Paper Details

Abstract

Due to the implementation of a relatively strict tobacco monopoly policy in China, every local government controls the tobacco planting area according to the allocated indicators. In the process of large-scale tobacco cultivation, how to quickly and cost-effectively predict the area of pre-planted tobacco by local tobacco farmers every year before seedling transplanting, timely detect phenomena of over-planting or under-planting, and taking corresponding remedial measures, is an important topic that needs to be studied. This work intends to comprehensively utilize spatial sampling, remote sensing interpretation, and deep learning techniques for estimating the large-scale tobacco planting area. Firstly, it combines Support Vector Machines (SVM) and medium-resolution multispectral satellite remote sensing images to estimate the area of film-mulched farmlands across the whole experimental regions, and calculates the proportion of film-mulched farmland through the existing farmland layer. Secondly, the whole experimental regions are divided into groups based on township administrative boundaries, and the most representative townships are selected for further analysis through cluster sampling. Then, within the representative township areas, spline sampling was conducted, and existing tobacco cultivation monitoring grids were selected as the spline sampling areas. An extraction method for film-mulched tobacco fields (FMTF) based on an improved Unet network (EMFMTFIUN) is proposed. The EMFMTFIUN network and high-resolution satellite images were used to extract FMTF in the spline sampling areas, distinguishing between FMTF and non- FMTF, and then calculating the proportion of FMTF in representative townships. Finally, based on the total area of farmlands, the proportion of film-mulched farmlands in the whole experimental regions, and the proportion of FMTF in representative townships, the area of FMTF and the total tobacco planting area in the whole experimental regions were calculated, and a corresponding SaaS (Software as a Service) cloud service platform for tobacco planting area estimation is constructed. Taking Guangze County in Fujian Province as an example, experimental results compared with other 7 representative algorithm models (DeeplabV3+, Hrnet, Pspnet, SegFormer, Attention-Unet, ResUnet and Unet++), using the EMFMTFIUN model to extract FMTF in the spline sampling regions (monitoring grids) in the county from 2022 to 2025 achieves higher accuracy (with an average of over 97.0%). The average estimation accuracy of the total tobacco cultivation area in the county from 2022 to 2025 reaches 96.91%, with an average estimation time consumption of 118.8 seconds. This effectively improves the accuracy and efficiency of large-scale tobacco planting area estimation, meets the relevant requirements of local tobacco management departments, and has good promotion and application value.

To Cite This Article