摘要
Abstract
In areas with complex planting structures and fragmented plots,conducting fine-scale monitoring of crop planting structures using medium and high-resolution remote sensing images still faces challenges such as mixed pixel interference and spectral similarity among crops.To systematically evaluate the applicability of different models in such scenarios,Xinjin District of Chengdu City was selected as the study area.Based on multi-source time-series images from Sentinel-2 and Landsat 8/9,the spectral-temporal characteristics of overwintering(small spring)and spring-sown(large spring)crops were analyzed across the blue,green,red,and near-infrared bands.Subsequently,four models-TempCNN,LSTM,Transformer,and Random Forest(RF)-were constructed to evaluate and compare their performance in planting structure extraction.The results show that TempCNN not only has high patch integrity and good spatial continuity in the identification of major crops such as wheat,rapeseed,rice,and corn,but also has significantly higher overall classification accuracy than other models.Although LSTM and Transformer perform well in some crop categories and have relatively complete patch recognition,their overall accuracy is still lower than that of TempCNN.In contrast,the extraction performance of RF is significantly inferior to the other three deep learning models.In conclusion,among the evaluated models,TempCNN performs best in extracting planting structures using medium-and high-resolution remote sensing images in areas with fragmented plots and complex planting structures.关键词
中高分辨率影像/遥感提取方法/种植结构/多源数据/时间序列/最佳算法Key words
medium and high resolution images/remote sensing extraction method/planting structures/multi-source data/time series/optimal algorithm分类
农业科技
