| Alfalfa（Medicago sativa）is an important forage source for grassland agricultural development and plays an important role in ensuring ecological and food safety,so it would be worthwhile to explore accurate and fast methods of alfalfa remote sensing identification and yield estimation.In this study,taking Hexi Region of Gansu Province as the study area,Sentinel-2 high-resolution remote sensing images and the Google Earth Engine（GEE）platform were used to establish a cloud-free normalized difference vegetation index（NDVI）time-series dataset while comprehensively considering phenological stages,growth characteristics,and the ecological environment.We proposed an effective method for alfalfa remote sensing feature extraction using the Findpeaks function of MATLAB to establish an NDVI-based alfalfa recognition and yield estimation method that automatically finds the“critical time nodes”.This method can determine the spatial distribution extent of cultivated alfalfa grassland in 2019 to2021 and thus enables fast and accurate macroscopic monitoring of alfalfa hay production in the study area.Results show that:1.By using Gee platform,the remote sensing image could be downloaded quickly and the cloud could be removed.The optimization effect of sliding interpolation for Alfalfa NDVI time series data was remarkable,it could effectively deal with the fault and outliers in the original NDVI time series curve,and avoided the"wrong classification"and"missed classification"caused by the outliers in Sentinel-2 satellite image data.2.Through the comparative analysis of the differences of identification methods,it was found that all the indicators of cultivated alfalfa identified by the Find Troughs method were better than the Find Peaks method.Using the optimized Sentinel-2 NDVI time series dataset as the data source,the cultivated alfalfa identified by the valley finding method has the best recognition accuracy in 2020,the Kappa coefficient is 0.81,and the overall accuracy is 92.5%.The user accuracy and producer accuracy were greater than 90%,and the position accuracy was 85.22%.The overall recognition accuracy in 3 years could reach 90.42%.The key areas for alfalfa cultivation in Hexi were located in Jiuquan city and Zhangye city,showed a concentrated contiguous planting pattern,accounted for 32%and 30%of the alfalfa planting area in the study area respectively;followed by Jinchang and Wuwei city,accounted for the alfalfa planting area in the study area 16%and 18%.From 2019 to 2021,the total area of cultivated alfalfa in Hexi region was 43497 ha,64720 ha and 57952 ha,respectively.3.The hay yield of cultivated alfalfa had a very significant correlation with NDVI,and the power function regression model had the highest coefficient of determination（R~2 were all greater than 0.6）,and there was a good linear relationship between the model predicted yield and the measured yield（R~2 were all greater than 0.6,the average relative error（MRE）was basically stable at about 20%）.The fitting function of the overall optimal model for the hay yield of cultivated alfalfa in Hexi area was y=9035.5x0.8339.The prediction accuracy of the yield estimation model was high,and it had good stability and validity.In terms of per unit yield,the average annual yield of cultivated alfalfa was the highest in 2020,with an average hay yield of 20206.40 kg/ha,followed by 2019,with an average hay yield of 19324.47 kg/ha,while in 2021,the average hay yield of cultivated alfalfa in the study area was 19099.33 kg/ha,which was the lowest in three years.In terms of total yield,the total yield of cultivated alfalfa in2020 was higher than that in 2019 and 2021,the total hay weight was 1139.82×10~3t,the second was 2021,the hay weight was 868.63×10~3t,and the lowest in 2019 was673.01×10~3t. |
