MixerSENet: Lightweight HSI Classifier

• MixerSENet is a lightweight framework for hyperspectral image classification that decouples spatial and channel mixing using PCA-reduced patches, point-wise and multi-scale depth-wise convolutions, and a squeeze-and-excitation block.
• It achieves competitive accuracy with significantly fewer parameters and lower computational costs compared to traditional 3D-CNNs and transformer-based architectures.
• The model maintains constant spatial resolution through patch-based processing, preserving local structure and yielding coherent classification maps even in challenging scenarios.

Searching arXiv for MixerSENet and closely related mixer/SE references. I’ll look up the cited arXiv entries to ground the article in current paper metadata. MixerSENet is a lightweight framework for hyperspectral image (HSI) classification that is designed to address computational efficiency and limited labeled data by processing fixed-size hyperspectral patches, explicitly decoupling spatial and channel mixing, and incorporating a squeeze-and-excitation (SE) block for channel-wise recalibration (Alkhatib et al., 1 Jun 2026). In the reported formulation, the model operates on PCA-reduced HSI patches of size $9 \times 9 \times 15$, preserves spatial resolution throughout the network, and combines $1\times1$ point-wise convolutions with multi-scale depth-wise convolutions to model spectral–spatial structure while keeping the parameter count low (Alkhatib et al., 1 Jun 2026).

1. Problem formulation and design rationale

An HSI is represented as a 3D data cube

$\mathbf{I}_{\text{Original} \in \mathbb{R}^{H \times W \times C},$

where $H$ and $W$ denote spatial dimensions and $C$ denotes the number of spectral bands, often exceeding 100 (Alkhatib et al., 1 Jun 2026). The classification task assigns a land-cover class label to each pixel. In the patch-based setting used by MixerSENet, a spatial neighborhood around a target pixel, such as a $9 \times 9$ patch, is extracted and used to predict the central pixel’s label (Alkhatib et al., 1 Jun 2026).

The motivation for MixerSENet is grounded in four stated difficulties of HSI classification: the spectral dimension is very high; labeled data are scarce and expensive to obtain; models must capture joint spectral–spatial structure; and computational resources may be limited in airborne, satellite, or edge-device deployments (Alkhatib et al., 1 Jun 2026). The paper positions existing model families accordingly. It states that 1D-CNNs focus only on spectral sequences, 2D-CNNs on spatial context, and 3D-CNNs and hybrids can capture joint spectral–spatial information but tend to be parameter- and compute-heavy and to overfit on small HSI datasets (Alkhatib et al., 1 Jun 2026). Transformer-based HSI models such as HSIFormer and SimPoolFormer are described as having high parameter counts, long inference times, and substantial labeled-data requirements, while Mamba-like models such as MorphMamba are characterized as computationally intensive in token processing and morphological operations, with relatively high inference time despite moderate parameter counts (Alkhatib et al., 1 Jun 2026).

Within that context, MixerSENet is presented as a mixer-style network inspired by MLP-Mixer and PolSARConvMixer, with efficient decoupling of spatial and channel mixing, depth-wise convolutions for inexpensive spatial mixing, point-wise convolutions for spectral mixing, and an SE block to emphasize informative spectral channels (Alkhatib et al., 1 Jun 2026). A plausible implication is that the architecture is intended not merely as a smaller substitute for 3D-CNNs or transformers, but as a reformulation of HSI classification around fixed-resolution local context and low-cost spectral–spatial operators.

2. Architectural organization

The reported high-level pipeline begins with PCA along the spectral dimension: $\mathbf{I}_{\text{Reduced} \in \mathbb{R}^{H \times W \times P}, \quad P \ll C,$ with $P=15$ in the experiments (Alkhatib et al., 1 Jun 2026). Overlapping patches of size $S \times S$, here $1\times1$0, are then extracted around each labeled pixel to form

$1\times1$1

This tensor is passed through a stack of repeated Mixer blocks, followed by an SE block and a classification head that outputs class probabilities for the central pixel (Alkhatib et al., 1 Jun 2026).

A central architectural property is constant spatial resolution. Depth-wise convolutions use padding, and there are no pooling or strided operations, so all intermediate feature maps retain spatial size $1\times1$2 (Alkhatib et al., 1 Jun 2026). The paper explicitly links this choice to preservation of spatial locality and continued access by the classifier to the full spatial neighborhood around the target pixel (Alkhatib et al., 1 Jun 2026). This suggests that MixerSENet prioritizes local structural fidelity over the hierarchical spatial abstraction commonly used in encoder-style vision backbones.

The model also explicitly separates channel and spatial processing. Channel mixing is performed by $1\times1$3 point-wise convolutions acting as per-pixel affine transformations across channels, in spirit similar to the channel MLP of MLP-Mixer; spatial mixing is performed by depth-wise convolutions per channel, with kernel sizes $1\times1$4, $1\times1$5, and $1\times1$6 to provide multi-scale receptive fields (Alkhatib et al., 1 Jun 2026). The paper describes this as analogous to alternating token-MLPs and channel-MLPs in MLP-Mixer, but implemented through convolutional operators that are more efficient on images and more natural for HSI data (Alkhatib et al., 1 Jun 2026).

3. Constituent blocks and mathematical formulation

After PCA and patch extraction, the initial tensor is

$1\times1$7

An initial point-wise convolution produces

$1\times1$8

with per-location affine transformation

$1\times1$9

where $\mathbf{I}_{\text{Original} \in \mathbb{R}^{H \times W \times C},$0 and $\mathbf{I}_{\text{Original} \in \mathbb{R}^{H \times W \times C},$1 (Alkhatib et al., 1 Jun 2026). The paper interprets this operation as a learnable spectral projection into a feature space of dimension $\mathbf{I}_{\text{Original} \in \mathbb{R}^{H \times W \times C},$2 (Alkhatib et al., 1 Jun 2026).

Spatial mixing is performed by depth-wise convolution. For channel $\mathbf{I}_{\text{Original} \in \mathbb{R}^{H \times W \times C},$3,

$\mathbf{I}_{\text{Original} \in \mathbb{R}^{H \times W \times C},$4

and with kernel $\mathbf{I}_{\text{Original} \in \mathbb{R}^{H \times W \times C},$5 of size $\mathbf{I}_{\text{Original} \in \mathbb{R}^{H \times W \times C},$6,

$\mathbf{I}_{\text{Original} \in \mathbb{R}^{H \times W \times C},$7

In MixerSENet, depth-wise convolutions with kernel sizes $\mathbf{I}_{\text{Original} \in \mathbb{R}^{H \times W \times C},$8, $\mathbf{I}_{\text{Original} \in \mathbb{R}^{H \times W \times C},$9, and $H$0 are applied as multi-scale spatial filters, while preserving the number of channels $H$1 because each input channel has its own filter (Alkhatib et al., 1 Jun 2026). The paper emphasizes that this extracts channel-specific spatial patterns at multiple scales with substantially fewer parameters than full 2D convolution (Alkhatib et al., 1 Jun 2026).

After spatial mixing, a second $H$2 convolution performs channel fusion: $H$3 A Mixer block is therefore summarized as

$H$4

$H$5

$H$6

and the network applies this transformation repeatedly: $H$7 with $H$8 treated as a key hyperparameter in the ablation study (Alkhatib et al., 1 Jun 2026).

The SE block refines the output feature maps by channel-wise attention. For each channel $H$9, with $W$0 and $W$1, the squeeze operation is global average pooling: $W$2 yielding

$W$3

The excitation stage is

$W$4

where $W$5, $W$6, $W$7 is a reduction ratio, and $W$8 is sigmoid (Alkhatib et al., 1 Jun 2026). Recalibration then applies

$W$9

or

$C$0

The paper states that this emphasizes channels that contribute more to discrimination and suppresses noisy or redundant spectral bands; for HSIs, where many bands are correlated, the SE block helps focus on informative spectral combinations learned by earlier convolutions (Alkhatib et al., 1 Jun 2026).

The variant without the SE block is termed MixerNet, while the full model is MixerSENet (Alkhatib et al., 1 Jun 2026). After SE recalibration, the final tensor

$C$1

is flattened and passed to a dense classifier: $C$2

$C$3

$C$4

with standard cross-entropy loss against ground-truth labels (Alkhatib et al., 1 Jun 2026).

4. Efficiency profile and training protocol

MixerSENet is explicitly characterized as lightweight. The paper reports 53,146 parameters for MixerSENet and 52,050 for MixerNet (Alkhatib et al., 1 Jun 2026). Comparative parameter counts are given as 397,586 for 3D-CNN, 142,690 for HybridKAN, 1,373,084 for HSIFormer, 771,122 for SimPoolFormer, and 67,650 for MorphMamba (Alkhatib et al., 1 Jun 2026). The interpretation provided in the paper is that MixerSENet is one of the smallest models among the baselines, with only MorphMamba in a similar range but with higher inference time and lower accuracy (Alkhatib et al., 1 Jun 2026).

The complexity table reports for MixerSENet: FLOPs $C$5, MACs $C$6, and inference time 2:32 (Alkhatib et al., 1 Jun 2026). The comparison values listed in the paper include 3D-CNN with FLOPs $C$7 and inference time 1:37; HSIFormer with FLOPs $C$8 and inference time 11:00; SimPoolFormer with FLOPs $C$9 and inference time 4:52; MorphMamba with FLOPs $9 \times 9$0 and inference time 7:32; and HybridKAN with FLOPs $9 \times 9$1 and inference time 3:25 (Alkhatib et al., 1 Jun 2026). All reported inference times were measured on Windows 10 with 64 GB RAM and an NVIDIA GeForce RTX 2080 GPU with 8 GB VRAM (Alkhatib et al., 1 Jun 2026).

The paper attributes efficiency to several specific architectural choices: depth-wise convolutions in place of full 2D or 3D convolutions, $9 \times 9$2 convolutions for linear-cost channel mixing, constant spatial resolution without pyramid or down/up-sampling machinery, shallow depth with diminishing returns beyond modest $9 \times 9$3, and PCA preprocessing that reduces spectral dimensionality from $9 \times 9$4 to $9 \times 9$5 (Alkhatib et al., 1 Jun 2026). A plausible implication is that the architecture is optimized around fixed local windows rather than global scene modeling, which is consistent with the patch-based workflow.

Training uses $9 \times 9$6 patches and 15 PCA components, Adam optimizer, learning rate $9 \times 9$7, batch size 32, and up to 100 epochs with early stopping based on validation accuracy (Alkhatib et al., 1 Jun 2026). The early-stopping rule checks validation accuracy after each epoch, stops training if validation accuracy does not improve for 10 consecutive epochs, and restores the weights with best validation accuracy (Alkhatib et al., 1 Jun 2026). The implementation is reported in Keras with TensorFlow backend (Alkhatib et al., 1 Jun 2026). The paper notes PCA, small model size, early stopping, and patch-based sampling as the effective overfitting controls, and explicitly states that dropout, weight decay, and normalization layers are not mentioned (Alkhatib et al., 1 Jun 2026).

5. Experimental protocol and empirical performance

The experiments use two benchmark datasets. Houston13 is the 2013 IEEE GRSS Data Fusion Contest dataset, used here in HSI-only form, with 15 land-cover classes including Healthy grass, Stressed grass, Synthetic grass, Tree, Soil, Water, Residential, Commercial, Road, Highway, Railway, Parking Lot1, Parking Lot2, Tennis court, and Running track (Alkhatib et al., 1 Jun 2026). Available training data are evenly split into training and validation, while the provided test set remains unchanged, giving approximate proportions of about 9% training, 9% validation, and 82% testing (Alkhatib et al., 1 Jun 2026). The QUH-Qingyun dataset is an urban HSI dataset with 6 classes: Trees, Concrete building, Car, Ironhide building, Plastic playground, and Asphalt road (Alkhatib et al., 1 Jun 2026). Its patches are randomly split into 5% training, 5% validation, and 90% testing (Alkhatib et al., 1 Jun 2026).

Evaluation uses Overall Accuracy (OA),

$9 \times 9$8

Average Accuracy (AA),

$9 \times 9$9

and the Kappa coefficient,

$\mathbf{I}_{\text{Reduced} \in \mathbb{R}^{H \times W \times P}, \quad P \ll C,$0

reported as $\mathbf{I}_{\text{Reduced} \in \mathbb{R}^{H \times W \times P}, \quad P \ll C,$1 (Alkhatib et al., 1 Jun 2026). Experiments are repeated 10 times with random splits, and mean $\mathbf{I}_{\text{Reduced} \in \mathbb{R}^{H \times W \times P}, \quad P \ll C,$2 standard deviation are reported; the best single run’s per-class accuracies are also tabulated (Alkhatib et al., 1 Jun 2026).

On Houston13, MixerSENet achieves OA $\mathbf{I}_{\text{Reduced} \in \mathbb{R}^{H \times W \times P}, \quad P \ll C,$3, AA $\mathbf{I}_{\text{Reduced} \in \mathbb{R}^{H \times W \times P}, \quad P \ll C,$4, and Kappa $\mathbf{I}_{\text{Reduced} \in \mathbb{R}^{H \times W \times P}, \quad P \ll C,$5 $\mathbf{I}_{\text{Reduced} \in \mathbb{R}^{H \times W \times P}, \quad P \ll C,$6 (Alkhatib et al., 1 Jun 2026). The reported baseline values are OA $\mathbf{I}_{\text{Reduced} \in \mathbb{R}^{H \times W \times P}, \quad P \ll C,$7 for 3D-CNN, $\mathbf{I}_{\text{Reduced} \in \mathbb{R}^{H \times W \times P}, \quad P \ll C,$8 for HybridKAN, $\mathbf{I}_{\text{Reduced} \in \mathbb{R}^{H \times W \times P}, \quad P \ll C,$9 for HSIFormer, $P=15$0 for SimPoolFormer, $P=15$1 for MorphMamba, and $P=15$2 for MixerNet (Alkhatib et al., 1 Jun 2026). The paper therefore states that MixerSENet improves OA over 3D-CNN by about 2.3 percentage points, outperforms the listed transformer-based baselines and MorphMamba, and improves over MixerNet by +1.24 OA, +0.79 AA, and +1.40 Kappa (Alkhatib et al., 1 Jun 2026). Reported class-wise highlights for MixerSENet include Synthetic Grass at 99.80%, Soil at 100.00%, Residential at 94.31%, Commercial at 78.35%, Road at 77.43%, Parking Lot1 at 86.07%, and Running Track at 99.37% (Alkhatib et al., 1 Jun 2026).

On QUH-Qingyun, MixerSENet achieves OA $P=15$3, AA $P=15$4, and Kappa $P=15$5 $P=15$6 (Alkhatib et al., 1 Jun 2026). The comparison values are OA $P=15$7 for 3D-CNN, $P=15$8 for HybridKAN, $P=15$9 for HSIFormer, $S \times S$0 for SimPoolFormer, $S \times S$1 for MorphMamba, and $S \times S$2 for MixerNet (Alkhatib et al., 1 Jun 2026). The paper states that MixerSENet achieves the best OA, AA, and Kappa among the compared methods on Qingyun (Alkhatib et al., 1 Jun 2026). Class-wise results noted in the paper include Car at 83.32% versus 61.53% for 3D-CNN and 50.80% for MorphMamba, Ironhide building at 99.81%, and Asphalt road at 96.06% (Alkhatib et al., 1 Jun 2026).

A concise summary of the headline results is as follows.

Dataset	MixerSENet result	Reported comparison context
Houston13	OA $S \times S$3; AA $S \times S$4; Kappa $S \times S$5	Higher OA than 3D-CNN, HybridKAN, HSIFormer, SimPoolFormer, MorphMamba, and MixerNet
QUH-Qingyun	OA $S \times S$6; AA $S \times S$7; Kappa $S \times S$8	Best OA, AA, and Kappa among the compared methods

The paper also reports qualitative classification maps for Houston13 and Qingyun. It states that MixerSENet produces smoother and more coherent maps, cleaner class boundaries, and fewer salt-and-pepper artifacts than the baselines, especially in residential and commercial areas in Houston13 and in roads, cars, and building roofs in Qingyun (Alkhatib et al., 1 Jun 2026). In Houston13, clouds obscure some regions in the RGB image, yet the paper reports relatively consistent labeling in cloud-affected areas (Alkhatib et al., 1 Jun 2026).

6. Ablations, interpretation, and relation to similarly named work

The principal ablations concern network depth $S \times S$9 and the SE block. On Qingyun with 1% training data, the depth study reports: $1\times1$00 gives OA 92.80, AA 86.08, Kappa 90.46, with 11,590 parameters; $1\times1$01 gives OA 93.96, AA 89.45, Kappa 92.18, with 51,270 parameters; and $1\times1$02 gives a slight drop to OA 93.85, AA 89.18, Kappa 91.93, with 61,190 parameters (Alkhatib et al., 1 Jun 2026). The paper’s interpretation is that increasing depth up to $1\times1$03 steadily improves performance, while $1\times1$04 shows mild overfitting, and parameter count grows linearly with $1\times1$05 (Alkhatib et al., 1 Jun 2026).

The SE ablation compares MixerNet and MixerSENet directly. On Houston13, the SE-equipped model gains +1.24 OA, +0.79 AA, and +1.40 Kappa; on Qingyun, it gains +0.66 OA, +0.92 AA, and +0.88 Kappa (Alkhatib et al., 1 Jun 2026). The paper interprets these results as evidence that channel reweighting is valuable in hyperspectral spectral–spatial modeling (Alkhatib et al., 1 Jun 2026). It further attributes the overall performance of MixerSENet to effective spectral–spatial feature modeling via decoupled mixing, reduced overfitting due to compact architecture and PCA, and channel attention that mitigates redundancy in spectral bands and highlights subtle spectral signatures (Alkhatib et al., 1 Jun 2026).

A common point of confusion is the resemblance between the names MixerSENet and SEMixer. These are distinct architectures for different modalities and tasks. MixerSENet is an HSI classification model based on point-wise and depth-wise convolutions plus an SE block (Alkhatib et al., 1 Jun 2026). By contrast, SEMixer is a long-term time series forecasting architecture built around an all-MLP mixer backbone with a Random Attention Mechanism (RAM) and a Multiscale Progressive Mixing Chain (MPMC) (Zhang et al., 18 Feb 2026). The similarity lies chiefly in the shared mixer-oriented design vocabulary and in the use of a semantics-enhancing component: in MixerSENet this role is played by channel-wise SE recalibration, whereas in SEMixer it is played by RAM operating over time patches (Alkhatib et al., 1 Jun 2026, Zhang et al., 18 Feb 2026). This suggests a broader pattern in contemporary mixer-style architectures, namely the augmentation of efficient mixing backbones with lightweight reweighting or semantics-enhancement modules tailored to the structure of the target modality.

Reported limitations of MixerSENet are also specific. The paper notes that depth-wise convolution can still struggle with highly correlated spectral bands, despite point-wise convolution and SE alleviating some issues; that the model is patch-based and scales to large scenes by sliding window, which is standard but may be slow for very large images; and that future work may investigate more expressive spectral mixing strategies, carefully controlled efficient transformers, and applications to multi-temporal HSI, other sensors, and fusion scenarios (Alkhatib et al., 1 Jun 2026). These points delimit the scope of the method: its contribution is an efficient local spectral–spatial classifier, not a complete solution to large-scene or globally contextual HSI modeling.