Research On Remote Sensing Monitoring For Fallow And Abandoned Cropland In The Farming-Pastoral Ecotone

In the past few decades,the continuous deterioration of the ecological environment and the largescale reclamation of natural grasslands have led to the complexity of land cover and land use patterns in the farming-pastoral ecotone in Inner Mongolia.The decline of soil fertility and the transfer of agricultural labor have made agricultural land more variable.Futhermor,the ecological restoration "Grain for Green" program has curbed the continuous deterioration of the ecological environment since launched in 2000.However,it has also aggravated the instability and complexity of the cropland landscape in the farmingpastoral ecotone in Inner Mongolia.Previous studies have shown that fallow and abandonment of cropland are common in arid and semi-arid regions,with extensive and concentrated contiguous distribution characteristics.It has become a new contradiction in land use,which has a long-term and farreaching impact on ecological security and food production.Therefore,remote sensing monitoring of fallow and abandoned cropland will contribute to evaluating the impact of complex cropland landscape on the ecological environment and global food security.The 50 counties of farming-pastoral ecotone in Inner Mongolia,covering a total area of 411997.10 square kilometres,has been chosen to monitor the pattern of the agricultural land in-depth using remote sensing technology.Adhering to the principle of mining in-depth for multi-sensor remote sensing data with Google’s powerful cloud computing platform,the research has been divided into the following four parts:(1)Landsat-5 TM annual images of 2000 were used for extracting cropland since 2000 is a year with less human intervention.However,because the retried cropland with unique landform,spectral and texture characteristics produced by the artificial afforestation plan of the "Grain for Green" program has withdrawn from agricultural production,it needs to be carried out from the cropland base map in 2000.Therefore,Sentinel-1 and Sentinel-2 data were used to establish the optimal classification datasets and extract the retired cropland in 2020.In addition,the prior knowledge data was utilized to mask out the water body and impervious surface from the cropland base map.Finally,the pure cropland pixels from2000 to 2020 will be obtained as the base map for the subsequent research.(2)The optimal classification datasets,including spectrum,index,texture,and terrain metrics,were farmed to identify non-active and active cropland.The annual sample data was divided into non-active cropland pixels and active cropland pixels using the automatic classification method based on setting a vegetation index threshold.The long-term Landsat 30-m resolution data from 2000 to 2020 and the Random Forest classifier have been utilized for obtaining the annual classification maps of non-active and active cropland.Furthermore,the Land Use Change Trajectory method was adopted for conducting a superimposed analysis on the classification maps from 2000 to 2020.The definition of fallow and abandonment of cropland announced by the United Nations Food and Agriculture Organization has been accepted as a criterion to generate a long-term series of the fallow,abandonment,and recultivation map of the farming-pastoral ecotone in Inner Mongolia.As a result,the spatiotemporal characteristics of the abandoned cropland were analyzed briefly.Furthermore,the annual cropland abandonment rate in 50 counties was taken as the explanatory variable.The annual natural and socio-economic elements were used as the driving factors to explain why cropland has been abandoned.The panel data were formed in State,and the Fixed Effect Model was adopted to regress the correlation between each factor and the cropland abandonment in statistics.(3)To explore the applicability of Sentinel data in the remote sensing monitoring of fallow and abandonment cropland,2020 was taken as an example to evaluate the classification performance in the three different scenarios with backscattering coefficient,optical data,and the combination of two kinds of data,respectively.As a result,the classification accuracy in three scenarios was systematically expressed in determining the optimal classification dataset and the optimal strategy with timeliness and accuracy for monitoring the fallow and abandonment cropland based on multi-sensor images.Moreover,the strategy was used in the next step to obtain a 10-m resolution map of the fallow and abandoned cropland using Sentinel multi-sensor remote sensing data from 2016 to 2021.The following conclusions were drawn in this research:(1)The cropland in 2000 and the cropland and retired cropland in 2020 were classified accurately on the Google Earth Engine.The classification accuracy of cropland and others reached 91.44% in research of 2000,and the F1 score of complex croplands reached 0.91.On the other hand,the classification accuracy of retired cropland also obtained high accuracy with Overall Accuracy equal to88.95% of 2020.The F1 scores of the retired cropland and complex cropland were 0.86 and 0.92,respectively.Therefore,retired cropland,water bodies and impervious surfaces provided by global land use maps were masked out from the cropland base map for the long-term series of remote sensing monitoring of the cropland fallow and abandonment.The 10-m resolution cropland map is used for the the same kind of research using Sentinel multi-sensor remote sensing images.(2)The Normalized Difference Vegetation Index(NDVI)difference between active and non-active cropland on typical samples randomly generated from the cropland base map was analyzed.The NDVI equal to 0.35 was used as a threshold to classify the samples into active cropland and non-active cropland automatically in the primary growing season;The classification accuracies of annual active and nonactive cropland maps are above 90%,the F1 scores of both two cropland types are above 0.90,except for a few years.The monitoring accuracy of fallow and abandoned cropland based on the Land Use Change Trajectory method is 79.29%.The accuracies of the continuously active cropland,fallow land,and abandoned cropland are 75%,100% and 62.86%,respectively.The panel data regression analysis framed using the time series of cropland abandonment rates and natural and socioeconomic factors in 50 counties of the study area showed that total evapotranspiration and volumetric soil water significantly positively correlated at the level 0.01 with cropland abandonment rates.While wind speed,total precipitation,and the drought index showed significant positive correlations at the level 0.05 with the cropland abandonment rate.The disposable income of permanent residents in rural areas and the proportion of arid fields showed negative impacts and no significant correlations with the abandoned cropland rate.Moreover,the proportion of agricultural employment has a significant negative correlation at the level0.05 with the cropland abandonment.(3)Based on Sentinel-1 and Sentinel-2 collaborative data and automatically generated samples,the Overall Accuracy of remote sensing monitoring on active and non-active cropland classification in 2020 reached 94.73%.The result shows that using Sentinel multi-sensor data only obtained from August can reach a high accuracy of 94%,which is the earliest period when non-active cropland can be accurately identified if the clear observation was sufficient.In addition,a classification accuracy of 79.20% can be obtained using Sentinel-1 radar data alone during the growing season in the study area.Although the noise is severe,it might have a comprehensive application prospect in the region with cloudy and rainy weather conditions.Furthermore,the accuracy of remote sensing monitoring of fallow and abandonment at 10-m resolution show improvement compared with the previous study based on the Landsat data.The accuracy has increased from 79.29% to 87.91%,and that of abandoned cropland has increased from 62.86% to75.38%.In general,the advantages of the cloud computing platform and multi-sensor remote sensing data were utilized to monitor cropland planting conditions and the challenges of fallow and abandonment monitoring have been discussed.Also,improving the automation of remote sensing monitoring on cropland fallow and abandonment was preliminarily tested in this article.The methodology for remote sensing monitoring of cropland fallow,abandonment and reclamation based on Landsat 30 resolution data and Sentinel multi-sensor data can be a solid foundation for future work.Also,it can be a valuable insight for research in this field.