresholds of phenological metrics for mapping autumn-harvest crops from 2009 to 2016

Note: DUR was defined as the number of days when NDVI was greater than 0.6; DTbegwas defined as the start date when NDVI was greater than 0.6; DTendwas defined as the date of the final stage when NDVI was greater than 0.6.

The thresholds listed in Tables 4 and 5 were used in the yearly summer-harvest and autumn-harvest decision-tree for crop classification. The thresholds every year were different, because the yearly phenological metrics for both summer-harvest and autumn-harvest crops were characteristic and different.

2.2 Mapping results and spatio-temporal analysis in crop planting patterns

2.2.1 Mapping results of summer-harvest crops

The classification maps for summer-harvest crops from 2009 to 2016 were showed in Fig.5. There were 8 images of summer-harvest classification results, and each image showed the yearly spatial distributions of summer-harvest crops. Light grey grid points represented oilseed rape, dark grey points were wheat and black points represented water bodies in Fig.5.

As could be seen from Fig.5, wheat was planted in the northern part of Jianli County and the oilseed rape was distributed in the southern part of the county from 2009 to 2016. The crop especially wheat was planted in the northwest part of Jianli County from 2009 to 2012. Wheat was also cultivated in northeast areas and still planted in small areas in the west from 2013 to 2016. Planting probability maps and the summer-harvest crop area were extracted from summer-harvest crop maps in Fig.5 by using the ENVI software to further acquire the spatio-temporal distribution information of summer-harvest crops.

2.2.2 The spatio-temporal analysis in the planting pattern of summer-harvest crops

Two planting probability maps of summer-harvest crops in Fig.6 were used for the spatial analysis of oilseed rape and wheat. One map was for the oilseed rape, and the other was for wheat. From Fig.6a and Fig.6b, it could be found that the planting probability of oilseed rape was much larger than wheat. Fig.6b showed that wheat was distributed mostly in northern and western parts of Jianli County, and the planting probability of wheat in other parts was very low. The reason for low planting probability of wheat is that planting wheat is time-consuming and laborious and farmers are unwilling to cultivate wheat.

Moreover, for the further spatio-temporal analysis in the planting pattern of summer-harvest crops, the area of oilseed rape and wheat from 2009 to 2016 was further extracted from eight classification maps in Fig.5 and the results were shown in Table 6. The extracted wheat area increased from 14.57×103hm2in 2009 to 15.8×103hm2in 2016. The wheat planting area from 2009 to 2016 increased slowly with small fluctuations because of climate and other influential factors. Meanwhile, the government-released data indicated that the prices of wheat increased from 2009 to 2014, and the prices of wheat remained stable from 2014 to 2016[30]. The area of crops, which farmers cultivate, was closely related to the price of the corresponding crops because farmers aimed to maximize profit, and how to select the crops to plant was according to the profit from crops. Therefore, farmers increased generally the planting area of wheat to gain greater benefits. It could also be seen from Table 6 that the planting area of oilseed rape increased from 91.6×103hm2in 2009 to 109.57×103hm2in 2012, but the area was reduced to 102.03×103hm2in 2015. The national temporary reserve purchase policy for oilseed rape was abolished in 2015, which resulted in the plunging price of oilseed rape, so the planting area of oilseed rape decreased in 2015[31].

Table 6 Extracted area of summer-harvest crops by remote sensing

The investigation for the study area indicated that the farmer’s dispute about planting wheat or oilseed rape was hot in Jianli County. There was a tendency to “dispose of oilseed and grow wheat”. Oilseed rape cultivation is labor-intensive and time-consuming with a low mechanization rate. Moreover, the economic benefit of planting oilseed rape was lower than planting wheat. With the young and middle-aged former farmers increasingly working in cities, more and more land was uncultivated in winter, so fewer farmers were willing to plant oilseed rape because of the high labor cost caused by this crop’s low agricultural mechanization level.

2.2.3 Mapping results of autumn-harvest crops

The classification maps for autumn-harvest crops from 2009 to 2016 were showed in Fig.7. There are eight images of autumn-harvest classification results, and each map showed the yearly spatial distributions of autumn-harvest crops. Light grey points represented middle-season rice, dark grey grid points were cotton and black points represented water bodies in Fig.7.

As could be seen from Fig.7, cotton was planted in the northern and the western parts of Jianli County while the middle-season rice was in the central parts of the county from 2009 to 2016. The crop especially cotton was mainly planted in the northern parts of Jianli County in 2015 and scattered in Jianli County in 2016. Planting probability maps and the autumn-harvest crop area were extracted from autumn-harvest crop maps in Fig.7 by using the ENVI software to further acquire the spatio-temporal distribution information of autumn-harvest crops.

2.2.4 The spatio-temporal analysis in the planting pattern of autumn-harvest crops

Two planting probability maps of autumn-harvest crops in Fig.8 were used for the spatial analysis of middle-season rice and cotton. One map was for the middle-season rice, and the other was for cotton. From Fig.8a and Fig.8b, we could find that the planting probability of middle-season rice was much larger than the cotton crop. Fig.8a showed that the middle-season rice crop was distributed mostly in the western part of Jianli County along the Yangtze River. The reason caused by the above phenomenon might be that farmers wanted to divert water from the Yangtze River for irrigation. It was also shown in Fig.8a and Fig.8b that the middle-season rice and cotton were planted in many same sites, and the reason might be that farmers sought profits and cultivated cotton as an economical crop in the same land parcel after the middle-season rice was harvested.

Moreover,the area of autumn-harvest crops from 2009 to 2016 was further extracted from eight classification maps in Fig.7, for the further spatio-temporal analysis in the planting pattern of autumn-harvest crops, and the results were shown in Table 7. The cotton planting area increased from 19.10×103hm2in 2009 to 25.97×103hm2in 2011. The cotton planting area was down to 11.85×103hm2in 2016. The government released data indicated that cotton prices rose from 2009 to 2011, stabilized from 2011 to 2014, and decreased obviously after 2014[32]. Changes in cotton areas were consistent with cotton price fluctuations. It was also concluded that profits drove farmers to decide which crop should be cultivated.

Table 7 Extracted area of autumn-harvest crops by remote sensing

It could also be found that the planting area of middle-season rice was on the increase from 2009 to 2013 and then decreased slowly after 2013. The government-released data indicated that the net profit of rice showed a downward trend from 2009 to 2015, but middle-season rice’s profit declined slowly compared with other types of rice[33]. The reason for this slow decline might be that the government set the standard of the lowest purchase price for the middle-season rice, which greatly reduced the risk of planting middle-season rice for farmers. This might guide farmers’ decisions and lead to small changes in the middle-season rice area.

From Table 7, we can compare the two-year area variation of cotton with middle-season rice. The area of cotton was down to 11.85×103hm2in 2016 from 15.86×103hm2in 2015, and the area of middle-season rice was down to 75.99×103hm2in 2016 from 77.90×103hm2in 2015. It could be concluded that the cotton area in 2016 was reduced by approximately 33% compared with the area in 2015, and this was bigger than middle-season rice. Meteorological data indicated that it had continuous rainfall and waterlogging in the summer of 2016, which led to the reduction of the middle-season rice area, and especially the big reduction of cotton.

2.3 Validation and discussion

2.3.1 Validation by statistical data

To validate and evaluate the accuracy of the classification method in this study, statistical data from the Statistical Yearbook of Jianli County was compared with the area data extracted from remote sensing images (Fig.9 and Table 8). Every figure expressed the area comparison of a specific crop. The statistical area data were expressed by the line of diamond points, and the extracted data was the line of square points, as shown in Fig.9.

The classification accuracy was also assessed by comparing the statistical data and the extracted data by remote sensing, as shown in Table 8 and Fig.9. As can be seen from Table 8, the average precision of wheat, oilseed rape, middle-season rice, and cotton arrived at 80.31%, 71.82%, 90.29% and 83.96%, respectively. The average accuracy rate of autumn-harvest and summer-harvest crops approached 81.60%. It could also be seen from Table 8 that the oilseed rape accuracy rate was around 70% from 2009 to 2016, and the cotton accuracy rate in 2014 and 2015 was about 65%. One reason might be that mixed pixels on remote sensing images resulted in poor recognition. Another reason was that the protective woods around the farmland located at the edge of the field might be misclassified to cotton or oilseed rape types.

Table 8 Classification accuracy assessment and validation by statistical data

As were shown in Fig.9 and Table 8, the extracted area of oilseed rape and middle-season rice was much higher than the statistical data and the extracted area of wheat was less than the statistical data. This classification results are affected by the following factors. The local crop plots were broken and mixed on a 30-meter resolution HJ image. Thus, the mixed pixels affected the accuracy of extracted maps about summer-harvest and autumn-harvest crops. Moreover, in some towns there may be deviations in the statistical values for middle-season rice and cotton and the statistical data overestimated the area of wheat, to some extent.

2.3.2 Validation by high-spatial-resolution images

Except for the validation by statistical data, the validation of the HJ-1 CCD classification maps based on high-spatial-resolution images were also presented in Fig.10 for summer-harvest map and Fig.11 for autumn-harvest map. The validation area was a part of Jianli County and the validation date of high-spatial-resolution images was specific, because the high-spatial-resolution satellites did not provide both high spatial and temporal resolution images, and the clear images only covered part of Jianli County.

In Fig.10 and Fig.11, the classification maps on the left were from HJ-1 CCD images, and the right validation maps were from high-spatial-resolution images. In Fig.10, the summer-harvest maps from HJ-1 CCD were validated by the ZY-1 02C image on Apr. 10th, 2013, the GF-1 image on Apr. 27th, 2014, the CBERS-04 image on May 9th, 2015, and the GF-1 image on Apr. 30th, 2016. In Fig.11, the autumn-harvest maps from HJ-1 CCD were validated by the ZY-1 02C image on Sep. 22nd, 2013, the GF-1 image on Sep. 22nd, 2014, the ZY-1 02C image on Sep. 16th, 2015, and the GF-1 image on Aug. 27th, 2016. A visual comparison showed that the HJ-1 CCD map results were consistent with the validation maps from the high-spatial-resolution images. To analyze the classification accuracy using high-spatial-resolution images further, the Kappa coefficients and overall accuracy values were used as the measure of agreement between the classification results and the validation maps, which were shown in Table 9.

It could be inferred from Table 9 that the spatial distributions of summer-harvest and autumn-harvest crops from HJ-1 CCD were consistent with the maps derived from the high-spatial-resolution images. The maximum accuracy rate arrived at 97.22% and the average accuracy rate reached 84%. The Kappa coefficient of autumn-harvest crop classification was only 0.50, and the reason might be that the HJ-1 CCD images used for the autumn-harvest crop map were not clear and covered by light clouds. In general, the classification method in this study was appropriately used to map crop spatial distribution on the county scale.

Table 9 Kappa coefficient and overall accuracy rate for HJ-1 CCD classification maps using high-spatial-resolution images

2.3.3 Discussion

Annual HJ-1 images were used to construct a time-series NDVI curve. The key metrics were extracted from the time-series NDVI curves for determining the threshold in the decision tree classification. The results demonstrated that different crop types could be identified from HJ-1 CCD images according to the time-series NDVI curve. Crop maps were effectively extracted with high overall accuracy, verifying the technical feasibility of crop classification using time-series remote sensing images. This conclusion was consistent with the study from Wardlow et al.[34].

In this study, the land-use change for different crops was evaluated based on multi-year crop classification maps that reflected the spatial distribution of dominant crops. Long time-series crop classification maps reflected the crop area change[35]. The HJ-1A/B CCD images were timely and widely available in China, thus meeting the needs of long time-series crop classification. This study was consistent with another study, which identified paddy rice in Yuanjiang City using time-series HJ-1 CCD imagery[36]. The potential of this method was also verified for the background investigations of crop acreage.

The appropriate selection of thresholds was important for the accuracy improvement of the classified results[37]. Generally, the thresholds in the decision tree were selected based on statistical data[38]. However, the statistical data in most existing crop mapping methods were collected from the field survey for a few years[39]. For crop mapping over many years, the high cost of yearly field survey collection limited the wide application of methods that combine remote sensing and field investigation. Jana et al.[40]generated training data for detecting thresholds used as rules in the decision trees by a random collection of polygons representing visually differing vegetation classes within each NDVI temporal profile. In contrast, in this study, the thresholds in the decision tree were extracted from the yearly time-series NDVI curve in typical planting areas, which was not highly dependent on the labor intensive and time-consuming field surveys. Accordingly, a perennial map of different crop types was also extracted and a small amount of field survey data could get accurate results. However, how to determine exactly thresholds with empirical characteristics was difficult, and the improper thresholds might lead to a decrease in accuracy rate. More scientific and precise methods to select thresholds would be the focus of subsequent research.

3 Conclusions

In this study, autumn-harvest and summer-harvest phenological metrics from the time-series NDVI curve were defined to represent the seasonal features. Typical planting areas of crops were obtained by visually interpreting Landsat 8 images. The thresholds of phenological metrics were extracted from the time-series NDVI curve in typical planting area. The thresholds were combined with the decision tree, and crop distribution maps of summer-harvest and autumn-harvest crops were extracted from 2009 to 2016 in Jianli County.

1) From the inversion crop area by remote sensing, the wheat area increased from 14.57×103hm2in 2009 to 15.80×103hm2in 2016. The oilseed rape area increased from 91.60×103hm2in 2009 to 109.57×103hm2in 2012, but the area was reduced to 102.03×103hm2in 2015. The cotton area increased from 19.10×103hm2in 2009 to 25.97×103hm2in 2011 and was down to 15.86×103hm2in 2015 and 11.85×103hm2in 2016. The middle-season rice area was also down to 75.99×103hm2in 2016 from 77.90×103hm2in 2015. When compared with the high-spatial-resolution imagery, the average classification accuracy rate was 84%, and the classification area matched up to 81.60% with the statistical crop area data.

2) By the spatio-temporal analysis of crops, it could be concluded that agricultural mechanization level, government policies, and meteorological conditions were three main factors influencing the area of summer-harvest and autumn-harvest crops. Fewer farmers were willing to plant oilseed rape because of the high labor cost caused by this crop’s low agricultural mechanization level. The reason for the slow decline of the middle-season rice area might be that the government set the standard of the lowest purchase price for middle-season rice, which greatly reduced the risk of planting middle-season rice for farmers. Meteorological data indicated that it had continuous rainfall and waterlogging in the summer of 2016, which led to the reduction of the middle-season rice area, and especially the big reduction of cotton.

3) Results showed that the main crop types by the crop classification method in this study were identified, and the analysis of spatio-temporal changes in the agriculture patterns at the county level was conducted. Small-scale and mixed-cultivation farmland in southern China restricted the accuracy of classification based on remote sensing data when distinguishing one kind of crop from other crops. Accordingly, this research method is simple and quick, but the classification accuracy needs to be improved, which will be addressed in future research.

[1] United Nations. Transforming our world: The 2030 agenda for sustainable development[EB/OL]. 2015, [2015-01-01]. https:// sustainabledevelopment. un. org/post2015/transformingourworld.

[2] Pringle M J, Michael S, Tindall D R. Multi-decade, multi-sensor time-series modelling-based on geostatistical concepts-to predict broad groups of crops[J]. Remote Sensing of Environment, 2018, 216: 183-200.

[3] Ozdogan M, Woodcock C E. Resolution dependent errors in remote sensing of cultivated areas[J]. Remote Sensing of Environment, 2006, 103(2): 203-217.

[4] 王红营，潘学鹏，罗建美，等. 基于遥感的华北平原农作物时空分布变化特征分析[J]. 中国生态农业学报，2015，23(9)：1199-1209. Wang Hongying, Pan Xuepeng, Luo Jianmei, et al. Using remote sensing to analyze spatiotemporal variations in crop planting in the North China Plain[J]. Chinese Journal of Eco-Agriculture, 2015, 23(9): 1199-1209. (in Chinese with English abstract)

[5] Crapper P. Errors incurred in estimating an area of uniform land cover using Landsat[J]. Photogrammetric Engineering and Remote Sensing. 1980, 46(10): 1295-1301.

[6] Crook F W. Underreporting of China’s cultivated land area: Implications for world agricultural trade[R]. International Agriculture and Trade Rep. No. RS-93-4 China: Situation and Outlook Series, 1993, 33-39.

[7] Mcnairn H, Champagne C, Shang Jiali, et al. Integration of optical and Synthetic Aperture Radar (SAR) imagery for delivering operational annual crop inventories[J]. ISPRS Journal of Photogrammetry & Remote Sensing, 2009, 64(5): 434-449.

[8] 陈颖姝，张晓春，王修贵，等. 基于Landsat8 OLI与MODIS数据的洪涝季节作物种植结构提取[J]. 农业工程学报，2014，30(21)：165-173. Chen Yingshu, Zhang Xiaochun, Wang Xiugui, et al. Extraction of crop planting structure in seasons prone to waterlogging using Landsat8 OLI and MODIS data[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2014, 30(21): 165-173. (in Chinese with English abstract)

[9] Brown J C, Kastens J H, Coutinho A C, et al. Classifying multiyear agricultural land use data from Mato Grosso using time-series MODIS vegetation index data[J]. Remote Sensing of Environment, 2013, 130: 39-50.

[10] Usman M, Liedl R, Shahid M A, et al. Land use/land cover classification and its change detection using multi-temporal MODIS NDVI data[J]. Journal of Geographical Sciences, 2015(12): 1479-1506.

[11] Chen Y L, Lu D S, Huang J F, et al. Mapping croplands, cropping patterns, and crop types using MODIS time-series data[J]. International Journal of Applied Earth Observation and Geoinformation, 2018, 69: 133-147.

[12] Upadhyay P, Kumar A, Roy P S, et al. Effect on specific crop mapping using WorldView-2 multispectral add-on bands: Soft classification approach[J]. Journal of Applied Remote Sensing, 2012, 6(1): 1-13.

[13] Hao P Y, Wang L, Niu Z. Potential of multitemporal Gaofen-1 panchromatic/multispectral images for crop classification: Case study in Xinjiang Uygur Autonomous Region, China[J]. Journal of Applied Remote Sensing, 2015, 9(1): 1-15.

[14] Yang H J, Pan B, Wu W F, et al. Field-based rice classification in Wuhua County through integration of multi-temporal Sentinel-1A and Landsat-8 OLI data[J]. International Journal of Applied Earth Observation & Geoinformation, 2018, 69: 226-236.

[15] 燕荣江，景元书，何馨. 农作物种植面积遥感提取研究进展[J]. 安徽农业科学，2010，38(26)：14767-14768.

[16] 刘庆生，黄翀，刘高焕，等. 基于关键期HJ卫星数据提取无棣县作物种植面积[J]. 中国农学通报，2014，30(26)：284-290. Liu Qingsheng, Huang Chong, Liu Gaohuang, et, al. Planting area extraction of a crop key growth period in Wudi County based on HJ satellite data[J]. Chinese Agricultural Science Bulletin, 2014, 30(26): 284-290. (in Chinese with English abstract)

[17] 魏新彩，王新生，刘海，等. HJ卫星图像水稻种植面积的识别分析[J]. 地球信息科学学报，2012，14(3)：382-388. Wei Xincai, Wang Xinsheng, Liu Hai, et al. Extraction of paddy rice coverage based on the HJ satellite data[J]. Journal of Geo-Information Science, 2012, 14(3): 382-388. (in Chinese with English abstract)

[18] 葛广秀，李卫国，景元书. 基于NDVI密度分割的冬小麦种植面积提取[J]. 麦类作物学报，2014，34(7)：997-1002. Ge Guangxiu, Li Weiguo, Jing Yuanshu. Area of winter wheat extracted on NDVI density slicing[J]. Journal of Triticeae Crops, 2014, 34(7), 997-1002. (in Chinese with English abstract)

[19] Hermosilla T, Ruiz L A, Recio J A, et al. Evaluation of automatic building detection approaches combining high resolution images and LiDAR data[J]. Remote Sensing, 2011, 3(6): 1188-1210.

[20] Souza J C M, Siqueira J V, Sales M H, et al. Ten-year Landsat classification of deforestation and forest degradation in the Brazilian Amazon[J]. Remote Sensing, 2013, 5(11): 5493-5513.

[21] 李敏，赵庚星，蔡明庆，等. 县域棉花信息遥感提取与棉田精确化管理分区研究[J]. 自然资源学报，2012，27(11)：1971-1980. Li Min, Zhao Gengxing, Cai Mingqing, et al. Extraction of cotton information using remote sensing and precision management zoning at county scale[J]. Journal of Natural Resources, 2012, 27(11): 1971-1980. (in Chinese with English abstract)

[22] 张晓忆，李卫国，景元书，等. 多种光谱指标构建决策树的水稻种植面积提取[J]. 江苏农业学报，2016，32(5)：1066-1072. Zhang Xiaoyi, Li Weiguo, Jing Yuanshu, et al. Extraction of paddy rice area by constructing the decision tree with multiple spectral indices[J]. Jiangsu Journal of Agricultural Sciences, 2016, 32(5): 1066-1072. (in Chinese with English abstract)

[23] Zhong L H, Hawkins T, Biging G, et al. A phenology-based approach to map crop types in the San Joaquin Valley, California[J]. International Journal of Remote Sensing, 2011, 32(22): 7777-7804.

[24] Zhang X C, Xiong Q X, Di L P, et al. Phenological metrics-based crop classification using HJ-1 CCD images and Landsat 8 imagery[J]. International Journal of Digital Earth, 2017, 11(12): 1219-1240.

[25] Wardlow B D, Egbert S L. Large-area crop mapping using time-series MODIS 250 m NDVI data: An assessment for the U. S. Central Great Plains[J]. Remote Sensing of Environment, 2008, 112(3): 1096-1116.

[26] 范锦龙，吴炳方. 复种指数遥感监测方法[J]. 遥感学报，2004(6)：628-636. Fan Jinlong, Wu Bingfang. A methodology for retrieving cropping index from NDVI profile[J]. Remote Sensing, 2004, 8(6): 628-636. (in Chinese with English abstract)

[27] Zhang H X, Li Q Z, Liu J G, et al. Crop classification and acreage estimation in North Korea using phenology features[J]. Giscience & Remote Sensing, 2017, 54(3): 1-26.

[28] 监利县人民政府. 监利统计年鉴[Z]. 监利县：监利县统计局，2009-2016.

[29] Cohen J A. Coefficient of agreement for nominal scales[J]. Educational & Psychological Measurement, 1960, 20(1): 37-46.

[30] 马文峰. 2016中国小麦市场分析回顾及2017年展望[J]. 粮食加工，2017，42(4)：5-11.

[31] 崔文超，马文杰. 湖北省油菜托市收购政策取消的影响研究[J]. 湖北农业科学，2017，56(17)：3386-3390. Cui Wenchao, Ma Wenjie. Influence of abolition of rape support and procurement policy in Hubei province[J]. Hubei Agricultural Sciences, 2017, 56(17): 3386-3390. (in Chinese with English abstract)

[32] 李想，孙一铮. 2017年中国棉花市场形势分析及2018年展望[J]. 农业展望，2018，14(4)：4-7. Li Xiang, Sun Yizheng. Chinese cotton market in 2017 and its prospect for 2018[J]. Agricultural Outlook, 2018, 4(12): 4-7. (in Chinese with English abstract)

[33] 周惠，曾晓虹. 2017年国内稻米市场形势分析及2018年展望[J]. 农业展望，2018，14(3)：4-9. Zhou Hui, Zeng Xiaohong. Domestic rice market in 2017 and its prospect for 2018[J]. Agricultural Outlook, 2018, 14(3): 4-9. (in Chinese with English abstract)

[34] Wardlow B D, Egbert S, Kastens J H. Analysis of time-series MODIS 250 m vegetation index data for crop classification in the U. S. Central Great Plains[J]. Remote Sensing of Environment. 2007, 108(3): 290-310.

[35] 潘学鹏，李改欣，刘峰贵，等. 华北平原冬小麦面积遥感提取及时空变化研究[J]. 中国生态农业学报，2015，23(4)：497-505. Pan Xuepeng, Li Gaixin, Liu Fenggui, et al. Using remote sensing to determine spatio-temporal variations in winter wheat growing area in the North China Plain[J]. Chinese Journal of Eco-Agriculture. 2015, 23(4): 497-505. (in Chinese with English abstract)

[36] Zhao J, Huang W J, Zhang Y H, et al. Inversion of leaf area index during different growth stages in winter wheat[J]. Spectroscopy & Spectral Analysis, 2013, 33(9): 2546-2552.

[37] Shahriar P M, Budde M, Rowland J. Mapping irrigated areas in Afghanistan over the past decade using MODIS NDVI[J]. Remote Sensing of Environment, 2014, 149: 155-165.

[38] Wang C Z, Fritschi F B, Stacey G, et al. Phenology-based assessment of perennial energy crops in North American Tallgrass Prairie[J]. Annals of the Association of American Geographers, 2011, 101(4): 742-751.

[39] Zhong L H, Gong P, Biging G S. Efficient corn and soybean mapping with temporal extendability: A multi-year experiment using Landsat imagery[J]. Remote Sensing. of Environment. 2014, 140: 1-13.

[40] Jana E, Christopher C, John L, et al. Reconstructing the spatio-temporal development of irrigation systems in Uzbekistan using Landsat time series[J]. Remote Sensing, 2012, 4(12): 3972-3994.

基于HJ-1 CCD的县域农作物种植结构提取和时空分析

张晓春1，曹泽群1，杨 聃3，王秋豪1，王修贵1，熊勤学2

（1. 武汉大学水资源与水电工程科学国家重点试验室，武汉 430072；2. 长江大学农学院，荆州 434025；3. 国网浙江省电力有限公司紧水滩水力发电厂，丽水 323000）

作物分类和时空变化监测信息可以为农业管理提供依据，多年作物种植结构图反映了作物种植方式的变化，对经济和社会分析起着重要作用。然而，用于绘制作物分布图的卫星影像不能同时具有高时间高空间分辨率，在提取作物种类复杂多样地区的种植结构图时，往往难以提供足够的作物生长周期内影像。该研究提出了一种既经济又高效的解决方案，即利用重复周期短的环境一号CCD（HuanJing-1 Charge-Coupled Device，HJ-1 CCD）图像和免费Landsat-8图像来提取中国监利县的作物种植区时空变化图。根据NDVI时间序列曲线定义了不同作物生育期物候指标例如归一化植被指数（Normalized Difference Vegetation Index，NDVI）的最大值、日期和天数等，用于作物分类。为了获取物候指标的阈值，首先从15m Landsat-8影像中提取典型种植区，然后利用典型种植区作物生长阶段NDVI时间序列曲线，得到物候指标中的NDVI阈值和时间阈值，再根据这些阈值制定了分类规则，并获得了2009—2016年作物分布图。根据多年主要作物分布图，分析不同作物的土地利用变化。最后利用高空间分辨率卫星图像和监利县统计年鉴中的作物面积数据对作物分类结果进行精度评估。与高空间分辨率图像相比，平均分类精度为84%，与统计作物面积数据相比，分类精度达到81.60%。结果表明，该研究为在像监利县这样复杂地区进行常规的作物分布制图提供了一种可行的分类方法。通过对夏收作物的时空动态变化分析可以发现，油菜农业机械化水平低、劳动力成本高，导致愿意种植油菜的农民较少。对于秋收作物，政府设定了中稻最低收购价标准，大大降低了农民种植中稻的风险，对农民种植秋收作物具有指导作用。

作物；遥感；决策树；归一化植被指数；环境一号CCD数据；时空分析

Zhang Xiaochun, Cao Zequn, Yang Dan, et al. Extraction and spatio-temporal analysis of county-level crop planting patterns based on HJ-1 CCD[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(6): 168-181. (in English with Chinese abstract)doi：10.11975/j.issn.1002-6819.2021.06.021 http://www.tcsae.org

张晓春，曹泽群，杨聃，等. 基于HJ-1 CCD的县域农作物种植结构提取和时空分析[J]. 农业工程学报，2021，37(6)：168-181. doi：10.11975/j.issn.1002-6819.2021.06.021 http://www.tcsae.org

2020-08-12

2020-10-20

National Key Research and Development Program of China (2018YFC1508301, 2018YFC1508302); National Natural Science Foundation of China (31871516); Hubei Natural Science Foundation (2019CFB507)

Zhang Xiaochun, PhD, Associate professor, research interests: agricultural remote sensing. Email: xczhang@whu.edu.cn

10.11975/j.issn.1002-6819.2021.06.021

S127

A

1002-6819(2021)-06-0168-14