Our ultimate goal is to establish an optimal classification metrics set that is suitable for the study area through these steps.In this study, Landsat-5 TM SR data were chosen to establish spectral and index metrics sets for 1990, 2001, and 2010, while the Landsat-8 OLI SR data were used to develop spectral and index metrics sets for 2019 . The yearly mean value composite images of 2010 and 2019 were taken as examples to analyze the spectral and index separability of different land use classes. Therefore, we took the annual reflectance mean value of varying land use samples were extracted from images in 2010 and 2019 to analyze the separability between active cropland , non-active cropland , retired cropland , natural grassland , impervious surface , forest and water body . It can be seen from Fig. 3 that IS, WB, AF, and FR were well separated on visible bands and indices based on both Landsat-5 TM and Landsat-8 OLI images. The mean reflectance values of NAC and RCL were different from other land use classes, round plastic pots while they were easily confused with GL. The reflectance of RCL was lower than the NAC and GL on the visible and shortwave infrared bands and higher than the NAC and GL on the NDISI and NDVI. Although NAC, RCL, and GL were not well separated on the visible bands, they well separated on the shortwave infrared bands and NDISI and NDVI based on the TM image.
However, Fig. 3d shows that the confusing three land use classes were not well separated on the six OLI image bands and three indices. The NCL and GL curves showed high overlap on spectral bands and indices based on both TM imagery and OLI imagery. There was a vast grassland area in the farming-pastoral ecotone in the northern foot of the Yinshan Mountains, and the NCL was also widely distributed. Since NCL and GL cannot be well separated in spectral bands and indices, especially RCL cannot be separated from other easily confused land use on OLI imagery. Therefore, to distinguish better RCL and NAC from other land use classes, it is necessary to add the texture metrics to classification methodology.GGP also promotes land use classes such as arid cropland, barren mountains, and desert grassland with low ecological functions to land with high ecological functions. Our study present, the retired cropland in the study area mainly includes three types of land use change trajectories: 1: converted from the cropland in 1990; 2: converted from the cropland cultivated after 1990, and 3: converted from the other land use classes. Their areas are respectively 51.97%, 25.02%, 20.10% of the total area of retired cropland in 2019. Trajectories of other types account for a relatively small area by 2.91% of the total area of retired cropland in 2019. See Table 4 for details. The retired cropland in the northern foot of the Yinshan Mountains mainly distributes in Duolun, Taipusi, Huade, Shangdu, Chahar Right Rear, Chahar Right Middle, and Siziwang County . The eastern region has a larger amount of precipitation, and the climate conditions are suitable for the survival of the shrubs planted by GGP. Furthermore, for the remaining four counties in the western region, due to drier climate conditions and a smaller proportion of cropland, there is also less cropland to be returned.
Since Wuchuan City has more high-quality cropland than other counties, it has the smallest area of retired cropland as shown in our result. Nearly 45% of the retired cropland converted from cropland in 1990, and around 25% of the retired cropland converted from cropland cultivated after 1990, which were in line with GGP’s original intention to retire some of the cropland with low producibility to plant wood or shrub with higher ecological service functions. What needs to be emphasized is that nearly 20% of the retired cropland was not marked as cropland in 1990, 2001, and 2010. It is possible that for ten years was used as a time interval, the retired cropland may be converted from cropland cultivated within two-time notes, or it may convert from other lands .To evaluate the vegetation dynamics after GGP, we divided the last three decades into three time periods: 1990–2000, 2001–2010, and 2011–2019, and calculated the long-term Landsat NDVI-mean value of retired cropland area and that of the entire study area. First, to evaluate the vegetation dynamics in the area of retired cropland, the mean value of NDVI per decade within the scope of retired cropland in 2019 was calculated. The result present in Fig. 7. It can be seen from Fig. 7 that the mean value of NDVI of retired cropland in the three periods increased gradually with mean values 0.1524, 0.1545, and 0.1728, respectively. Moreover, the growth rate from 2011 to 2019 is higher than the rate from 2001 to 2010. This result indicates a significant vegetation restoration in retired cropland areas over the last three decades. However, over the previous ten years, vegetation restoration showed a considerable increase related to the launch of GGP in the study area and its strengthening during the last decade. Second, to evaluate the vegetation dynamics of the entire study area, each decade’s NDVI-mean value was calculated. A correlation curve with the change of the retired cropland area and the change of the cropland area was made and is presented in Fig. 8. During the past 30 years, the NDVI-mean value of the entire study area showed an increasing trend, as shown in Figs. 8 and 9, with values 0.1559, 0.1562, and 0.1749, respectively.
The NDVI-mean value of the study area did not change greatly between 1990 and 2010, but as the percentage of retired cropland in the total area increased considerably from 2011 to2019 , the NDVI-mean value of the study area has increased significantly from 0.1562 in 2010 to 0.1749 in 2019 as well . The vegetation restoration in the entire study area was also accompanied by a decrease in the percentage of cropland in the total area . The result shows that, on a large area, the decrease of cropland and the increase of retired cropland associated with the vegetation restoration to a certain extent. The retired cropland can show a more stable NDVI value on remote sensing imagery than cropland and with fewer seasonal changes. However, the change of NDVI value in the study area is affected by many factors, such as climate change and natural ecology. The long-term NDVI-mean value was used in this study shows a significant correlation with decreasing cropland and increasing retired cropland, as shown in Fig. 8.The farming-pastoral ecotone environment in the northern foot of the Yinshan Mountains in Inner Mongolia is severely damaged due to drought and endangered ecological environment. The land use in this area has also changed greatly because of the change in a natural environment and artificial afforestation program such as GGP. In the past 20 years, with the rapid development of remote sensing technology, the study of evaluating the ecological environment of farming-pastoral ecotone in the northern foot of the Yinshan Mountains has attracted extensive attention. In this study, the RF-GEE classifier and multi-metrics were addressed to identify the changes of cropland and retired cropland in the northern foot of the Yinshan in Inner Mongolia over the past 30 years. The influence of the multi-metrics set on classification accuracy, the relationship between LUCC and vegetation restoration in the study area are the key points of our research. 4.1. The accuracy of cropland and retired cropland mapping The training data quality is one of the critical factors to obtain satisfactory classification results .
The fact that active cropland and non-active cropland exit simultaneously in the northern foot of the Yinshan Mountains was fully considered when the cropland sample was selected. The sample of non-active cropland accounted for 20–30%. Although nonactive cropland and retired cropland are easily confused with natural grassland in the study area , the highest OA and Kappa coefficients of cropland classification were obtained in this study . The highest F1 score of croplands was 0.94, which is higher than the accuracy of previous land cover mapping studies focused on entire Inner Mongolia or global scale based on Landsat data . The accuracy of our study can support subsequent research about complex cropland use patterns, such as remote sensing monitoring of fallow and abandoned cropland in the farming-pastoral ecotone. On the other hand, the highest F1 score of the retired cropland was 0.75, with lower accuracy than the cropland and others. Retired cropland’s spectral and index metrics have weak separability with those confusing land use classes ,hydroponic bucket and the texture metrics of retired cropland are similar to that of cropland. In the research on land use in InnerMongolia, there are very few studies that classify retired cropland as a single land use class, and most of these previous remote sensing land use classification researches in the northern foot of the Yinshan Mountains adopts visual interpretation method. The precision of cropland can reach more than 0.9. Chun used Landsat TM data to visually interpret the land use change in Wuchuan County in the northern foot of the Yinshan Mountains, with accuracy over 0.9 and concluded that the cropland in this area has a continuous decreasing trend. Wang took the year 2000 when human intervention was minimal as an example, using supervised classification and visual interpretation to produce a city-level land use map of Ulanqab city in the middle part of the northern foot of the Yinshan Mountains. In this research, the UA of arid cropland and irrigated cropland were higher than 0.95, and the PA of arid cropland and irrigated cropland were higher than 0.80. However, the methods introduced by Chun and Wang are time-consuming and labor-intensive and do not have the advantages to the large-scale research. Most of the land use classification in large-scale studies has not mentioned the retired cropland, while there are remote sensing results of the shrubland or sparse shrubland on a large scale, with the accuracy ranging from 0.30 to 0.70 .
Although the “Grain for Green” project has been implemented in Inner Mongolia for 20 years, the existing land use research in Inner Mongolia classifies the retired cropland into the category of woodland. It does not certify it as a separate type of land use, except for studies with smaller spatial scales. A study has explained this problem from another view; Yin et al. classified the degraded cropland in Inner Mongolia as a single land category. The UA of degraded cropland varied from 0.42 to 0.70 in different years. The PA was between 0.65 and 0.96. Nevertheless, still retired cropland is differing from degraded cropland even in the ecological transition zone. From the results in Section 3.1, it can be seen that the vegetation coverage of the retired cropland in arid and semi-arid regions is not high. This phenomenon is expected in the farming-pastoral ecotone in the northern foot of the Yinshan Mountains in Inner Mongolia. Therefore, classify the retired cropland as a single land use type has a great significant when monitor the LUCC in farming-pastoral ecotone of northern China using remote sensing technology.Human land use is a dominant driver of the greening earth . GGP is a typical case of Human-intervention in land use, especially in the farming-pastoral ecotone. Previous studies have shown that GGP has improved China’s ecology and of Inner Mongolia . By the end of the 20th century, the vast grasslands in the farming-pastoral ecotone in northern China started to be cultivated and approximately doubling the cropland area. A large amount of cropland was often accompanied by extensive land degradation . Inner Mongolia is considered one of the most severely degraded regions in China. Therefore, almost all national environmental protection land restoration projects were launched first in Inner Mongolia, which became the Chinese province with the highest investment in ecological restoration programs . Land degradation directly affects the region’s vegetation dynamics, which is particularly prominent in the Mongolian Plateau .