基于DINOv3动态LoRA微调的牧场多组分生物量估算

1. Introduction

Pasture biomass estimation is a critical task in livestock production systems and precision agriculture. Accurate knowledge of the quantity and composition of available feed directly influences decisions on stocking rates, rotational grazing schedules, and pasture recovery management. Grazing systems cover approximately 25% of the global land surface [1], and suboptimal feed management leads to economic losses and environmental degradation, including soil erosion, biodiversity loss, and carbon depletion.

Existing methods for biomass quantification each present significant limitations. The destructive clip-and-weigh technique, while accurate, is labor-intensive and infeasible at scale. Non-destructive proximal sensors such as rising plate meters and capacitance probes offer faster readings but suffer from poor reliability under heterogeneous vegetation conditions and cannot differentiate biomass by species or condition [1]. Satellite-based remote sensing enables broad-scale monitoring but lacks the spatial resolution to capture fine-grained compositional information and still requires manual calibration.

The specific problem addressed in this paper is the simultaneous estimation of five pasture biomass components from high-resolution top-view RGB images under conditions of limited labeled data, non-standard image geometry, and significant distributional variation across geographic regions and seasons. This multi-output regression task is complicated by heavily right-skewed target distributions, zero-inflated clover measurements, and a training set of only 357 annotated images spanning 19 sites across four Australian states.

Large pretrained Vision Transformers (ViTs) such as DINOv3 [2] offer powerful visual representations learned from billions of images. However, deploying these models on small downstream datasets presents a fundamental tension: full fine-tuning updates hundreds of millions of parameters from fewer than 400 labeled examples, making overfitting nearly inevitable, while keeping the backbone entirely frozen forgoes the opportunity to adapt representations to domain-specific patterns such as vegetation density, canopy texture, and species-level discrimination. This tension is especially acute for ViTs, whose self-attention layers learn highly distributed representations that are sensitive to even small perturbations in early layers.

Low-Rank Adaptation (LoRA) [3] offers a principled middle ground by constraining weight updates to low-rank subspaces, dramatically reducing trainable parameter count while preserving the pretrained structure. However, standard LoRA applies a uniform rank across all layers, treating early and late transformer blocks identically. This ignores a well-established property of deep networks: lower layers encode generic, broadly transferable features (edges, textures, color statistics), while deeper layers encode abstract, task-specific representations [4]. In a low-data regime, uniform adaptation allocates capacity to early layers that may need little change and may under-adapt late layers where domain-specific patterns must be learned.

The central question of this paper is: how should adaptation capacity be distributed across transformer layers when fine-tuning a large pretrained ViT under severe data scarcity? We propose a dynamic, depth-aware LoRA schedule in which the rank r and scaling factor $\alpha$𝛼 grow exponentially from the first to the last transformer layer. This provides smooth, continuous control over the degree of adaptation: near-zero perturbation of early layers that encode reusable visual primitives, and progressively stronger specialization in deeper layers where task-relevant compositional patterns must be formed. We observe that this design choice yields a consistent improvement in weighted $R^{2}$𝑅 2 over uniform LoRA, and that it helps preserve pretrained feature quality that otherwise degrades under full fine-tuning and aggressive uniform adaptation in low-data settings.

To instantiate and evaluate this adaptation strategy, we build a complete pipeline for pasture biomass estimation that additionally addresses two practical challenges. First, the dataset images have a 2:1 aspect ratio that conflicts with the square input assumption of pretrained ViTs; we partition each image into two square halves processed as a dual-view stream with a contrastive alignment loss enforcing feature coherence. Second, we introduce a progressive resolution training strategy in which LoRA adapters are trained and merged at successively higher resolutions ($384 \rightarrow 512 \rightarrow 768$384→512→768), which stabilizes the optimization landscape for adapter training. The final system incorporates metadata integration, test time augmentation, and multi backbone ensembling as further supporting components.

The primary contribution of this work is the depth-aware LoRA adaptation scheme and the accompanying empirical analysis demonstrating its advantages over uniform LoRA, full finetuning, and frozen baselines for adapting large ViTs under data scarcity. As secondary contributions, we introduce the dual-view spatial partitioning strategy for handling rectangular imagery and the progressive resolution training curriculum that stabilizes adapter optimization. The complete framework is evaluated on the CSIRO Image2Biomass benchmark [1], where it achieves a cross-validated weighted $R^{2}$𝑅 2 of 0.81.

The remainder of this paper is organized as follows. Section 2 reviews related work on biomass estimation, vision transformers, and parameter-efficient finetuning. Section 3 describes the proposed methodology. Section 4 presents the experimental setup. Section 5 reports quantitative results and ablation studies. Section 6 discusses findings and limitations. Section 7 concludes.

2. Related Work

2.1. Pasture Biomass Estimation

Traditional approaches to pasture biomass estimation fall into three broad categories: destructive sampling, proximal sensing, and remote sensing. Destructive harvesting remains the gold standard for accuracy but is prohibitively expensive and cannot scale beyond small experimental plots [1]. Proximal sensors such as rising plate meters, capacitance probes, and active optical sensors provide faster non destructive readings, but their accuracy degrades under heterogeneous canopy structures and varying moisture conditions [5]. Satellite and UAV-based remote sensing offers spatial coverage but typically operates at coarse resolution and cannot separate biomass by species or phenological state [6].

More recently, image-based approaches using convolutional neural networks have been explored for biomass estimation from UAV-RGB imagery [7,8]. These works primarily employ standard architectures such as AlexNet and ResNet on aerial orthoimages, treating biomass prediction as a single output regression task. Semi-supervised methods using synthetic data and approximate labels have also been proposed to mitigate the annotation bottleneck [9]. More recently, generative AI assisted preprocessing has been proposed as a complementary route to the same problem: Rana et al. [10] use prompt-guided generative image reconstruction to convert visually heterogeneous vegetation imagery into annotation ready representations that preserve plant morphology while reducing background complexity, facilitating the creation of higher-quality training data for downstream models. This preprocessing stage direction is complementary to the adaptation stage contribution of the present work. However, prior work has largely focused on total biomass rather than component level estimation, and few studies have leveraged modern foundation models or addressed the specific challenges of non standard image geometry in pasture imagery.

Deep learning approaches to pasture and forage biomass estimation remain an active area: Cândido et al. [11] integrate proximal and satellite remote sensing with machine learning for pasture biomass estimation; Azubuike et al. [12] apply transfer learning and stacking ensembles to smartphone imagery across commercial dairy farms; and Mandal [13] recently benchmarks a systematic range of vision foundation model backbones, from EfficientNet-B3 to DINOv3 ViT-Large, together with several cross view fusion mechanisms, on the same CSIRO Image2Biomass dataset used in this work. These studies confirm that CNN- and transformer based deep learning is an active and growing direction for non-destructive biomass estimation, while also underscoring the gap our work addresses: parameter-efficient adaptation of large foundation models under extreme data scarcity (357 training images) and non standard image geometry, conditions that remain largely unaddressed in this recent literature.

2.2. Vision Transformers and Self-Supervised Representation Learning

Vision Transformers (ViTs) [14] have emerged as a dominant architecture for visual representation learning, particularly when combined with self-supervised pretraining on large scale datasets. The DINO family of methods [2,15,16] demonstrates that self-supervised ViTs trained with self distillation objectives produce general purpose visual features that transfer effectively to diverse downstream tasks including classification, segmentation, and depth estimation, often without any finetuning. DINOv2 [15] was trained on a curated dataset of 142 million images and showed strong linear probing and k-NN performance. DINOv3 [2] scales this further to 1.7 billion images and a 7 billion parameter teacher model, introduces Gram anchoring to address dense feature degradation, and produces improved dense and global representations across over sixty benchmarks. In dense linear probing of the frozen backbone, DINOv3 reaches 55.9 mIoU on ADE20k and 86.6 mIoU on PASCAL VOC, compared to 49.5 and 83.1 for DINOv2 [2], a gap the authors attribute to Gram anchoring’s correction of patch level feature degradation that otherwise emerges under long self-supervised training schedules.

The strength of these pretrained features makes them attractive for domains with limited labeled data, such as agricultural imagery. However, directly deploying frozen features may perform poorly when the target domain differs substantially from the pretraining distribution, as is the case for close range pasture photographs that exhibit fine grained texture variation not well represented in standard image datasets. This motivates the need for controlled adaptation strategies that can bridge the domain gap without degrading the pretrained representation.

2.3. Parameter-Efficient Fine Tuning and LoRA for Vision Transformers

Full finetuning of large pretrained models on small downstream datasets is computationally expensive and prone to overfitting. This problem is particularly acute for ViTs, whose self attention layers learn globally distributed representations: perturbations to early layers can propagate through the network and degrade overall feature quality [4]. Parameter-efficient finetuning (PEFT) methods address this by introducing a small number of trainable parameters while keeping the majority of the pretrained model frozen.

LoRA [3] constrains weight updates to low-rank decompositions $\Delta W = B A$Δ 𝑊=𝐵 𝐴, where $B \in \mathbb{R}^{d \times r}$𝐵∈ℝ 𝑑×𝑟, $A \in \mathbb{R}^{r \times k}$𝐴∈ℝ 𝑟×𝑘, and $r \ll min \left(\right. d , k \left.\right)$𝑟≪min(𝑑,𝑘). This reduces the parameter count by orders of magnitude while maintaining task performance and introduces no additional inference latency since the low rank updates can be merged into the original weights. Originally proposed for language models, LoRA has since been applied to ViTs for image classification [3], medical imaging [17], and other visual tasks.

However, standard LoRA applies a uniform rank across all layers, which does not account for the well documented functional specialization of transformer layers: lower layers capture generic, transferable features while deeper layers capture increasingly task specific representations [4,18]. In low data settings, uniform adaptation may allocate unnecessary capacity to early layers where the pretrained representation is already appropriate, while under adapting deeper layers where domain specific patterns must be learned. Recent work has explored layer wise learning rate schedules and selective layer unfreezing as alternatives, but these approaches provide coarse, discrete control compared to the smooth, continuous modulation possible with a rank schedule.

Our work introduces a depth aware LoRA scheme that addresses this gap. By increasing the rank and scaling factor with layer depth, we achieve an allocation of adaptation capacity that is better matched to the layer wise feature hierarchy of ViTs. This approach appears especially beneficial when the training set is extremely small, where each degree of freedom in the adaptation must be carefully managed to avoid overfitting.

Despite the availability of strong pretrained vision models and efficient adaptation techniques, there remains a gap in their application to agricultural biomass estimation under the specific constraints of non standard image geometry, limited data, and multi output regression. Our work addresses this gap by combining DINOv3 with dynamic depth-aware LoRA, dual view spatial partitioning, and progressive resolution training.

3. Methodology

The overall pipeline consists of five components: (1) dual-view spatial partitioning of rectangular images, (2) feature extraction using a DINOv3 backbone with dynamic depth aware LoRA adaptation, (3) contrastive alignment of dual-view embeddings, (4) metadata integration via a dense sub network, and (5) multi output regression with Huber loss. The system is trained using a progressive resolution curriculum and evaluated through five fold stratified group cross validation. An overview of the architecture is shown in Figure 1.

Figure 1. Overview of the proposed pipeline. The input image ($2000 \times 1000$2000×1000) is split into left and right square halves, each patchified and passed through a shared DINOv3 backbone with dynamic depth aware LoRA. As shown in the inset, the LoRA rank and scaling factor follow $r_{ℓ} = 2^{1 + 8 \left(\right. ℓ - 1 \left.\right) / \left(\right. L - 1 \left.\right)}$𝑟 ℓ=2 1+8(ℓ−1)/(𝐿−1) and $\alpha_{ℓ} = 2^{2 + 7 \left(\right. ℓ - 1 \left.\right) / \left(\right. L - 1 \left.\right)}$𝛼 ℓ=2 2+7(ℓ−1)/(𝐿−1), ranging from $r = 2 , \alpha = 4$𝑟=2,𝛼=4 at the first layer to $r = 512 , \alpha = 512$𝑟=512,𝛼=512 at the deepest layer. The pooled embeddings from both streams are concatenated and mapped to five biomass predictions (Grass, Clover, Dry Dead, GDM, Total).

3.1. Dual-View Spatial Partitioning

The input images have dimensions of approximately $2000 \times 1000$2000×1000 pixels (width × height), representing a $70 \times 30$70×30 cm pasture quadrat captured from above. DINOv3 models are pretrained on square images with a fixed patch size (14 or 16 pixels), and directly resizing a 2:1 aspect ratio image to a square format would compress the horizontal axis and distort spatial relationships between vegetation features.

Instead, each image $\mathcal{I} \in \mathbb{R}^{3 \times H \times W}$ is partitioned along the width dimension into two equal square halves:

$$
\mathcal{I}^{L} = \mathcal{I} \left[\right. : , : , 0 : W / 2 \left]\right. , \mathcal{I}^{R} = \mathcal{I} \left[\right. : , : , W / 2 : W \left]\right. .
$$

(1)

Both halves are resized to the target resolution (e.g., $768 \times 768$) and processed through the shared DINOv3 backbone in the same batch. The resulting CLS token embeddings $\mathbf{f}^{L} , \mathbf{f}^{R} \in \mathbb{R}^{d}$ are concatenated to form the combined visual representation $\mathbf{f}_{vis} = \left[\right. \mathbf{f}^{L} ; \mathbf{f}^{R} \left]\right. \in \mathbb{R}^{2 d}$.

3.2. Backbone and Dynamic Depth-Aware LoRA Adaptation

We use DINOv3 ViT-Large (304 M parameters, $d = 1024$, $L = 24$ layers) and DINOv3 ViT-Huge (632 M parameters, $d = 1280$, $L = 32$ layers) as backbone models. Both are initialized from publicly available self-supervised checkpoints and kept largely frozen during training.

3.2.1. The Case for LoRA over Full Fine-Tuning

With only 357 training images and backbone sizes exceeding 300 M parameters, full finetuning creates a severely over parameterized optimization problem. In our experiments, full finetuning of even the last four transformer blocks produced unstable training dynamics and degraded validation performance relative to a frozen baseline (CV $R^{2}$ of 0.70 vs. 0.72). This is consistent with the broader observation that self-supervised ViT representations are holistic: each layer’s output depends on the representations learned in all preceding layers, so perturbation of early layers may disrupt downstream feature representations. LoRA constrains updates to low-rank subspaces, providing plasticity for domain adaptation while limiting the risk of catastrophic forgetting of pretrained features.

3.2.2. Why Uniform LoRA Is Insufficient

Standard LoRA applies the same rank r and scaling factor $\alpha$ to every layer. While this is a reasonable default when data is abundant, it becomes less effective in low data settings for two reasons. First, early layers in DINOv3 encode generic visual primitives (edges, local textures, color statistics) that are already well suited to pasture imagery; adapting these layers consumes trainable capacity that could be better allo