BS-Mamba for Black-Soil Area Detection On the Qinghai-Tibetan Plateau

TL;DR

This work tackles detecting extremely degraded black-soil grasslands on the Qinghai-Tibetan Plateau from UAV imagery. It introduces the QTP-BS dataset and a novel BS-Mamba network that fuses a Spatial Mamba Block-based long-range context model with convolutional features in a two-branch encoder. The model optimizes with a combined loss ${L}=\lambda_{1}\mathcal{L}_{CE}+\lambda_{2}\mathcal{L}_{IoU}$ and predicts two-channel maps ${Y}=[Y^{blk},Y^{mat}]$ for black-soil patches and mattic epipedon, achieving state-of-the-art results on two test sets. The work enables accurate assessment of black-soil extent to guide restoration planning, and the dataset and code are publicly available.

Abstract

Extremely degraded grassland on the Qinghai-Tibetan Plateau (QTP) presents a significant environmental challenge due to overgrazing, climate change, and rodent activity, which degrade vegetation cover and soil quality. These extremely degraded grassland on QTP, commonly referred to as black-soil area, require accurate assessment to guide effective restoration efforts. In this paper, we present a newly created QTP black-soil dataset, annotated under expert guidance. We introduce a novel neural network model, BS-Mamba, specifically designed for the black-soil area detection using UAV remote sensing imagery. The BS-Mamba model demonstrates higher accuracy in identifying black-soil area across two independent test datasets than the state-of-the-art models. This research contributes to grassland restoration by providing an efficient method for assessing the extent of black-soil area on the QTP.

Figures (11)

• Figure 1: Comparison of our BS-Mamba model and a variant of Mamba, Mamba-Unet, in detecting black-soil areas. In both gravel-type and poisonous-plant-type regions, our BS-Mamba model consistently outperforms Mamba-Unet. Detection challenges include omissions, inaccuracies, and unclear boundaries. These issues arise mainly due to the similar color tones between black-soil patches and mattic epipedon (first row) and the high diversity of invasive weeds (second row). The results demonstrate that our BS-Mamba model effectively mitigates these challenges.
• Figure 2: The BS-Mamba network utilizes a spatial Mamba block (SMB) to capture long-range dependencies and a convolutional block to retain spatial structure. The two-branch encoder provides the decoder with skip connections from each branch, enabling it to capture both long-range dependencies and fine-grained texture details in black-soil patches and the mattic epipedon.
• Figure 3: Four scanning strategies are illustrated: (a) 'H' indicates horizontal scanning. (b) 'V' represents vertical scanning. (c) '2' is scanning within a local $2 \times 2$ window. (d) '2F' is scanning in the flipped direction within a $2 \times 2$ window.
• Figure 4: Scanning strategies in four stages of BS-Mamba: each stage contains two SMBs, with each SMB's SSM configured to scan sequence tokens in four distinct directions. In each stage, the first row represents the scanning configuration for the first SSM, while the second row corresponds to the second SSM in each stage.
• Figure 5: Different black-soil types of weeds and poisonous plants: The first row showcases original UAV-captured images highlighting diverse weeds and poisonous plants in black-soil areas. The second row shows the ground truth labels. The third row presents the segmentation results generated by the BS-Mamba model, demonstrating its ability to accurately delineate these challenging features.