TMPS: Target-Aware Metric Learning

• The paper demonstrates that prioritized sampling of scarce target-domain samples boosts metric learning performance, yielding a 7.3 macro F1 improvement over source-only baselines.
• TMPS integrates target-domain emphasis into embedding optimization, effectively mitigating domain gaps in applications like plant disease diagnosis.
• The framework optimally tunes the sampling probability (p ≈ 0.7) to balance source diversity and avoid overfitting, making it adaptable to various fine-grained classification tasks.

Target-Aware Metric Learning with Prioritized Sampling (TMPS) is a metric learning framework designed for high-robustness in scenarios where access to labeled target-domain data is extremely limited. The paradigm, as formalized for plant disease diagnosis, employs prioritized sampling of scarce target-domain examples during metric embedding optimization to bridge domain gaps that standard classification or metric learning methods cannot address efficiently (Nogami et al., 14 Oct 2025).

1. The TMPS Framework

TMPS formalizes metric learning for domain adaptation where the standard pipeline fails due to environmental, contextual, or acquisition shifts between source and deployment domains. It operates by mapping images $x$ to low-dimensional embeddings $f(x)$ and enforcing similarity structure via Euclidean distances in the embedding space. The core innovation is a prioritized sampling scheme: when constructing the comparison set for metric loss computation, each class representative $x_i$ is drawn from the target-domain sample pool $X^t$ with probability $p$ and from the source-domain pool $X^s$ with probability $1-p$. This mechanism enables the feature space to adapt explicitly to the structure of the target domain and leverages limited target data maximally.

The similarity distribution over $c$ classes for an input $x$ is defined as:

$$P(x; x_1,\ldots,x_c)_i = \frac{\exp(-\|f(x) - f(x_i)\|^2)}{\sum_{j=1}^c \exp(-\|f(x) - f(x_j)\|^2)}$$

The metric loss is the cross-entropy between the ideal one-hot label vector $I(x)$ and this distribution:

$$L(x; x_1,\ldots,x_c) = H_{CE}(I(x), P(x; x_1,\ldots,x_c))$$

This loss is incorporated as a regularizer into the full model objective. Prioritized sampling is formalized as:

$$x_i = \begin{cases} x_i^{t} & \text{with probability}~p,~x_i^{t} \in X^t \ x_i^{s} & \text{with probability}~1-p,~x_i^{s} \in X^s \end{cases}$$

2. Implementation and Algorithmic Details

The practical implementation follows these steps:

1. Gather large labeled data $X^s$ for training and a limited labeled target-domain set $X^t$.
2. Use a backbone network (EfficientNetV2-S, pre-trained; inputs resized to $512 \times 512$ px).
3. For each training sample, sample $x_i$ for each class according to the prioritized probability $p$.
4. Compute $P(x; x_1,\ldots,x_c)$ and $L(x; x_1,\ldots,x_c)$, then combine with the standard classification objective.
5. Optimize jointly.
6. The selection probability $p$ is tuned (optimal $p \approx 0.7$ reported).

The tuning of $p$ is crucial: low $p$ reverts to standard metric learning, high $p$ risks overfitting due to limited target diversity.

3. Experimental Setup and Key Results

TMPS was validated using a dataset of 223,073 images covering 21 diseases (plus healthy controls) from three crop species and 23 agricultural fields. The target setting had only 10 labeled images per disease from the deployment domain, whereas source images were abundant.

TMPS was compared against several baselines:

• Standard source-only training
• Conventional metric learning without prioritized sampling
• Combined-data models (“All-Train”)
• Fine-tuning on target-domain data

Results:

• TMPS achieved an average macro F1 improvement of 7.3 points over the source-only baseline and 3.6 points over fine-tuned models, with 18.7 (over baseline) and 17.1 (over conventional metric learning) point gains on specific configurations.
• Prioritized sampling (p ≈ 0.7) was crucial for optimal adaptation; lower and higher values diminished effectiveness due to underutilization or overfitting.

4. Mechanistic Insights and Rationale

The TMPS framework's underlying hypothesis is that even extremely small sets of target-domain samples, if strategically prioritized in the metric learning process, can anchor the embedding space to key deployment conditions. Unlike conventional approaches that blend source and target data or fine-tune on target examples, TMPS specifically regularizes the feature space to be sensitive to target samples throughout training—mitigating the domain gap effect without sacrificing the diversity of source information.

This approach offers robustness against “domain gap” phenomena, including differences in leaf morphology, symptoms, and backgrounds in plant disease imaging.

5. Applications Beyond Plant Disease Diagnosis

The TMPS methodology is applicable wherever labeled target data is rare and domain shift is pronounced:

• Medical imaging (inter-institutional variance)
• Remote sensing (sensor/environment differences)
• Industrial visual inspection (changing production conditions)
• Mobile health applications (personal device adaptation)

Any fine-grained classification scenario subject to large environment-induced distributional shifts can potentially benefit.

6. Relationship to Related Sampling and Metric Learning Techniques

TMPS is distinguished from classical metric learning by its explicit, probabilistic prioritization of target-domain samples in constructing the optimization pairs. This principled sampling rule extends beyond random or hard-negative selection techniques and contrasts with methods that simply combine or fine-tune on limited target data. Unlike proxy-based methods or adaptive loss modulations, TMPS targets domain adaptation directly via sample selection probability.

7. Future Directions

Possible future directions for TMPS include:

• Developing a principled or theoretically-guided mechanism for tuning $p$ as a function of data diversity or estimated domain gap.
• Integrating advanced data augmentation (including generative approaches) to enrich both source and target samples.
• Evaluating TMPS under varying degrees of domain shift and with alternative backbone architectures.
• Generalizing TMPS to more complex scenarios, such as multi-domain or continually evolving target domains.

---

In summary, Target-Aware Metric Learning with Prioritized Sampling (TMPS) formalizes a robust, flexible, and effective strategy for metric-based domain adaptation. By leveraging prioritized inclusion of scarce target-domain samples, it substantially improves classification metrics in the face of domain gaps and limited deployment data, as empirically demonstrated for large-scale plant disease diagnosis (Nogami et al., 14 Oct 2025).