The Stem Block (green section in Figure 2) serves as the entry point of TinyWeedNet, responsible for low-level feature extraction and spatial downsampling of the input image. It applies a single 3×3 convolution layer with stride s=2 and C0=24 output channels, followed by batch normalization and a ReLU activation in Equation 1:

This operation reduces the spatial resolution by a factor of two. The relatively narrow channel width (C₀ = 24) ensures a lightweight feature representation that minimizes computation in subsequent layers while maintaining sufficient expressive capacity for early-stage encoding. Such an compact stem design is particularly beneficial for TinyML deployment, as it balances representational richness and on-chip efficiency, enabling fast inference on low-power MCUs.

The MSC block (Figure 2 blue part) aims to enhance receptive-field diversity while maintaining computational efficiency. Natural scenes in agricultural environments often exhibit significant intra-class variations and complex backgrounds—e.g., overlapping leaves, varying weed sizes, and inconsistent lighting. To effectively handle these spatial complexities, the MSC employs four parallel convolutional branches with distinct receptive fields and a pooling pathway to capture features across multiple spatial scales.

Given the input feature map F0∈ℝ24×112×112 from the Stem block, the MSC block processes it through four concurrent branches shown in Equations 2–5:

Here, the first branch with 1 × 1 kernels performs local feature compaction and inter-channel mixing, the 3 × 3 and 5 × 5 branches extract mid- and large-scale spatial structures, and the fourth branch incorporates a 3×3 max-pooling operation to capture contextual cues such as background contrast before projection via a 1 × 1 convolution. All convolutional outputs are batch-normalized and activated using ReLU to ensure stable training and non-linear feature transformation.

Finally, the outputs of the four branches are concatenated along the channel dimension as follow in Equation 6:

To maintain parameter uniformity across hardware deployments, the total output channels are fixed at 36, with each branch contributing roughly one quarter of the total (as determined using the _make_divisible function in the implementation). This design allows the MSC to capture both fine-grained edge patterns and coarse-scale weed morphology while preserving the same spatial resolution as the Stem output. These kernel scales were empirically selected to capture the diverse weed morphologies ranging from fine leaf veins to broad canopy textures. By aggregating multi-scale features early in the network, the subsequent layers can focus on high-level abstraction without sacrificing low-level texture fidelity—an essential property for visual tasks executed on resource-constrained MCUs.

The Inverted Residual block constitutes the core feature transformation unit of TinyWeedNet, combining efficient depthwise separable convolutions with lightweight channel attention to achieve trade-off between high accuracy and efficiency (Figure 2 orange part). Each IR block adopts an “expand–depthwise–project” pattern, which was first introduced in MobileNetV2 and later optimized for edge deployment Sandler et al. (2018). Given an input feature map U∈ℝCin×H×W, the IR block expands the channel dimension by a factor of E=4, applies spatial filtering through a depthwise convolution, recalibrates the resulting features via a Channel Attention mechanism, and finally projects the output back to the target dimension Cout. The value of E is determined by the hyperparameter sensitivity analysis discussed later. Residual connections are included when the input and output tensors share identical shapes ( s=1 and Cin=Cout) to facilitate gradient propagation.

The complete transformation can be expressed as in Equations 7–10:

where Ze ∈ ℝ(4Cin)×H×W, Zd ∈ ℝ(4Cin)×(H/s)×(W/s), V ∈ ℝCout×(H/s)×(W/s).

When the stride s=1 and the channel dimensions remain unchanged, a residual connection is added shown in Equation 11:

This shortcut enhances gradient flow and stabilizes the optimization process, allowing deeper feature extraction without accuracy degradation. By operating primarily in the channel domain, the IR block minimizes the computational cost associated with conventional full convolutions, reducing the number of multiply–accumulate operations while retaining representational richness.

To further improve feature discrimination, a lightweight CA (R = 8) mechanism is embedded after the depthwise stage in each IR block (Figure 2 gray part). CA adaptively re-weights channel responses by modeling inter-channel dependencies using global pooling operations and a shared two-layer 1 × 1 convolutional bottleneck. The value of R is determined by the hyperparameter sensitivity analysis discussed later. For an input feature tensor Zd, the attention response is compress as in Equations 12 and 13:

where W1∈ℝCR×C and W2∈ℝC×CR are 1×1 convolutional layers functioning as an efficient two-layer perceptron, σ(·) denotes ReLU, and σg(·) is the sigmoid activation. The attended feature map is obtained as Za=Zd⊙s, where ⊙ represents channel-wise multiplication. With a reduction ratio R=8, this mechanism imposes minimal overhead while enabling the network to emphasize weed-related features (e.g., leaf shape, vein contrast) and suppress background noise.

TinyWeedNet stacks five IR blocks sequentially, progressively reducing the spatial resolution while expanding channel depth. The first four IR blocks employ a stride of s=2 for spatial downsampling, and the final block keeps s=1 to preserve high-level spatial context.

This hierarchical design enables gradual abstraction from low-level spatial textures to high-level semantic representations, ensuring robust weed background separation under varying field conditions. The embedded CA modules at each stage further enhance intra-class compactness and inter-class separability, leading to improved classification robustness on resource-constrained devices.

The Final Projection and Classification Head (Figure 2 pink part) aggregates the high-level representations extracted by the preceding IR blocks and transforms them into class-level predictions. After the last IR stage, the feature tensor FIR5∈ℝ120×7×7 undergoes a 1×1 convolutional projection that expands the channel dimension to Cf=240 for enhanced representational richness shown in Equation 14:

This 1 × 1 layer performs linear channel mixing and feature compression without altering spatial dimensions, allowing the network to consolidate multi-channel semantic descriptors from previous layers into a compact yet expressive feature space. Batch Normalization stabilizes the activation distribution, and the ReLU activation introduces non-linearity while preserving computational simplicity for MCU deployment.

Subsequently, global average pooling (GAP) aggregates the spatial information of Ff into a 240-dimensional feature vector shown in Equation 15:

This operation effectively converts each feature map into a single representative statistic, thus eliminating spatial redundancy and ensuring translation invariance—a desirable property for real-world weed imagery where plants appear at varying positions and scales.

To prevent overfitting and improve generalization, a dropout layer with a probability of p = 0.2 is applied before the final classifier. The resulting feature vector is then passed through a fully connected (FC) layer to produce the final prediction vector y^ over the nine target weed categories shown in Equation 16:

This minimalistic classification head provides an trade-off between inference latency and predictive accuracy. By combining global pooling with a single dense layer, TinyWeedNet avoids the parameter overhead of multi-layer classifiers while maintaining discriminative capability. Such a streamlined design, together with integer-quantization compatibility, ensures that the network can be deployed on low-power STM32 MCUs for real-time weed classification tasks.

Illumination changes are simulated using four photometric transformations: brightness adjustment, contrast scaling, gamma correction, and white-balance shift.

Brightness adjustment is defined as in Equation 17.

where β∈{0.10, 0.20, 0.30} controls the intensity of brightness variation after image normalization to [0,1]. These values are selected to represent mild to strong exposure changes commonly observed in outdoor agricultural environments.

Contrast scaling is formulated as in Equation 18.

where μ denotes the mean pixel intensity and α∈{0.8, 0.6, 0.4} represents progressively decreasing contrast levels. This range captures gradual contrast degradation caused by overcast weather or sensor limitations while avoiding unrealistic visual artifacts.

Gamma correction is applied as in Equation 19.

with γ∈{0.8, 1.2, 1.6}. These values are chosen to approximate non-linear illumination effects under shadowed and high-glare conditions frequently encountered in field imaging.

White-balance shift is modeled through channel-wise scaling as in Equation 20:

where (sR,sG,sB)∈{(1.1,1.0,0.9),(1.2,1.0,0.8),(1.3,1.0,0.7)}. These settings simulate color temperature variations caused by different sunlight spectra throughout the day.

Weather-related image degradation is approximated using Gaussian blur, which models reduced image sharpness due to motion, defocus, light rain, or atmospheric haze.

Gaussian blur is defined as in Equation 21.

where Gσ denotes a Gaussian kernel. Kernel sizes of {3, 5, 7} are employed to represent increasing levels of degradation severity. These values are selected to reflect realistic blur conditions encountered by mobile agricultural platforms, while preserving the semantic structure of the target objects.

To approximate variations in soil appearance and background color distributions without introducing new textures or datasets, a hue shift is applied in the HSV color space as in Equation 22:

where Δh∈{±5∘,±10∘,±15∘}. This perturbation serves as a lightweight proxy for changes in soil color, vegetation background, and regional color distributions, enabling an initial robustness assessment under controlled background-related domain shifts.

The expansion ratio E determines the internal channel width within each IR block, thus shaping the model’s feature capacity and non-linearity. When E increases from 3 to 6, both accuracy and F1-score improve steadily because wider intermediate layers capture more diverse spatial patterns. However, this benefit comes with a steep rise in computational cost and memory usage. For example, the configuration with E = 6 attains the best accuracy (97.29%) but incurs a ninefold increase in inference latency (118.34 ms versus 12.67 ms) and a 26-fold rise in flash memory compared with E = 3. Beyond a certain point, further expansion yields little accuracy gain but severely limits deployability on MCUs. In practice, a moderate setting of E = 4 offers the best balance—achieving 96–97% accuracy while keeping inference time below 90 ms and flash usage under 2 MB. This trend echoes prior findings on lightweight vision networks Sandler et al. (2018), where moderate expansion preserves accuracy without sacrificing efficiency.

The reduction ratio R in the channel attention (CA) module governs the degree of channel compression within the squeeze–excitation branch. A small R (e.g., 4) retains more intermediate channels and strengthens attention responses to subtle inter-class variations such as leaf texture or vein details. Yet, this comes at the expense of a larger parameter count and memory demand—for instance, increasing from 38.8 K parameters at R = 16 to 51.5 K at R = 4. In contrast, a large R (e.g., 16) overly compresses the feature descriptors, leading to minor but consistent accuracy drops (up to 2–3%). Across all configurations, R = 8 provides the most balanced outcome: F1-scores remain above 94% while parameters are reduced by roughly one quarter compared to R = 4. Configurations with R = 8 also show fast inference (12–115 ms) and stable energy use, confirming that moderate compression achieves the best trade-off between attention expressiveness and MCU-level efficiency. This observation further validates the lightweight attention strategy adopted in TinyWeedNet.

The number of stem channels S controls the capacity of the initial convolution to extract low-level edges and texture features. Increasing S from 8 to 24 leads to clear improvements in accuracy and F1-score (typically 1–3%), as richer early representations enhance gradient flow and feature discrimination. However, these benefits come with quadratic growth in both parameters and flash memory, from 0.196 MiB to 1.588 MiB, and require more data to be stored externally (up to 7.81 MiB), which adds energy overhead. Empirically, S = 24 represents a practical upper bound, offering strong feature extraction while staying within the 2 MB on-chip flash limit. Configurations beyond E = 6 or S = 24 exceeded STM32H7 storage capacity, emphasizing that architecture design in TinyML must be co-optimized with hardware constraints.

When jointly considering accuracy, latency, and on-chip memory limits, the configuration (E4 R8 S24) defining the proposed TinyWeedNet architecture demonstrates the most balanced performance among all 27 tested combinations. It achieves 97.26% accuracy and 96.64% F1-score while maintaining a moderate model size (1.815 MB) and an inference latency (89.40 ms) for agricultural robotic vision systems. Unlike typical brute-force neural architecture searches, this configuration was a priori derived from design principles emphasizing multi-scale diversity, moderate expansion, and lightweight attention. The subsequent sensitivity experiments validate that this combination yields the best accuracy–efficiency–energy synergy, with per-inference energy of 39.08 mJ - 8× lower than the heaviest configuration (ID 21). Furthermore, the relatively small standard deviations across all trials (≤1%) indicate statistical robustness, confirming that TinyWeedNet’s performance remains stable under random initialization and environmental noise.

Overall, the hyperparameter sensitivity analysis not only verifies TinyWeedNet’s architectural choices but also provides generalizable insights for TinyML model design. In particular, (1) moderate expansion (E = 4-5) prevents accuracy saturation while ensuring low latency, (2) mid-level attention reduction (R = 8) balances precision and overhead, and (3) a strong but bounded stem (S = 24) offers maximal texture sensitivity without memory overflow. These findings collectively demonstrate that the TinyWeedNet configuration is hardware-aware, energy-efficient, and empirically validated for MCU deployment in weed classification scenarios.

Removing the multi-scale convolution branch (“w/o MS Conv”) leads to a drop in accuracy (from 97.26% to 94.29%) and F1-score (from 96.64% to 91.96%), confirming that the multi-scale design (integrating kernels of 1 × 1, 3 × 3, and 5 × 5) is essential for capturing spatial features at multiple receptive-field resolutions. Interestingly, both inference time and energy consumption remain almost unchanged (Δ0.48 ms, Δ0.15 mJ), implying that the MS block introduces negligible computational cost relative to its gain in representational richness. In terms of parameter efficiency, the baseline achieves 204.8%/M accuracy-per-parameter (APP), while “w/o MS Conv” drops to 198.5%/M, highlighting that MS not only improves raw accuracy but also increases accuracy density per parameter. The energy-per-MACC also remains stable (≈1.25×10⁻¹⁰ J/MACC), demonstrating that the gain is architectural rather than computational.

Disabling the channel attention mechanism (“w/o Attention”) decreases F1-score to 90.75%, illustrating a loss of discriminative ability in feature reweighting. Although inference becomes slightly faster (86.9 ms) with lower energy consumption (38.0 mJ), this accuracy–efficiency trade-off is unfavorable for real deployments. Moreover, the variant unexpectedly increases model parameters (820K vs. 475K), confirming that the lightweight CA module not only strengthens feature expressiveness but also imposes a regularizing effect, allowing the backbone to remain compact without losing performance. From an energy–delay perspective, the Energy–Delay Product (EDP) improves only marginally (3.30×10³ vs. 3.49×10³ mJ·ms), while accuracy drops significantly, making the baseline configuration more Pareto-optimal. Overall, the CA module with reduction ratio R = 8 offers the best synergy between attention strength and hardware efficiency, ensuring both high accuracy and stable runtime.

Replacing all depthwise separable convolutions with standard 3×3 convolutions (“w/o DepthwiseConv”) results in a catastrophic increase in parameter count from 0.48 M to 5.10 M, inflating model size over tenfold (1.8 MB → 19.7 MB) and MACC operations by 5.13×. While the accuracy remains high (96.68%), the inference latency rises from 89.3 ms to 593.4 ms, and energy consumption surges from 39.1 mJ to 259.2 mJ. This translates to a 44× increase in EDP (1.54 × 105 mJ·ms), making this variant less attractive for MCU applications. The per-MACC energy also deteriorates (0.162 nJ/MACC vs. 0.125 nJ/MACC), implying that standard convolutions magnify memory traffic and data-movement overhead. This confirms that depthwise separable convolutions are the dominant contributor to TinyWeedNet’s computational efficiency, maintaining high accuracy while minimizing latency and energy cost.

Removing the final 1 × 1 projection layer before global average pooling (“w/o Final Conv1×1”) leads to an accuracy reduction from 97.26% to 94.69% with a slight decrease in model size (1.81 MB → 1.69 MB). This projection consolidates high-level semantic features and improves inter-class separability prior to classification. Its computational overhead is minimal (0.3ms difference) and energy consumption remains virtually unchanged (38.7 mJ), demonstrating that its inclusion provides substantial accuracy gains with negligible runtime penalty. Although its accuracy-per-parameter (212.8%/M) seems higher numerically due to the reduced parameter count, the absolute performance loss of over 2.5% indicates that this layer is essential for final-stage feature fusion and stable convergence.

The ablation experiments also reveal non-linear interactions among components:

These interactions highlight that removing modules based solely on FLOP reduction can be misleading in TinyML contexts; instead, optimizing for on-chip memory residency, energy-per-MACC stability, and accuracy-per-parameter density yields better deployment efficiency.