Cross-region plastic greenhouse segmentation and counting using the style transfer and dual-task networks

Abstract

Plastic Greenhouse (PG) is a vital component of protected agriculture, that can enhance crop yields and quality by altering the local microclimate conditions. Obtaining the number and distribution of PG based on high-resolution remote sensing images is important for agricultural policymaking. However, due to the variability and interconnection of PGs in high-resolution images, counting and mapping them accurately is challenging. In addition, variation in data distribution across regions restricts the reuse of the trained model beyond the study area, impeding the cross-regional application from the source region to the target region. To address this limitation, this paper proposes a novel Cross-Regional Segmentation and Counting framework (CRSC) that integrates the unsupervised Style Transfer Network (STNet) and dual task-based Segmentation Counting Network (SCNet) to perform simultaneous PG segmentation and counting across different regions. The STNet is adopted as a target-adaptive data augmentation strategy to generate fake training samples from the source training samples. The SCNet is designed as a dual-task structure to perform the PG segmentation and PG counts simultaneously. Specifically, for each target region, we use the STNet to generate the fake training samples and then train a SCNet based on both source training samples and fake training samples. Subsequently, we leverage the trained SCNet to predict the segmentation and counting results of PGs, yielding the PG map and counting for each target region. The validation results on five target regions indicate that our proposed CRSC framework can achieve stable improvement in cross-regional segmentation and counting of PGs, especially in cases where the labeled samples in the source region are limited.