基于Sentinel-2A的棉花种植面积提取及产量预测

及时、准确预测棉花产量在棉田经营管理、农业决策制定等方面具有重要的价值和意义。为了提高棉花产量预测精度并确定估产的最佳生育时期，该研究利用谷歌地球引擎（Google Earth Engine，GEE）获取2020年Sentinel-2A的3个时间段影像，采用随机森林（Radom Forest, RF）、支持向量机（Support Vector Machine, SVM）、决策树（Classification and Regression Tree, CART）进行棉花种植区域提取，利用顺序向前选择（Sequential Forward Selection, SFS）和偏最小二乘算法（Partial Least Squares Regression, PLSR）确定棉花产量预测最佳生育时期，最终形成莫索湾垦区棉花产量预测分布图。结果表明，1）RF分类效果最佳，农田与非农田分类总体精度为0.94，Kappa 系数为0.89；棉田与非棉田分类总体精度为0.92，Kappa 系数为0.83。2）红边波段（B6）在3个生育时期中与产量相关性较好，相关系数随着生育时期的递进而增加，分别为0.37、0.47、0.53。3）基于PLSR构建的产量预测模型中，铃期预测效果最佳（决定系数R2=0.62，均方根误差RMSE=625.5 kg/hm2，相对误差RE=8.87%），优于吐絮期（R2=0.51，RMSE=789.45 kg/hm2，RE=11.06%）和花期（R2=0.48，RMSE=686.4 kg/hm2，RE=9.86%），铃期为棉花产量预测的最佳生育时期。该研究利用GEE和Sentinel-2A影像数据，为新疆莫索湾垦区棉花种植面积提取及产量预测提供一种新的思路，可为合理水肥配置、精准种植、农作物生长过程监测提供数据支撑。

Cotton planting area extraction and yield prediction based on Sentinel-2A

Timely and accurate prediction of cotton yield has been one of the most important steps in cotton field management and decision making. The cotton yield can be recognized as a continuous and dynamic process, particularly depending mainly on the canopy structure, biomass, and chlorophyll content. In most previous studies, the vegetation index with the highest correlation can often be used to establish a regression equation with the cotton yield as a prediction model. However, a single vegetation index cannot effectively explain the complexity of cotton yield prediction, leading to the error and instability of the model during the subsequent application. In this study, multiple vegetation indices were selected to explore the promising potential in cotton yield prediction. Cotton in the Mosuwan Reclamation Region of Xinjiang, China was taken as the research object. Sentinel-2A image data was set as the data source in 2020. These remote sensing images were obtained to extract the cotton planting area for the yield prediction by combing with the Google Earth Engine (GEE) cloud platform and machine learning. A two-step grading cotton mapping strategy was adopted to extract the cotton from the images. In the first step, 17 features (12 spectral, 3 vegetation index, and 2 DEM features) were selected to extract the farmland by excluding non-crops. In the second step, three classification methods (Random Forest, RF), Support Vector Machine (SVM), and decision tree (CART)) were used to screen the four indicators (Overall Accuracy (OA), Producer''s Accuracy (PA), User''s Accuracy (UA), and Kappa Coefficient (KC)), where the best was selected to extract the cotton farmland. Then, the cotton fields at the florescence, boll, and boll opening stages were extracted from the remote sensing images with eight bands at visible, near-infrared, and red edge ranges. Fourteen vegetation indexes (six canopy structure indexes and eight chlorophyll related indexes) under these bands were calculated via the Sequential Forward Selection (SFS) for different growth stages. SPXY (Sample set Partitioning based on joint X-Y distances) sample classification was selected to divide into the training and prediction set, in terms of the features and the cotton yields. The prediction models of cotton yield were constructed at different growth stages using Partial Least Squares Regression (PLSR). The best growth period was determined to compare the accuracy of the models for the cotton yield prediction. Finally, the independent samples were selected to verify the model. The prediction model was then applied to the extracted cotton planting area map to predict the distribution map of cotton yield. The results show that: 1) the RF was the best classification. The cropland-non cropland classification PA, UA, OA, and KC were 0.92, 0.96, 0.94, and 0.89, respectively, which were significantly better than those of SVM and CART. The PA and UA of cotton-non cotton field classification reached 0.95 and 0.87, respectively, whereas, the OA and KC were 0.92, and 0.83, respectively. The RF classified and actual areas of the cotton field were 60 400, and 64 866.7 hectares, respectively, with a relative error of 6.9 %. 2) The red edge band (B6) was set as the first selected feature in the three growth periods, indicating an excellent correlation with the yield, where the correlation coefficient (0.37, 0.47, and 0.53) increased with the three growth period. 3) The boll stage was the best growth stage for the cotton yield prediction (determination coefficient R2=0.62, root mean square error RMSE=625.5 kg/hm2, relative error RE=8.87%) using PLSR, which was better than that at boll opening (R2=0.51, RMSE=789.45 kg/hm2, RE=11.06%) and florescence (R2=0.48, RMSE=686.4 kg/hm2, RE=9.86%) stages. Consequently, the high-performance computing power was achieved by the GEE and Sentinel-2A image data, further determining the cotton yield prediction model at a regional scale. The finding can provide a new idea for the cotton planting area extraction and yield prediction, particularly for the monitoring of cotton crop growth in precision planting.