Mapping the Spatial Distribution of Tobacco using Multi-modal Satellite Imagery and Deep Learning

Accurate mapping of high-value economic crops, such as tobacco, in complex mountainous regions is essential for sustainable precision agriculture and regional land-use management. However, identifying tobacco plots remains challenging due to spectral confusion among objects and insufficient segmentation accuracy in complex terrains encountered in traditional tobacco remote sensing image semantic segmentation. This study presents a deep learning framework designed to overcome these limitations by synergizing Unmanned Aerial Vehicle (UAV) imagery with multi-temporal satellite data.

We propose a novel semantic segmentation model. Specifically, by introducing a channel-spatial attention module, we enhance the feature discrimination between tobacco plants and background crops/bare land; by incorporating an adaptive convolution module, we improve the model's adaptability to complex terrains. To validate the model's performance, a dedicated semantic segmentation dataset for tobacco remote sensing imagery was constructed. Results on this dataset demonstrate that the proposed model outperforms mainstream segmentation models such as U-Net and DeepLabv3+, achieving an improvement of 5% in mean Intersection over Union (mIoU).

The framework offers a scalable, automated solution for monitoring economic crops in heterogeneous environments, providing critical spatial intelligence for crop yield estimation and agricultural policy-making in challenging mountainous terrains.