Cropland Extraction Method Of Remote Sensing Image Based On Multi-source Data Fusion

Cropland is an important type of land cover,and refined cropland extraction is of great significance to agricultural production management.Remote sensing technology can provide timely and accurate observation data in cropland areas.However,there are differences in the imaging mechanisms of different remote sensing sensors,which makes it difficult for the existing models to be applied to different image data sources at the same time,and the generalization ability is poor.In addition,remote sensing data of a single time phase can only obtain cropland information at a certain time,and cannot make full use of the phenological characteristics during cropland planting,so multi-temporal continuous observation is required to improve the accuracy of cropland extraction.Since complete multi-temporal data is difficult to obtain,studying a suitable time window selection method is an important basis for the research of multi-temporal cropland extraction algorithm.At present,deep learning technology is widely used in the field of remote sensing,and it has become an increasingly important trend to explore cropland extraction methods based on deep learning and multi-source remote sensing image fusion.However,in the extraction of cropland with multi-source data fusion,there are still problems such as insufficient training sample data and weak generalization ability of network model,which involve key technologies such as window selection of image acquisition time of multisource data and structural design of network model.In order to solve these problems,this paper proposes a cropland extraction model based on multi-source data fusion,which aims to use single-phase high-resolution aerial images and multi-temporal low-resolution satellite images to improve the accuracy of cropland extraction while controlling the cost of remote sensing data acquisition.The main work is as follows:(1)In order to determine the optimal time and quantity of satellite image acquisition for multi-source data fusion,this paper collected single-time aerial images and multitemporal Sentinel-2 images of the same area during the crop growing season.By traversing all the combinations of multi-source data,the Random Forest(RF)method was used to extract cropland for each data combination,and multiple classification accuracy evaluation indicators were used for accuracy assessment.Combined with the correspond-ding classification effect of single-time Sentinel-2 satellite data,the optimal acquisition time window of the image in multi-source data fusion was analyzed.The analysis results show that the "a single aerial image and dual-temporal satellite images during the crop growing season" can be overlaid as the best time window for image acquisition,which can balance the cost of acquiring multi-source data and the accuracy of cropland extraction well.(2)Facing the cropland extraction task based on multi-source data fusion,this paper designs a multi-source data fusion module and a pyramid feature recalibration module,combines the optimal acquisition time window of images in multi-source data fusion,constructs a three-input multi-source data fusion cropland extraction network MOUDCS,and uses the combined loss function lovasz-ce to train the network model and optimize the classification results for multiple evaluation indicators,so as to solve the problems of model prediction bias caused by the imbalance of spatial distribution of cropland areas.This enables high-resolution fine extraction of cropland.(3)To address the issue of insufficient training sample dataset in multiple-source data fusion-based cropland extraction tasks,this paper collected large-scale remote sensing image datasets from both domestic and international sources,and constructed two multi-source data fusion cropland sample datasets(Fujian dataset and AI4 Boundaries dataset).Verification experiments on the MOUDCS network model were conducted on these datasets.The experimental results show that the proposed method achieved high cropland extraction accuracy on different datasets,with the classification results of cropland in the Fujian dataset reaching an Io U index of 0.7466,which is 6.96% higher than the baseline(cropland classification using only aerial images),and m Io U of 0.8611,which is 3.96% higher.On the AI4 Boundaries dataset,the classification results for cropland reached an Io U index of 0.7513 and an m Io U of 0.7971,which improved by1.08% and 1.05%,respectively,compared to the baseline.Furthermore,through comparison with current mainstream multimodal data fusion methods and attention mechanism methods,the effectiveness of the proposed method was further verified.