Monitoring Cropland Abandonment in Southern China from 1992 to 2020 Based on the Combination of Phenological and Time-Series Algorithm Using Landsat Imagery and Google Earth Engine

Abstract

Cropland abandonment is one of the most widespread types of land-use change in Southern China. Quickly and accurately monitoring spatial-temporal patterns of cropland abandonment is crucial for food security and a good ecological balance. There are still enormous challenges in the long-term monitoring of abandoned cropland in cloud and rain-prone and cropland-fragmented regions. In this study, we developed an approach to automatically obtain Landsat imagery for two key phenological periods, rather than as a time series, and mapped annual land cover from 1989 to 2021 based on the random forest classifier. We also proposed an algorithm for pixel-based, long-term annual land cover correction based on prior knowledge and natural laws, and generated cropland abandonment maps for Guangdong Province over the past 30 years. This work was implemented in Google Earth Engine. Accuracy assessment of the annual cropland abandonment maps for every five years during study period revealed an overall accuracy of 92–95%, producer (user) accuracy of 90–96% (73–87%), and Kappa coefficients of 0.81–0.88. In recent decades, the cropland abandonment area was relatively stable, at around 50 × 10⁴ ha, while the abandonment rate gradually increased with a decrease in the cultivated area after 2000. The Landsat-based cropland abandonment monitoring method can be implemented in regions such as southern China, and will support food security and strategies for maintaining ecological balance.

1. Introduction

Cropland abandonment is a widespread phenomenon that can be defined as the process by which arable land that was previously cultivated is gradually restored to natural vegetation, without agricultural management for at least two to five years [1,2,3,4,5]. Since the industrial revolution in the mid-19th century, a large area of cultivated land has been abandoned in Europe and other developed regions worldwide [6,7,8,9,10]. In China, abandonment of farmland has occurred in more than 100 counties due to the rapid economic development and urbanization that started in the 1990s [11,12]. However, cropland abandonment not only has an impact on biodiversity and carbon storage [13,14,15,16,17,18], but leads to waste of cultivated land and decline in agricultural production [7,19,20]. The average arable land area per person in China is only half of the world’s average [21], and the reduction of arable land and continuous increase in cropland abandonment has threatened national food security [22]. There is an urgent need to formulate policies with a scientific basis to address this and ensure social stability. Therefore, it is necessary to monitor cropland abandonment over the long term so as to provide data support for exploring the driving mechanism of cropland abandonment.

Due to the complexity of the cropland abandonment process, it is difficult to map spatial and temporal changes therein using existing abandoned cropland data. For example, abandoned cropland recorded by “on-the-spot photographs” can easily be confused with fallow land, and therefore may not reflect the spatial-temporal pattern of abandonment. Remote sensing-based technology is the most effective and reliable way of mapping abandoned cropland and estimating the rate of abandonment [23,24]. However, as the intermediate stage between cultivated land and forest, abandoned cropland with regenerated grasses or shrubs in different succession phases is similar to grassland [25] and sparse woodland, and it is difficult to differentiate between these land cover types with static images [5,26]. Time series images can capture the dynamic features of different land cover types, and high temporal resolution data [e.g., Moderate Resolution Imaging Spectroradiometer (MODIS)] have been widely applied to map cropland abandonment [25,27,28,29]. However, abandoned cropland in hills and scattered plots is usually too fragmented to be extracted from coarse-resolution (250 m to 1 km) data. Cropland in China is highly fragmented, with an average field size of 1.52 mu (1 mu ≈ 666.667 m²). The particularly small (0.9 mu) fields in hill regions may lead to large classification errors due to the mixed pixels in coarse-resolution data [19,30]. Moreover, MODIS data do not cover the cropland abandonment that occurred in China in the 1980s. Fortunately, the resolution (30 m) and temporal depth of Landsat satellite images overcome the limitations of MODIS, and can thus be used to monitor long-term cropland abandonment at a fine spatial scale [2,31,32].

Landsat imagery has been widely used to map cropland abandonment, but the complex inter- and intra-annual spectral diversity of cropland presents challenges [28,33]. Existing monitoring studies based on Landsat data have mainly focused on two approaches. The first approach is mainly used to detect land cover conversion [34,35,36,37], with the reliability principally depending on the accuracy of annual land cover classifications [23]. However, this approach is usually applied to typical land cover classes and few studies have focused specifically on abandoned cropland. A recent study monitoring cropland abandonment based on annual land cover types considering vegetation characteristic for land cover types. Early successional grass and shrubs can be used to monitor changes from cultivated land to abandoned cropland [2]. However, errors may occur in land cover maps for individual years, including misclassification of grassland and uncultivated fields. Methods that identify cropland abandonment based on annual cropland maps have lower accuracy and long-term monitoring is particularly challenging [33]. The second approach is to map cropland abandonment by monitoring vegetation changes based on time series spectral data, such as the normalized difference vegetation index (NDVI) and bare soil index (BSI) [2,38,39,40,41]. However, there are several limitations to such time series algorithms. First, due to the complexity of crop types in activated croplands, time series spectra vary widely [2]. Therefore, it is difficult to identify the transition to abandoned farmland. Second, dynamic characteristics are significantly affected by missing data in cloudy and rainy areas of southern China, which may lead to large errors in such regions.

Furthermore, the various definitions of cropland abandonment make it difficult to evaluate the reliability of the different approaches to mapping it [2]. Long-term annual land cover classification schemes including uncultivated cropland can be applied to adjust cropland abandonment maps. However, as in annual classification schemes, during the initial vegetation succession stage, as well as during the transition toward woody vegetation after abandonment for several years, abandoned cropland vegetation can easily be confused with cultivated land or woody vegetation. Grassland may also be confused with uncultivated cropland due to their similar phenological characteristics in annual land cover classification schemes, leading to overestimates of abandoned cropland. Moreover, when there are many missing data points or outliers in individual years, land cover types may be erroneously classified. Ideally, cropland abandonment should be mapped over the long term on a year-by-year basis correction according to prior knowledge, natural laws, and clear definitions of land cover types.

Given the need for cropland abandonment information and the difficulty in monitoring it over the long term in cloud-prone and cropland-fragmented regions, this study attempted to develop an approach to map abandoned cropland from 1992 to 2020 in southern China based on Landsat imagery and the Google Earth Engine (GEE). Our goal is to develop an approach for annual land covers classification based on phenology and mapping cropland abandonment with time-series algorithm for annual land covers maps.

2. Materials and Methods

2.1. Study Area

Guangdong Province (Figure 1) is located in the southern part of mainland China (20°13′–25°31′E and 109°39′–117°19′N). It contains 21 cities, has a land area of 179,800 km², and is part of the East Asian monsoon zone (which has central subtropical, southern subtropical, and tropical climates from north to south). Guangdong Province has abundant light, heat, and water resources to meet crop growth requirements in China. However, the landform types are complex and diverse, with 70% of the area being mountainous and hilly, and the cropland being relatively fragmented. According to statistical data for 1990, the area of cropland in Guangdong Province was 252.88 × 10⁴ ha, while in 2019 it was 190.19 × 10⁴ ha; thus, there has been a significant decrease in the area of cropland. Moreover, with the introduction of migrant workers in the 1990s, the abandonment of cropland has increased, which may ultimately affect the stable development of society [5]. Furthermore, existing time-series based cropland abandonment mapping algorithms still face challenge in southern China due to the cloudy and rainy weather there. Therefore, Guangdong Province is an appropriate study area for mapping cropland abandonment over recent decades.

2.2. Data

2.2.1. Landsat 4/5/7/8 Imagery and Pre-Processing

Landsat series satellites with 30-m spatial resolution jointly managed by NASA and USGS are mainly used to detect the resources and environment of earth. For monitoring cropland abandonment in Guangdong Province over the last 30 years, Landsat surface reflectance (SR) data from 1989 to 2021 were collected from GEE, including 14 Landsat-4 scenes from 1989 to 1993, 4734 Landsat-5 scenes from January 1989 to April 2012, 4814 Landsat-7 scenes from May 1999–2021, and 2499 Landsat-8 scenes from March 2013–2021 (Figure 2a). The Scan Line Corrector (SLC) of Landsat-7 malfunctioned in 31 May 2003, resulting in the scanning pattern to record wedge-shaped scan-to-scan gaps along the satellite ground track, obtaining approximately 78% of the data of the normal scene area for acquired images [42].

The Landsat-8 SR dataset was processed to the Level 1 Precision Terrain level for compatibility with Landsat-4/5/7. Image areas of poor quality, including those affected by cloud, cloud shadow, snow/ice, and cirrus, were identified and masked using the FMask algorithm [43]. The average numbers of good-quality observations from 1989 to 2021 (in 0.5° latitude increments) in Guangdong Province are shown in Figure 2b. For the periods 2000–2009 and 2013–2021, there were abundant good-quality observations (average of ~15 scenes for each 0.5° latitude). Good-quality data were least abundant for 2012 (2–4 scenes per 0.5° latitude). As for the pixel-based spatial distribution of good-quality observations number in 2009 (Figure 2c) and 2019 (Figure 2d), there were more than nine good-quality scenes in most areas.

2.2.2. CNLUCC

China Multi-period Land Use Land Cover Change Remote Sensing Monitoring Dataset (CNLUCC) in 1990, 1995, 2000, 2005, 2010, 2015, and 2020 verified by field inspection and manually revised were obtained from the Resource and Environmental Science Data Registration and Publishing System, which were one of the one of the reference bases for expending training samples and verifying our proposed approach.

2.2.3. Field Surveys

Our field surveys for different land covers were conducted with different distribution patterns in Guangdong province in 2019. Field surveys data for “Cultivated field” (53 sample points) randomly distributed in various regions in Guangdong Province due to the complex and diverse crops planted there, while that of “Uncultivated field”, “Woody vegetation”, “Water”, built-up land and bare land of “Other land-cover” (171, 40, 35, 30, and 35 sample points, respectively) were mainly distributed in three prefecture-level cities (Meizhou, Maoming and Zhanjiang) in Guangdong province obtaining from the special survey projects. The field survey data were used to construct the land covers feature library including spectral feature and shape, texture, and color, as well as brightness in the remote sensing images for generating expanded training samples and testing the accuracy of our proposed annual land covers classification method.

2.2.4. The Regions of Interests (ROIs) for Approach Training and Validation from Field Survey and CNLUCC Data

Our method for mapping cropland abandonment was based on five self-generated annual land cover classifications referring to the classification system CNLUCC. “Uncultivated field” was defined as cropland not planted with crops or grassland covered with herbs or shrubs in the year of interest, such that it was overgrown with herbaceous vegetation, belonging to the category of cropland (grassland) in CNLUCC. “Cultivated field” was defined as actively cultivated cropland in the year of interest. “Woody vegetation” was defined as woody vegetation. “Water” included rivers, lakes, reservoirs, seas, and other water bodies. “Other land cover” was defined as any land not covered by the above four categories, including built-up and bare land.

The Regions of Interests (ROIs) for training the land cover classification model were taken from 2019 and 2009. ROIs in 2019 including field survey ROIs and ex-pended ROIs. The expanded ROIs were developed based on the field survey data and the CNLUCC data (Figure 3) [36].

Firstly, we obtained two types of basic data including field survey data in 2019 from Guangdong Province and the CNLUCC of Guangdong province in 2010 and 2020 (as a reference for expanded ROIs in 2009 and 2019, respectively). Secondly, we constructed a land covers features library based on the basic data, including the spectral characteristics of each land cover from field survey data, the characteristics of VHR images, and Landsat true and false color images, such as shape, texture, color, and brightness. CNLUCC data were used as one of the references for developing ROIs. Thirdly, we generated the expanded ROIs based on the land covers features library from the second step. See “Development of ROIs” in Figure 3 for the detailed process. We preliminarily judged a land cover type based on the characteristics of the VHR image, and further checked whether it matched the spectral characteristics and CNLUCC. If it did not match, the sample point was discarded. If it matched, it was further judged based on the multi-season Landsat true and false color images. Note that our artificially expanded ROIs must completely match the characteristics of VHR and Landsat true and false color images, as well as the spectral characteristics and CNLUCC, before they can be determined as the final ROIs. To reduce the effects of sensor differences, ROIs for 2009 were modified based on the ROIs from 2019 and expanded following the flowchart in Figure 3. In total, the ROIs in 2019 (2009) were “Uncultivated field” [1108 (788) pixels], “Cultivated field” [666 (1399) pixels], “Woody vegetation” [1572 (2400) pixels], “Water” [235 (264) pixels], and “Other land cover” [376 (347) pixels] (Figure 4). To verify the reliability of our expanded ROIs and land covers classification approach, 30 sample points for each land covers randomly obtained from field ROIs in 2019 were used for verifying the land cover classification model and analyzing the misclassification for each land cover. The rest of field surveys ROIs (“Uncultivated field” 141 pixels, “Cultivated field” 23 pixels, “Woody vegetation” 10 pixels, “Water” 5 pixels, and “Other land cover” 33 pixels) combined with all the manually expanded ROIs from both 2009 and 2019 (“Uncultivated field” 1725 pixels, “Cultivated field” 1952 pixels, “Woody vegetation” 3104 pixels, “Water” 464 pixels, and “Other land cover” 660 pixels) were used to train the land covers classification model. As for the distribution of ROIs, “Cultivated field” areas distributed in various regions of Guangdong were sampled; these areas contained different types of crops. The ROIs of other land covers with high homogeneity of characteristics and small intra-annual variation were distributed in regions around the field survey ROIs in Meizhou, Zhanjiang, and Maoming (Figure 4).

We developed an algorithm for annual land cover classification in Guangdong from 1989 to 2020 based on key phenological data. Furthermore, we developed an algorithm that can correct for land cover misclassification; it was used to map cropland abandonment from 1992 to 2020 based on prior knowledge and natural laws. The algorithm was implemented in GEE. The workflow is shown in Figure 5.

2.3. Algorithm

2.3.1. Algorithm for Annual Land Cover Classification

Time series-based methods are difficult to apply in the cloud-prone, cropland-fragmented region of southern China. To solve this problem, we proposed a phenology-based algorithm to automatically index data for key phenological periods, and we used it in combination with an RF classifier for annual land cover classifications from 1989 to 2021.

Time series-based remote sensing methods have been widely applied to map cropland abandonment. However, the time series characteristics of cultivated land are difficult to determine in areas with many different crops. Moreover, Guangdong is a cloudy and rainy area, and a lack of suitable data makes it difficult to construct land cover time series characteristics. In this study, we were dedicated to obtaining dimensionality-reduced Landsat data for key phenological periods, which can effectively distinguish the five land cover types. To our knowledge, all land cover can be broadly categorized into two groups by judging whether it is covered with vegetation. As for our self-generated land cover types, “Uncultivated field,” “Cultivated field,” and “Woody vegetation” are the categories that covered with vegetation, while “Water” and “Other land cover” are the categories with no vegetation covered. These two categories can be distinguished based on their different degree of vegetation coverage, while the category covered with vegetation should be further analyzed, especially with regard to the annual vegetation changes. The NDVI is a typical spectral index that can reflect vegetation coverage and changes [44], while the NDSI is sensitive to bare soil [45]. Therefore, NDVI and NDSI time series data were used to analyze the vegetation coverage and vegetation changes of these five land covers. After analysis, it can be found that Landsat data for the combination of two key phenological periods can well distinguish the vegetation coverage and vegetation changes of the five land cover types.

One key phenological period is the “least vegetative phase” (LVP). During these periods, bare soil occurs in “Cultivated field” (Figure 6b) before and after the crops are planted, and the NDVI usually reaches a minimum while the NDSI reaches a maximum. For “Uncultivated field” in Guangdong (Figure 6a), herbaceous vegetation is generally sparse at the beginning and end of the year, when the temperature is lowest and there is no field management. The “Woody vegetation” (Figure 6c) class mainly included evergreen forests due to the warm and humid climate in southern China (https://en.wikipedia.org/wiki/Evergreen_forest, accessed on 24 December 2022); this class always shows fewer vegetation changes and higher vegetation coverage throughout the year than others, and its NDSI in this phase is higher than the other two land covers with vegetation-cover due to the less exposed soil (Figure 6a–e). The other key phenological period is the “peak vegetative phase” (PVP). Typically, the “Water” (Figure 6d) and “Other land cover” (Figure 6e) areas had little vegetation cover, and their NDVI was low and showed little variance throughout the year. For the other three land cover types, the vegetation density and NDVI were highest during the peak vegetative phase. Furthermore, these two key phenological periods presented that, among the category covered with vegetation, “Cultivated field” has the largest annual vegetation change, followed by “Uncultivated field”, and “Woody vegetation” changed the slightest. Therefore, data for LVP and PVP were integrated for annual land cover classification. Pixel-based Landsat data for LVP was composed by using the maximum annual value of NDSI minus the NDVI [(NDSI-NDVI)-max] as an index for each pixel, while that for PVP was composed by using the maximum annual NDVI (NDVI-max) as an index for each pixel.

$$\mathrm{NDVI}=\frac{{\rho }_{nir}-{\rho }_{red}}{{\rho }_{nir}+{\rho }_{red}}$$

(1)

$$\mathrm{NDSI}=\frac{{\rho }_{swir}-{\rho }_{green}}{{\rho }_{swir}+{\rho }_{green}}$$

(2)

To improve land cover classification, blue, green, red, near-infrared (NIR), and shortwave-infrared (SWIR) bands were used. Figure 7a–c show that the blue, green, and red bands can distinguish “Uncultivated field” from “Water” and “Other land cover”. The NIR band is also important for detecting vegetation changes [46,47,48]; Figure 7d shows that it could distinguish “Uncultivated field” from all other types of land cover except “Woody vegetation”. Meanwhile, SWIR differentiated “Uncultivated field” from “Water” and “Other land cover” in both the PVP and LVP; the other two land cover types could only be differentiated based on SWIR in the LVP.

Spectral indices (Figure 7f–i) were also used to facilitate classifications. Two vegetation indices (NDVI and EVI [49]) and land surface water index (LSWI) [50] clearly distinguished “Uncultivated field” from other land cover types, while the NDSI could only distinguish “Water” and “Other land cover”. The difference in annual NDVI (NDVIdiff) and EVI (EVIdiff) play an important role in annual land cover classification, and reflect changes in vegetation intensity for all land cover types.

$$\mathrm{LSWI}=\frac{{\rho }_{nir}-{\rho }_{swir}}{{\rho }_{nir}+{\rho }_{swir}}$$

(3)

$$\mathrm{EVI}=2.5\ast (\frac{{\rho }_{nir}-{\rho }_{red}}{{\rho }_{nir}+6.0{\rho }_{red}-7.5{\rho }_{blue}+1})$$

(4)

$${\mathrm{NDVI}}_{\mathrm{diff}}=NDV{I}_{max}-NDV{I}_{min}$$

(5)

$${\mathrm{EVI}}_{\mathrm{diff}}=EV{I}_{max}-EV{I}_{min}$$

(6)

Pixel-based images containing the selected sensor bands and spectral indices of PVP and LVP were combined with ground truth data and used as input for training the RF model (Figure 8). The RF algorithm, proposed by Breiman in 2001 [51], has been widely used for land cover classification. It improves decision tree algorithms that integrate multiple decision trees. Multiple decision trees bifurcate and “recurse” the input features of RF models simultaneously, enabling them to determine the most critical features according to information entropy. The final input features of the RF model are determined by voting for their importance. Our land cover classification model was trained by a random forest algorithm. Based on an accuracy and efficiency test with varying numbers of trees and different default parameters, the final RF model was trained using 200 trees. The trained RF model was used to distinguish five land cover types in the PVP and LVP, based on which annual land cover maps were generated from 1989 to 2021 for Guangdong Province.

2.3.2. Algorithm for Mapping Cropland Abandonment

The cropland abandonment mapping was based on annual land cover classifications, such that a clear definition of cropland abandonment and accurate land cover classifications were needed. We defined abandoned cropland as cropland without agricultural management (“Uncultivated field”) for at least 2 consecutive years, which avoided misclassification of year-round fallow fields and took consideration into the situation of the increasing population and decreasing area of cropland in China. A reduction of annual classification errors in the period 1989–2021 can improve the reliability of abandoned cropland mapping, and we proposed an algorithm for time-series correction [36] with a sliding window to map cropland abandonment based on prior knowledge and natural laws, as stated above (Figure 9). We are committed to correcting the pseudo land use changes that presented in the classification results caused by erroneous data (e.g., the mutation of “Water” into other land cover in individual years can be corrected in Step 1) as well as human activities/natural disaster (e.g., “Woody vegetation” misclassified as “Cultivated field” or “Uncultivated field” due to large changes in vegetation caused by human logging or typhoons can be corrected in Step 2) or different land covers with similar characteristics (e.g., the grass land or shrub vegetation classified into “Uncultivated field” in the annual land cover maps can be ruled out the possibility that they were mapped as abandoned cropland according Step 4). Note that, in order to correct the abnormal land type changes in the annual land cover maps more accurately, our time-series corrections algorithm was implemented based on the annual land cover classification with five land cover types, which is due to the fact that: refinement of land cover types provided the ability to be corrected according their respective characteristics; under some natural or human-activity influences, the “Woody vegetation” will be misclassified as the other two land cover types covered by vegetation, so it is necessary to correct it separately. However, our final maps only extracted abandoned cropland and cultivated cropland but merged “Woody vegetation,” “Water,” and “Other land cover” into one class. The correction and mapping procedures were as follows.

Step 1: We composed the resultant land cover maps from 1989 to 2021, separated each land cover type for a single map (labeled 1 for “Yes” and 0 for “No”), and used a sliding window of 2–4 years for correction. According to the definition of abandoned cropland, the “correction rules” for “Uncultivated field” were different from the other four land cover types. For “Uncultivated field”, all pixels with values of 1 in < 2 consecutive years were changed to 0, where “0 1 0” was converted to “0 0 0” with a sliding window of 3 years, and “0 1” to “0 0” in the last two years. Considering the stability of the other four land cover types, converted “0 1 0 0” to “0 0 0 0” with a sliding window of 4 years, “0 1” to “0 0”, “1 0” to “1 1” in the last two years, and “0 1 0” to “0 0 0” in the first three years, aiming to correcting the abrupt changes in land covers in individual years caused by remote sensing aberrant information.

Step 2: After separate correction, all long-term land cover types were combined in each year from 1989 to 2021. “Cultivated field,” “Uncultivated field,” “Woody vegetation,” “Water,” and “Other land cover” were labeled as 0–4, respectively.

Step 3: “Woody vegetation” comprised forest or thicket that was cut down or replanted after felling, and this type of vegetation coverage can change over time, thus woody vegetation can easily be misclassified as “Cultivated field” or “Uncultivated field.” Therefore, “2 1 0” was converted to “2 2 0,” “0 1 2” to “0 2 2” (sliding window of 3 years), and “2 1” to “2 2” in the last two years.

Step 4: After applying the above corrections, we mapped cropland abandonment. Land was identified as abandoned cropland only when at least 2 consecutive years of the “Uncultivated field” classification occurred after the “Cultivated field” classification (conversion of “1 1 0” to “0 0 0” in the first three years each pixel). Aiming to eliminate the grassland or shrubs that did not turn into abandoned cultivated land, as well as the commission error of sparse forests in “Woody vegetation” to “Uncultivated field,” ensured that the source of abandoned cropland is cultivated land. Therefore, cropland abandonment mapping stared in the second year, i.e., 1990. Finally, all land cover types except “Uncultivated field” and “Cultivated field” were merged into one class, i.e., land cover, types 2–4 were all assigned a value of 2. Due to uncertainty caused by the start and end of the study period, cropland abandonment maps with three categories (0 is “cultivated cropland” (CC), 1 is “abandoned cropland” (AC) and 2 is “other land cover” (OLC)) of land cover were generated from the fourth year (1992) to the penultimate year (2020).

2.4. Accuracy Assessment for Cropland Abandonment Maps

To validate the reliability of our algorithm, the accuracy of the cropland abandonment maps was assessed based on stratified random sampling using GEE (Figure 10). First, based on Landsat pixels, “validation points” were generated in Guangdong Province for AC, CC, and OLC. The size of the validation points was determined based on the expected overall accuracy (OA) of the cropland abandonment maps. Samples size allocations in AC, CC, and OLC depended on the proportions of their areas, but each stratum should have at least 75 samples (

Table 1. The number of random ground reference data of Abandoned Cropland (AC), Cultivated Cropland (CC) and Other Land-Covers (OLC) for study areas in 1995, 2000, 2005, 2010, 2015 and 2020.

Land Cover Types	1995	2000	2005	2010	2015	2020
AC	75	75	75	75	75	75
CC	99	102	104	75	75	75
OLC	363	357	356	399	395	388
Total	573	563	568	549	545	538

). The sampling points were verified as followed processes (take 1995, for example) (Figure 11). The validation for each sample point followed the development-based logic and details of expanded ROIs in Section 2.2.4 (Figure 3). The resultant maps of CC and OLC types were judged on the bases in 1995, while AC should be verified with three consecutive years (1994, 1995, and 1996). For the AC verified in 1995, the sample point was firstly judged in 1995. If the judgement for AC in 1995 was “No,” this sample point was misclassified as AC. The sample point should be further analyzed in 1994 and 1996 while the judgement was “Yes.” Only when the sample point was identified as AC for at least one year among 1994 and 1996 (1994 √ and 1996×; 1994× and 1996 √; 1994 √ and 1996 √), it was correctly identified. After verification, Kappa coefficients, OA, UA, and PA were calculated based on the confusion matrix. Because the sensors differed throughout the study period, verification was performed every 5 years from 1992 to 2020 (in 1995, 2000, 2005, 2010, 2015, and 2020).

3. Results

3.1. Accuracy Assessment of Annual Land Cover Classification Model

Based on the field survey ROIs in 2019 for approach validation, our annual land cover classification model had an overall accuracy of 0.87, while the UA (PA) for “Uncultivated field,” “Cultivated field,” “Woody vegetation,” “Water,” and “Other land-cover” were 0.79 (0.77), 0.84 (0.90), 0.81 (0.87), 0.97 (0.93), and 0.93 (0.87), respectively (

Table 2. Accuracy assessment of annual land cover classification model based on field surveys data.

Land Cover Types		Field Surveys Data						User’s Accuracy	Producer’sAccuracy
		Uncultivated Field	Cultivated Field	Woody Vegetation	Water	Other Land-Cover	Total
Map	Uncultivated field	23	1	3	0	2	29	0.79	0.77
	Cultivated field	3	27	1	0	1	32	0.84	0.90
	Woody vegetation	4	2	26	0	0	32	0.81	0.87
	Water	0	0	0	28	1	29	0.97	0.93
	Other land-cover	0	0	0	2	26	28	0.93	0.87
	Total	30	30	30	30	30	150	OA = 0.87

), which showed that the classification approach had certain reliability, but there were still different degrees of identification errors occurring to each land cover. Among which, both the UA and PA for “Uncultivated field” were lowest, while that was followed by “Woody vegetation”. It was mainly due to the fact that the shrubs included in “Uncultivated field” being the intermediate stage from herbaceous vegetation to woody vegetation actually had similar vegetation coverage and changes per annual year as sparsely forested vegetation in “Woody vegetation”. This situation can be further corrected based on our time-series annual land-covers maps corrections algorithm in Section 2.3.2; only when “Uncultivated field” is followed by “Cultivated field” for at least 2 years can it be identified as abandoned cropland, so the errors where sparsely forested vegetation is misclassified as “Uncultivated field” can be corrected based on the algorithm. The omission of “Cultivated field” mainly occurred for the cropland that had been abandoned for several years and naturally evolved into a shrub-dominated field. Furthermore, errors also may occur for each land cover in annual classification due to the missing data of key phenological periods (e.g., the commission errors of “Cultivated field” to “Uncultivated field” or “Woody vegetation” or “Other land-cover”) or remote sensing aberrant information (e.g., the commission errors between “Water” and “Other land-cover”) or sudden changes in “Woody vegetation” caused by human factors. These errors can be relieved based on the time-series correction approach for every land-cover that would be further corrected year-by-year.

3.2. Accuracy Assessment of Cropland Abandonment Maps

Based on the land cover classification model trained with 2009 and 2019 sample data, annual land cover classification maps were generated from 1989 to 2021, and cropland abandonment was mapped from 1992 to 2020 based on long-term land cover classifications corrected annually (on the basis of prior knowledge and natural laws).

The confusion matrix and results of the accuracy assessment of cropland abandonment maps in 1995, 2000, 2005, 2010, 2015, and 2020 are shown in

Table 3. Accuracy assessment of the cropland abandonment maps for study area in 1995, 2000, 2005, 2010, 2015, and 2020 based on random ground reference data.

Year	Land Cover Types		Reference				User’s	Producer’s
			AC	CC	OLC	Total	Accuracy	Accuracy
1995	Map	AC	56	4	15	75	0.75	0.90
		CC	3	85	11	99	0.86	0.90
		OLC	3	5	355	363	0.98	0.93
		Total	62	94	381	537	OA = 0.92	Kappa = 0.85
2000	Map	AC	55	2	18	75	0.73	0.92
		CC	2	90	10	102	0.88	0.96
		OLC	3	2	352	357	0.99	0.93
		Total	60	94	380	534	OA = 0.93	Kappa = 0.88
2005	Map	AC	60	4	11	60	0.80	0.91
		CC	3	90	11	3	0.87	0.89
		OLC	3	7	346	3	0.97	0.94
		Total	66	101	368	66	OA = 0.93	Kappa = 0.85
2010	Map	AC	55	2	18	75	0.73	0.93
		CC	1	62	12	75	0.83	0.86
		OLC	3	8	388	399	0.97	0.93
		Total	59	72	418	549	OA = 0.92	Kappa = 0.81
2015	Map	AC	65	0	10	75	0.87	0.96
		CC	3	64	8	75	0.85	0.94
		OLC	0	4	391	395	0.99	0.95
		Total	68	68	409	545	OA = 0.95	Kappa = 0.88
2020	Map	AC	63	3	9	75	0.84	0.90
		CC	4	63	8	75	0.84	0.85
		OLC	3	8	377	388	0.97	0.96
		Total	70	74	394	538	OA = 0.93	Kappa = 0.82

. The highest OA was 0.95 (2015), while the lowest was 0.92 (1995, 2010); the highest kappa coefficient was 0.88 (2000, 2015) and the lowest was 0.81 (2010). The PA and UA of OLC were higher than those for AC and CC (PA, 0.93–0.96; UA, 0.97–0.99). For AC and CC, the PA ranged from 0.90 to 0.96 and 0.85 to 0.96, respectively. For CC and AC, the UA ranged from 0.90 to 0.96 and 0.73 to 0.87, respectively. The lower UA for AC was related to the fact that AC covered with dense bushes that had been abandoned for several years have the similar phenological characteristics with the shrubby vegetation in OLC, leading to misclassification of OLC as AC. The uncertainty is further analyzed in Section 4.2.

3.3. Spatial-Temporal Patterns of Cropland Abandonment

The spatial-temporal pattern of cropland abandonment maps for large-scale and details of typical areas in Guangdong were presented for every five years from 1995 to 2020.

From large-scale perspective (Figure 12), the extent of cropland abandonment increased significantly after 2000, accompanied by a decrease in cultivated land. As shown in Figure 12b,f, the most obvious increase in cropland abandonment and decrease in cultivated cropland occurred in the northern part of western Guangdong (WG), the entire Pearl River Delta (PRD), the central part of eastern Guangdong (EG), and the eastern part of northern Guangdong (NG).

From the detail view of typical area of cropland abandonment, Figure 13a1,a2 show the spatial-temporal changes in plain areas over a 6-year period. Abandonment mainly occurred in fragmented cropland, and the area thereof was small compared to the large area of contiguous cropland. Figure 13b1,b2 show that cropland decreased sharply in the PRD from 2005 to 2010 due to land-use transformation from cropland to fishponds and built-up land. As shown in Figure 13c1,c2, a large area of cropland in the mountainous regions had been abandoned by 2020 due to the fragmentation of cultivated land, and complex terrain in hilly and mountainous areas make it difficult to meet mechanized agricultural management, while the management is high-cost and labor-intensive work, and, with the movement of people from the rural areas to urban, there are not enough people available to carry out the agricultural work there.

3.4. Extent of Cropland Abandonment during the Last 30 Years

The area of cropland abandonment, cultivated area, and rate of abandonment in Guangdong Province from 1992 to 2020 are shown in Figure 14. Over the last 30 years, the area of abandoned cropland has remained relatively stable, at around 50 × 10⁴ ha, although the abandonment rate increased gradually with the decrease in cultivated area since 2000. From 1992 to 1995, due to large-scale construction and changes in the agricultural structure in rural areas, the cropland area declined significantly, while the abandoned area increased to over 60 × 10⁴ ha and the abandonment rate to 16%. With the introduction of cropland protection policies and convictions for illegal occupation of cropland, destruction of arable land, unlawful granting of land ownership, and illegal transfer of land, the abandoned area decreased from 1996 to 2000. The rate of abandonment decreased to < 10% in 2000 but gradually increased thereafter, reaching 20% between 2000 and 2017. From 2017 to 2019, the rate of cropland abandonment exceeded 20%, which is the highest during the last 30 years. However, the area of cropland either remained unchanged or increased slightly from 2017 to 2020, while the rate of abandonment in 2020 decreased. The main reason for this decrease was the outbreak of coronavirus 2019 at the end of 2019, which highlighted the need to ensure food security and self-sufficiency [52].

4. Discussion

4.1. Reliability of the Cropland Abandonment Mapping Algorithm

This study proposed a feasible and reliable algorithm for mapping cropland abandonment over the past 30 years in a cropland-fragmented and cloud-prone region of southern China, based on a phenological classification system and long-term classification corrections using Landsat imagery and GEE. The proposed method for mapping cropland abandonment was based on non-time series data and applied over the period 1992–2020. This was possible due to the availability of massive, long-term Landsat remote sensing data and sufficient computing power, as well as the stability of evergreen woody vegetation in south China.

First, the existing methods based on high temporal resolution but coarse-resolution images met the requirements for analyzing time series characteristics. However, for cloud-prone areas of highly fragmented cropland, such as southern China, coarse-resolution pixels can lead to large errors. In addition, remote sensing data does not cover the period when cropland abandonment began in China, i.e., the 1990s [5,53]. Long-term Landsat-SR imagery with 30 m spatial resolution captured since 1989 can overcome the limitations of moderate-resolution images. The high number of good-quality observations (>12 scenes for every 5° increment in latitude) in most years allowed us to analyze two key phenological periods.

Second, our land cover classification method was implemented based on a phenological analysis of different land cover types, especially ones associated with vegetation. First, we defined land cover types based on NDVI and NDSI time series data, which provide information on vegetation cover and changes therein. Landsat-SR image features for LVP and PVP were composed according to the NDSI-NDVI-max and NDVI-max, respectively. For annual land cover classifications, an RF classifier was used with LVP and PVP data serving as input. Unlike existing time series-based methods, our algorithm may alleviate the problem of variation in time series characteristics for the same land cover types, which can arise when there are few data in cloud-prone areas; this may lead to land type misclassification (Figure 15).

Compared to previous studies on changes in time series spectral indices [25], we incorporated more features, including sensor bands and spectral indices, to improve the accuracy of land cover classifications. Finally, dimensionality reduction of the data in this study reduced the computational burden, such that it became possible to map cropland abandonment over the long term.

Third, we mapped cropland abandonment over 30 years and calculated the abandonment rate; corrections of time-series algorithm were applied based on prior knowledge. With consideration of the effects of data error and the stable characteristics of woody vegetation in southern China, the annual land cover classification corrections performed in this study enabled the production of more reliable and intuitive maps of annual abandoned cropland. Moreover, our algorithm could exclude types of grassland that may lead to misclassifications, thus improving the accuracy of the cropland abandonment maps. We also determined the duration of abandonment in every year (Figure 16), which is important to balance food security with ecological considerations [21]. The typical duration of abandonment for given areas in Guangdong Province from 1992 to 2020 was 2–5 years, and newly abandoned cropland was seen every year. The longest duration of abandonment was 18 years, but this area was very small. Cropland abandonment has been reasonably well controlled in Guangdong, although longer cropland abandonment times were associated with lower soil carbon sequestration capacity, which greatly reduced ecological benefits and affected food security.

Fourth, GEE is a cloud-based computing platform for processing satellite imagery and other observation data. It provides a large database of satellite imagery, immense computational power, and numerous algorithms for geospatial analysis. In this study, we searched for and collected cloudless Landsat data from GEE for the annual LVP and PVP analyses. Because the LVP and PVP differed among land cover types over the study area at the pixel scale, this step would have been almost impossible without using GEE. Moreover, the RF classifier for land cover could be used directly in conjunction with the GEE platform. Furthermore, due to its high-performance parallel computing capabilities, efficient processing and analysis of large-scale, long-term geospatial data can be conducted to map cropland abandonment using the GEE platform. Our proposed method to map cropland abandonment based on land cover maps over a 32-year period required only ~30 min for annual-scale mapping of a 19.8 × 10⁴ km² area for Guangdong Province at 30 m spatial resolution.

4.2. Potential Sources of Uncertainty

The accuracy of the proposed cropland abandonment mapping algorithm may be affected by several factors. Firstly, 30 m spatial resolution Landsat imagery used in this study may lead to a part of errors occurring to the mixture pixels of the extremely fragmented cropland. Probabilistic methods may alleviate such problems. Secondly, the expanded ROIs for training land cover classification model may have a very small number of wrong samples due to the uncertainty of human judgment and remote sensing data, which may lead to a slight decrease in the accuracy of the model. Thirdly, the lower temporal resolution for Landsat imagery with 8~16 days in cloud and rain-prone southern China still presented uncertainty for our phenology-based land-cover classification algorithm. The composite data of LVP and PVP for different land cover types reflect their vegetation coverage as well as the degree of vegetation change. However, when cultivated or abandoned fields lack Landsat imagery during the LVP (bare soil for a cultivated field and withered vegetation for an abandoned field), they may be misclassified as “Woody vegetation” due to their characteristic of denser vegetation and fewer vegetation changes. To alleviate the errors caused by this type of uncertainty, the annual land cover maps were further corrected based on the long-term corrections algorithm. Fourthly, in a small number of croplands abandoned for a long time, the vegetation has gradually evolved from grass only to shrubs, or even woody vegetation, often leading to misclassification as “Woody vegetation”. Finally, our long-term correction algorithm has a degree of uncertainty. For classification as abandoned cropland, abandonment for ≥2 consecutive years was required, which led to classification uncertainty in the first year of the study period (1989) due to a lack of data prior to that year. To reduce uncertainty, cropland abandonment maps were thus obtained from the fourth year (1992) onward. Similarly, cropland abandoned in the final year (2021) of our study for the first time (“0 1”) could not be classified as such because there was no subsequent abandonment data. This may possibly lead to some underestimation of the area of abandoned cropland and overestimation of the area of cultivated land. Some of the land cover classification error caused by this uncertainty cannot be easily corrected, which will affect the accuracy of cropland abandonment maps.

4.3. Implications of This Study and Future Work

Since the reform and opening up of China, due to rapid economic development, cropland abandonment has occurred to varying degrees across the country. The government has also continuously introduced policies for arable land protection, resulting in differences in the rate of cropland abandonment among areas. Accurate monitoring of the temporal and spatial patterns (and duration) of cropland abandonment is important to ensure food security and a good ecological balance. However, cropland analyses are highly complex because this land use type is affected by both human activity and the natural environment [30]. Furthermore, to meet the demand for development in some areas, and because of the influence of climate change, some cultivated land must have fallow periods, which can lead to misclassification of abandoned cropland.

Overall, the identification of large-scale cropland abandonment in China is important but challenging. The phenological characteristics of land cover types covered with vegetation (including cultivated land, abandoned land, and woody vegetation) differ by latitude and terrain. The classification algorithm introduced in this study is only applicable to southern China, where woody plant cover is relatively stable. Further research is required, including determination of the different phenological characteristics of land cover types among regions; this will enable the application of knowledge-based methods and large-scale remote sensing data. In addition, approaches for mapping cropland abandonment that predict land cover types and correct for errors therein (based on time series data) should be implemented to reduce uncertainty and improve mapping accuracy [23,28,33].

5. Conclusions

Long-term monitoring of cropland abandonment is necessary to determine the drivers of land cover changes, and to achieve a good ecological balance and social stability. Due to the extensive economic development of Guangdong Province in recent decades, land cover has changed substantially. Existing time series-based approaches for monitoring cropland abandonment in southern China face difficulties, such as cloudy and rainy weather and fragmented cropland patterns. We proposed a land cover classification method extracting key phenological characteristics, but not requiring complete time series data; this approach was implemented for cropland abandonment mapping in southern China. Moreover, based on prior knowledge, long-term correction of aspects of cropland abandonment was achieved, including when and where the abandonment began and the annual rate and duration thereof. Our method not only overcomes the difficulty of applying time series methods to cloudy and highly fragmented regions, but also provides detailed spatiotemporal information that could inform ecological protection and food security policies. Our results indicated that the cropland abandonment mapping algorithm was generally reliable, generating accurate maps in most study years. Our method was mainly applied in southern China, where woody vegetation dominates evergreen forests; further research is needed for large-scale mapping of more complex vegetation.