Paper deep dive
A Parameter-efficient Convolutional Approach for Weed Detection in Multispectral Aerial Imagery
Leo Thomas Ramos, Angel D. Sappa
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 7/20/2026, 4:39:28 AM
Summary
The paper introduces FCBNet, a parameter-efficient deep learning model for weed segmentation in multispectral aerial imagery. It utilizes a fully frozen ConvNeXt backbone to minimize trainable parameters by over 90%, integrated with a novel Feature Correction Block (FCB) to refine features for a lightweight FPN-based decoder. Evaluated on WeedBananaCOD and WeedMap datasets, FCBNet achieves mean Intersection over Union (mIoU) exceeding 85%, outperforming models like U-Net, DeepLabV3+, and SegFormer, while demonstrating superior computational efficiency with training times between 0.06 and 0.2 hours.
Entities (10)
Relation Signals (9)
FCBNet → usesbackbone → ConvNeXt
confidence 95% · The architecture is based on a fully frozen ConvNeXt backbone
FCBNet → containscomponent → Feature Correction Block
confidence 92% · the proposed Feature Correction Block (FCB), which leverages efficient convolutions for feature refinement
FCBNet → achievesmetric → mIoU
confidence 90% · showing that FCBNet outperforms models... in terms of mIoU, exceeding 85%
FCBNet → evaluatedon → WeedBananaCOD
confidence 90% · FCBNet is evaluated on the WeedBananaCOD and WeedMap datasets
FCBNet → evaluatedon → WeedMap
confidence 90% · FCBNet is evaluated on the WeedBananaCOD and WeedMap datasets
FCBNet → usesdecoder → Feature Pyramid Network
confidence 88% · integrated with a Feature Pyramid Network (FPN)-based decoder
FCBNet → outperforms → SegFormer
confidence 85% · showing that FCBNet outperforms models such as... SegFormer... in terms of mIoU
FCBNet → outperforms → DeepLabV3+
confidence 85% · showing that FCBNet outperforms models such as... DeepLabV3+... in terms of mIoU
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We introduce FCBNet, an efficient model designed for weed segmentation. The architecture is based on a fully frozen ConvNeXt backbone, the proposed Feature Correction Block (FCB), which leverages efficient convolutions for feature refinement, and a lightweight decoder. FCBNet is evaluated on the WeedBananaCOD and WeedMap datasets under both RGB and multispectral modalities, showing that FCBNet outperforms models such as U-Net, DeepLabV3+, SK-U-Net, SegFormer, and WeedSense in terms of mIoU, exceeding 85%, while also achieving superior computational efficiency, requiring only 0.06 to 0.2 hours for training. Furthermore, the frozen backbone strategy reduces the number of trainable parameters by more than 90%, significantly lowering memory requirements.
Tags
Links
- Source: https://arxiv.org/abs/2603.06655v1
- Canonical: https://arxiv.org/abs/2603.06655v1
Trouble viewing inline? Open PDF directly →
Full Text
54,319 characters extracted from source content.
Expand or collapse full text
A Parameter-efficient Convolutional Approach for Weed Detection in Multispectral Aerial Imagery Leo Thomas Ramos1,2 (ltramos@cvc.uab.cat) Angel D. Sappa1,2,3 (sappa@ieee.org) 1Computer Vision Center 2Universitat Autònoma de Barcelona 3ESPOL Polytechnic University Abstract We introduce FCBNet, an efficient model designed for weed segmentation. The architecture is based on a fully frozen ConvNeXt backbone, the proposed Feature Correction Block (FCB), which leverages efficient convolutions for feature refinement, and a lightweight decoder. FCBNet is evaluated on the WeedBananaCOD and WeedMap datasets under both RGB and multispectral modalities, showing that FCBNet outperforms models such as U-Net, DeepLabV3+, SK-U-Net, SegFormer, and WeedSense in terms of mIoU, exceeding 85%, while also achieving superior computational efficiency, requiring only 0.06 to 0.2 hours for training. Furthermore, the frozen backbone strategy reduces the number of trainable parameters by more than 90%, significantly lowering memory requirements. Code available at: hidden_for_review 1 Introduction Weeds are unwanted plant species that grow alongside crops and compete for essential resources such as nutrients, water, sunlight, and space [22]. As a result, weeds are recognized as one of the primary factors contributing to yield reduction across a wide range of cropping systems [22, 5], reducing overall agricultural efficiency. At a broader scale, these effects result in substantial economic losses and negatively impact productivity [28], profitability [23], and the stability of supply chains [21]. Traditional detection relies on manual field inspection [6], where farmers visually identify weed presence across cultivated areas. However, agricultural fields often extend over large areas, making this process labor-intensive [21], time-consuming [19], and difficult to perform consistently [23]. This reduces the frequency and spatial coverage of inspections, increasing the risk of delayed or incomplete detection [12]. Accordingly, this lack of precise monitoring affects subsequent management decisions [34, 1], since effective treatment depends on accurate knowledge of weed location and distribution within the field [1]. To address these limitations, Deep Learning (DL)-based approaches have gained significant attention [21, 42], particularly semantic segmentation methods [21, 15]. These models are trained on large datasets and learn discriminative patterns from visual features such as colour, texture, and shape [24]. Consequently, they enable automatic identification and precise delineation of weeds at the pixel level [18], providing detailed information about their presence and spatial distribution. This facilitates large-scale analysis and supports more accurate and consistent monitoring compared to manual inspection [32]. Nonetheless, DL models face practical limitations. Achieving high accuracy commonly requires complex architectures with millions of trainable parameters [2], resulting in high computational and memory requirements during training [28, 27]. Also, deployment conditions impose strict constraints, since the models are often required to operate in open and remote agricultural environments [28] without access to high-performance computing infrastructure [14], and are typically integrated into platforms such as UAVs and drones with limited capacity [35]. Furthermore, multispectral/multimodal imagery is frequently employed in this domain [36, 31, 7, 37], which increases the computational requirements and data processing complexity. Several strategies exist to obtain efficient models, but they have specific trade-offs. Model distillation transfers the knowledge from a large model to a smaller one with reduced latency [8], but it often requires a pre-trained high-capacity teacher and an distillation stage that introduces non-negligible overhead [4, 39]. Transfer learning adapts representations learned from a source domain to a target domain [45, 20], yet segmentation often requires fine-tuning a significant portion of the model [46, 13], since features for classification do not preserve the spatial and boundary information required for mask reconstruction [43, 44]. Likewise, model freezing reduces memory usage and the number of trainable parameters [40], but introduces a mismatch between fixed feature representations and the reconstruction requirements of the decoder. This often requires introducing heavy adaptation blocks to achieve competitive accuracy [11], which undermines the intended efficiency of using a frozen backbone. Therefore, developing accurate yet efficient models remains an open research challenge. Based on the above, we propose a DL approach for weed detection, aiming to achieve high accuracy while maintaining training efficiency and low latency. To this end, we adopt the model freezing paradigm, as it is one of the most effective methods for reducing the number of trainable parameters and computational cost. Hence, we employ a fully-frozen ConvNeXt [17] backbone integrated with a Feature Pyramid Network [16] (FPN)-based decoder for segmentation. To handle the existing mismatch, we introduce Feature Correction Blocks (FCBs), which are inserted after each ConvNeXt extraction stage to refine the features passed to the decoder. These blocks are composed of pointwise and depthwise convolutions to preserve efficiency, along with group normalization to stabilize training independently of batch size. The proposed model is evaluated on two aerial image datasets for weed detection, covering both RGB and multispectral scenarios. The results demonstrate superior performance compared to established and state-of-the-art models. Furthermore, the use of ConvNeXt enables leveraging its stage-based design to incorporate the same number of FCBs across all its variants, regardless of their complexity. Consequently, a reduction of more than 90% in the number of trainable parameters is achieved, while maintaining superior performance and low latency. The contributions of this work are as follows: • We propose FCBNet, an efficient model for weed detection. • We introduce the Feature Correction Block, which refines feature representations from a frozen backbone with minimal computational cost. • We show that the proposed freezing strategy reduces the number of trainable parameters by more than 90%. • We show that FCBNet achieves low latency and fast training without compromising segmentation performance. 2 Related work Numerous DL approaches for weed detection have been developed in recent years. Increasing attention has been given to model performance, although not without important trade-offs. For instance, [9] proposes a semi supervised adversarial framework based on a U-Net with a ResNet50 backbone and a convolutional discriminator, enabling the use of unlabelled UAV RGB images through pseudo labelling. The method reduces annotation requirements, but the training process requires additional iterations. The work by [35] uses a densely connected feed forward Deep Neural Network (DNN) combining handcrafted color and texture descriptors from UAV RGB images to improve weed discrimination. While it achieves high performance, the DNN training involves strenuous and prolonged processing. Similarly, [37] proposes CGS-YOLO for multispectral weed detection that incorporates CARAFE upsampling, attention mechanisms, and an additional small target detection layer to improve small instance identification. While this improves discrimination capability, parameter count and complexity increase due to the additional head and feature fusion operations. In the same way, [31] proposes a ResNet-50-backboned U-Net combined with a fully connected Conditional Random Field (CRF) for boundary refinement. The method introduces a composite 3-channel input derived from spectral transformations to reduce computational cost. However, the CRF refinement increases post processing complexity. Likewise, [10] proposes a multi-class detection approach using U-Net and hybrid variants with pre-trained encoders (VGG19, MobileNetV2, and InceptionResNetV2). While deeper backbones achieve higher performance, they substantially increase model size and computational cost. Other methods focus on improving efficiency to enable real-world deployment, although this often introduces performance limitations. The work by [27], for example, proposes a weed detection system based on superpixel segmentation and lightweight CNN classifiers, targeting deployment on embedded devices such as Raspberry Pi. It explores custom shallow architectures, pruning, and quantization to reduce model size and latency. However, aggressive quantization introduced accuracy degradation. SC-Net [15] uses a U-Net based model with multi-scale convolutional blocks and attention based feature fusion to improve weed localization while reducing computational cost. Strip convolutions enlarges the receptive field with lower complexity, enabling efficient feature extraction. However, this degrades shallow feature representation and reduce performance in small regions, while the model size remains non-trivial. Similarly, [47] proposes a U-Net network with a reduced VGG backbone to improve computational efficiency. The model is pre-trained using synthetic images and later fine-tuned on real data, enabling accurate, low-latency segmentation. However, this limits feature extraction, reducing suitability for complex scenarios. As seen, progress is driven either by strategies and components that push identification performance, or by efficiency oriented designs intended to reduce deployment cost. However, accuracy gains are often accompanied by increased training time or parameter count that raise resource demands. Conversely, approaches that prioritize compactness can reduce latency and memory footprint, but usually exhibit performance degradation under challenging conditions, highlighting the persistent trade-off between computational efficiency and robust weed identification. 3 Model design The proposed model follows an encoder-decoder design, as shown in Fig. 1. It uses a fully frozen ConvNeXt backbone that produces four multi-scale feature maps, and inserts FCBs after each extraction stage to refine and correct the features. The corrected stage features are then integrated by a lightweight FPN-based decoder operating over the same four levels, which reconstructs a high-resolution representation. This representation is subsequently processed by a compact segmentation head to generate the final segmentation mask. We refer to this model as FCBNet. Below, we present each component of the architecture in detail. Figure 1: Overview of FCBNet proposed in this work. 3.1 Encoder FCBNet’s encoder is based on ConvNeXt as its underlying backbone. ConvNeXt is a Convolutional Neural Network (CNN) that builds upon ResNet50 by introducing a stage-wise design, layer normalization, GELU activations, and larger convolutional kernels (7×7). This enables effective hierarchical representation learning with controlled computational cost. A ConvNeXt block can be seen in Fig. 2. Figure 2: Comparison between ResNet, Swin Transformer, and ConvNeXt blocks. The selection of ConvNeXt in this work is motivated by its architectural design. ConvNeXt was developed with the goal of achieving performance comparable to Transformers, which currently represent the state-of-the-art in computer vision. However, Transformers exhibit quadratic complexity and require large amounts of data and training time to achieve strong performance. In response, ConvNeXt modernizes conventional CNNs through a Transformer-inspired design while preserving the efficiency advantages of CNNs. Table 1: ConvNeXt variants and their details. Model Channels (per stage) Depths (per stage) Params (M) ConvNeXt-tiny [96, 192, 384, 768] [3, 3, 9, 3] 28 ConvNeXt-small [96, 192, 384, 768] [3, 3, 27, 3] 50 ConvNeXt-base [128, 256, 512, 1024] [3, 3, 27, 3] 89 ConvNeXt-large [192, 384, 768, 1536] [3, 3, 27, 3] 198 Based on the above, ConvNeXt is selected as the feature extractor of FCBNet, as it provides an effective balance between the representation capability of Transformers and the computational efficiency of CNNs, which is consistent with the objective of our approach. ConvNeXt is available in four variants, as shown in Table 1. In this work, ConvNeXt-base (Fig. 3) is used as the default configuration of FCBNet. Figure 3: ConvNeXt-base architecture. Adapted from: [26] Feature Correction Block As mentioned, ConvNeXt is an efficient and high-performance architecture well-suited for feature extraction. However, it has a non-trivial number of parameters, particularly in its larger variants, which can limit its suitability for deployment in resource-constrained environments. Furthermore, a segmentation model also requires the inclusion of a decoder and a segmentation head, which further increases the overall model complexity. To maximize the efficient use of ConvNeXt, we employ it in a fully frozen configuration. This results in a substantial reduction in the number of trainable parameters, but also causes the feature representations to remain optimized for the original task and unable to be refined during training for the new domain. This, in turn, causes the decoder to operate on suboptimal feature representations, limiting its ability to accurately reconstruct the segmentation mask. To address the mismatch between the fixed encoder features and the decoder requirements, we introduce the Feature Correction Block (FCB). FCB is a lightweight residual module designed to operate on convolutional feature maps. Its objective is to refine and adapt the features extracted by the backbone, enabling more effective integration by the decoder while preserving computational efficiency. A FCB comprises seven layers in a bottleneck-like structure: a Pointwise convolution (PWConv), followed by Group Normalization (GroupNorm) with GELU activation, a Depthwise convolution (DWConv), a second GroupNorm + GELU, and a final PWConv. In addition, each FCB incorporates a residual connection that links the input and output of each block, that allows it to introduce corrective modifications while preserving the original feature information. PWConv and DWConv are a design choice for computational efficiency. PWConv (1×1) enables channel-wise feature projection with minimal overhead, while DWConv captures spatial context at significantly lower cost than standard convolutions. Likewise, GroupNorm is used to stabilize optimization without relying on large batch sizes, which is particularly useful under resource constrained training settings. Mathematically, FCB implements a transformation f(⋅)f(·) over the input tensor x∈ℝB×C×H×Wx ^B× C× H× W, as in Eq. 1: y=x+αf(x),y=x+α f(x), (1) where f(⋅)f(·) operates in a lower dimensional embedded space, applies spatial processing, and produces a correction term that is added to the input through a residual connection scaled by the learnable parameter α∈ℝα . The transformation f(x)f(x) follows a bottleneck structure composed of efficient convolutional operations. First, the input tensor with C channels is projected into an intermediate space of dimensionality C′C defined in Eq. 2: C′=max(Cmin,⌊Cr⌋),C = (C_ , Cr ), (2) where r corresponds to the bottleneck ratio controlling the channel reduction, and CminC_ is a lower bound imposed to ensure sufficient representational capacity. This projection is implemented using a bias free 1×1 convolution, followed by GroupNorm and a GELU activation, as in Eq. 3: z1=ϕ(GroupNorm1(Conv1×1C→C′(x))),z_1=φ (GroupNorm_1 (Conv_1× 1^C→ C (x) ) ), (3) where, ϕ(⋅)φ(·) denotes the GELU activation function. Next, a DWConv with kernel size k×k× k is applied to model spatial context efficiently, again followed by GroupNorm and GELU, as shown in Eq. 4: z2=ϕ(GroupNorm2(DWConvk×k(z1))).z_2=φ (GroupNorm_2 (DWConv_k× k(z_1) ) ). (4) Finally, the representation is projected back to the original channel dimensionality using another 1×11× 1 convolution, as in Eq. 5: f(x)=Conv1×1C′→C(z2).f(x)=Conv_1× 1^C → C (z_2 ). (5) In our implementation, the number of GroupNorm groups is dynamically selected to exactly divide the number of channels C′C , preventing invalid configurations when C′C is small or not divisible by the preferred number of groups. The entire refining process is achieved using lightweight operations, avoiding the heavy parameter and computational overhead of alternative mechanisms like attention modules. Within the frozen ConvNeXt encoder, one FCB is inserted immediately after each feature extraction stage, allowing progressive correction of the multi-scale representations. Since all ConvNeXt variants share the same four-stage hierarchical structure, the number of FCBs remains constant regardless of the backbone complexity. 3.2 Decoder and head To fuse the features extracted by the encoder, we employ a lightweight FPN-based decoder. It integrates four multi-scale feature maps into a single, high-resolution representation optimized for segmentation. Additionally, the design is streamlined to maintain computational efficiency. Specifically, it includes four 1×1 convolutions that project the feature maps from each encoder stage into a shared channel dimensionality. These are followed by four smoothing blocks, each consisting of a 3×3 convolution, BatchNorm, and GELU activation. These blocks are used during the top-down fusion process, where lower resolution features are upsampled using bilinear interpolation and combined with higher resolution features through element wise summation. Two additional smoothing blocks form a refinement stage applied at the highest resolution to further consolidate the decoded features before prediction. Figure 4: Structure of the Feature Correction Block. (a) Smoothing block. (b) Segmentation head. Figure 5: Smoothing block and head used in FCBNet decoder. Finally, a compact segmentation head converts the decoder features into class predictions. The head uses a 3×3 convolution followed by BatchNorm and GELU, applies dropout for regularization, and then uses a final 1×1 convolution to map features to the desired number of output channels. The resulting logits are upsampled to the original input resolution using bilinear interpolation, producing the final full-resolution segmentation mask. 4 Experiments 4.1 Datasets and preprocessing To evaluate the proposed model, two aerial imagery datasets, WeedBananaCOD [38] and WeedMap [30], for weed detection are selected. WeedBananaCOD [38] is an aerial dataset for camouflaged weed detection. The images were acquired over an approximately 4.3 hectare banana field located in the coastal region of Ecuador. The site is characterized by dense and homogeneous vegetation, creating challenging conditions where weeds are often visually camouflaged within the surrounding crop. Images were captured using a MicaSense RedEdge-M camera mounted on a DJI Mavic 2 Pro UAV. The dataset is multispectral, including RGB and near infrared (NIR) bands, and comprises a tile of size 5571×5855 pixels. The images were divided into 512×512 patches using the official preprocessing code provided by the authors. WeedMap [30] is one of the best-known datasets for weed detection. It was acquired in sugar beet fields located in Switzerland and Germany. It consists of two subsets: one captured with a RedEdge-M camera mounted on a DJI Inspire 2 UAV, providing five spectral bands (RGB, NIR, and red edge (RE)), and another by a Sequoia camera mounted on a DJI Mavic Pro UAV with RGB-NIR configuration. In this work, the RedEdge-M subset111Note that WeedMap contains multiple semantic classes; for this study, all non-weed classes were merged into the background to enable binary segmentation. (5964 images of 480×360 pixels) was selected, as it contains a larger number of images with weed presence, while the Sequoia subset includes a higher proportion of other vegetation classes. 4.2 Experimental setup Training is conducted in a Python environment using cross entropy loss and the AdamW optimizer. The learning rate is set to 1e-3 and scheduled using a OneCycle policy. Models were trained for 100 epochs with a batch size of 64 on an NVIDIA A100 GPU with 40 GB of memory. For comparison, several semantic segmentation models were trained under the same configuration, including established architectures and weed specific methods such as U-Net [29], DeepLabV3+ [3], SegFormer [41], SK-UNet [25], and WeedSense [33]. Performance is evaluated using the main semantic segmentation metrics, including intersection over union (IoU) for weed and background classes, and mean IoU (mIoU). Computational efficiency is assessed by measuring inference time (average latency per image), training time, total number of parameters, number of trainable parameters, and GFLOPs. The experiments were conducted using both RGB and full spectral configurations. To support additional channels, the first layer of each model is expanded accordingly. 4.3 Results and discussion Table 2 evaluates the impact of the α parameter on model performance. Results show a modest but consistent improvement as α increases, with the best performance achieved at α = 0.07 across both RGB and multispectral configurations. This indicates that stronger feature correction improves representation quality. Beyond 0.07, performance degrades, suggesting that excessive correction disrupts feature stability rather than enhancing discrimination. Also, α has no effect on efficiency, as parameter count, GFLOPs, and latency remain constant. Table 2: Ablation of the FCB α parameter for feature refinement. Alpha Dataset Modality IoU↑BG_BG IoU↑W_W mIoU↑ Inference time (s) Training time (h) Total params. (M) Trainable params. (M) GFLOPS 0.01 WeedBananaCOD RGB 0.990 0.730 0.860 0.0236 0.101 91.277 2.685 15.156 RGB-NIR 0.991 0.754 0.873 0.0237 0.106 91.279 2.685 15.190 WeedMap RGB 0.985 0.531 0.758 0.0141 0.129 91.277 2.685 10.657 RGB-NIR 0.986 0.531 0.759 0.0144 0.137 91.279 2.685 10.680 RGB-NIR-RE 0.986 0.536 0.761 0.0147 0.144 91.281 2.685 10.704 0.03 WeedBananaCOD RGB 0.990 0.732 0.861 0.0253 0.103 91.277 2.685 15.156 RGB-NIR 0.991 0.747 0.869 0.0279 0.108 91.279 2.685 15.190 WeedMap RGB 0.983 0.529 0.756 0.0173 0.132 91.277 2.685 10.657 RGB-NIR 0.986 0.531 0.759 0.0175 0.140 91.279 2.685 10.680 RGB-NIR-RE 0.986 0.535 0.761 0.0179 0.146 91.281 2.685 10.704 0.05 WeedBananaCOD RGB 0.990 0.722 0.856 0.0254 0.102 91.277 2.685 15.156 RGB-NIR 0.992 0.760 0.876 0.0281 0.109 91.279 2.685 15.190 WeedMap RGB 0.985 0.527 0.756 0.0158 0.130 91.277 2.685 10.657 RGB-NIR 0.985 0.530 0.758 0.0164 0.138 91.279 2.685 10.680 RGB-NIR-RE 0.985 0.532 0.759 0.0169 0.145 91.281 2.685 10.704 0.07 WeedBananaCOD RGB 0.991 0.733 0.862 0.0227 0.099 91.277 2.685 15.156 RGB-NIR 0.991 0.762 0.877 0.0298 0.105 91.279 2.685 15.190 WeedMap RGB 0.985 0.534 0.760 0.0141 0.128 91.277 2.685 10.657 RGB-NIR 0.986 0.545 0.766 0.0145 0.134 91.279 2.685 10.680 RGB-NIR-RE 0.987 0.547 0.767 0.0147 0.142 91.281 2.685 10.704 0.09 WeedBananaCOD RGB 0.990 0.713 0.851 0.0215 0.104 91.277 2.685 15.156 RGB-NIR 0.991 0.749 0.870 0.0240 0.105 91.279 2.685 15.190 WeedMap RGB 0.984 0.532 0.758 0.0169 0.130 91.277 2.685 10.657 RGB-NIR 0.985 0.531 0.758 0.0177 0.143 91.279 2.685 10.680 RGB-NIR-RE 0.986 0.537 0.762 0.0172 0.144 91.281 2.685 10.704 The ablation of the FCB bottleneck ratio is shown in Table 3. The results indicate that a configuration of 2 represents the optimal balance point. Due to the inverse nature of this parameter in the FCB block, increasing this ratio reduces model complexity and accelerates computation, as higher values decrease the number of internal channels. However, ratios above 2 lead to performance degradation. The excessive reduction in internal dimensionality limits the model’s representational capacity, preventing the block from capturing the fine details required for segmentation and resulting in a gradual loss of accuracy. Table 3: Ablation of the of the FCB bottleneck ratio. Bottleneck ratio Dataset Modality IoU↑BG_BG IoU↑W_W mIoU↑ Inference time (s) Training time (h) Total params. (M) Trainable params. (M) GFLOPS 1 WeedBananaCOD RGB 0.991 0.739 0.865 0.0248 0.107 92.682 4.090 16.260 RGB-NIR 0.991 0.752 0.872 0.0306 0.112 92.684 4.090 16.293 WeedMap RGB 0.985 0.517 0.751 0.0167 0.135 92.682 4.090 11.432 RGB-NIR 0.984 0.520 0.752 0.0169 0.139 92.684 4.090 11.456 RGB-NIR-RE 0.985 0.537 0.761 0.0171 0.148 92.686 4.090 11.480 2 WeedBananaCOD RGB 0.991 0.733 0.862 0.0227 0.099 91.277 2.685 15.156 RGB-NIR 0.991 0.762 0.877 0.0298 0.105 91.279 2.685 15.190 WeedMap RGB 0.985 0.534 0.760 0.0141 0.128 91.277 2.685 10.657 RGB-NIR 0.986 0.545 0.766 0.0145 0.134 91.279 2.685 10.680 RGB-NIR-RE 0.987 0.547 0.767 0.0147 0.142 91.281 2.685 10.704 3 WeedBananaCOD RGB 0.988 0.728 0.858 0.0220 0.096 90.812 2.221 14.882 RGB-NIR 0.990 0.753 0.872 0.0276 0.102 90.814 2.221 14.915 WeedMap RGB 0.985 0.526 0.756 0.0137 0.121 90.812 2.221 10.464 RGB-NIR 0.985 0.538 0.762 0.0138 0.126 90.814 2.221 10.487 RGB-NIR-RE 0.986 0.543 0.764 0.0140 0.129 90.816 2.221 10.511 4 WeedBananaCOD RGB 0.990 0.729 0.859 0.0214 0.094 90.583 1.991 14.747 RGB-NIR 0.991 0.740 0.865 0.0258 0.097 90.585 1.991 14.780 WeedMap RGB 0.985 0.513 0.749 0.0127 0.118 90.583 1.991 10.369 RGB-NIR 0.985 0.527 0.756 0.0127 0.123 90.585 1.991 10.393 RGB-NIR-RE 0.985 0.530 0.758 0.0129 0.124 90.587 1.991 10.416 5 WeedBananaCOD RGB 0.989 0.688 0.838 0.0214 0.090 90.449 1.857 14.692 RGB-NIR 0.989 0.702 0.846 0.0248 0.095 90.451 1.857 14.725 WeedMap RGB 0.985 0.517 0.751 0.0125 0.114 90.449 1.857 10.330 RGB-NIR 0.985 0.525 0.755 0.0123 0.125 90.451 1.857 10.354 RGB-NIR-RE 0.986 0.530 0.758 0.0128 0.124 90.453 1.857 10.377 The ablation study on the kernel size (Table 4) of the DWConvs in the FCB shows that a 3×3 configuration achieves the best performance. Although larger kernels (5×5 and 7×7) expand the receptive field and capture broader context, this becomes counterproductive within the FCB block. Since the backbone (ConvNeXt) already extracts large scale contextual information (7×7 kernels), using larger kernels in the FCB appears to introduce redundancy that degrades local detail sharpness. Additionally, increasing the kernel size leads to higher latency, training time, and GFLOPs without improving performance. Table 4: Ablation of the depthwise kernel size in the FCB block. Kernel size Dataset Modality IoU↑BG_BG IoU↑W_W mIoU↑ Inference time (s) Training time (h) Total params. (M) Trainable params. (M) GFLOPS 1 WeedBananaCOD RGB 0.989 0.696 0.842 0.0225 0.096 91.269 2.677 15.141 RGB-NIR 0.988 0.677 0.833 0.0296 0.103 91.271 2.677 15.174 WeedMap RGB 0.986 0.535 0.760 0.0130 0.123 91.269 2.677 10.646 RGB-NIR 0.985 0.533 0.759 0.0137 0.135 91.271 2.677 10.669 RGB-NIR-RE 0.986 0.537 0.761 0.0139 0.140 91.273 2.677 10.693 3 WeedBananaCOD RGB 0.991 0.733 0.862 0.0227 0.099 91.277 2.685 15.156 RGB-NIR 0.991 0.762 0.877 0.0298 0.105 91.279 2.685 15.190 WeedMap RGB 0.985 0.534 0.760 0.0141 0.128 91.277 2.685 10.657 RGB-NIR 0.986 0.545 0.766 0.0145 0.134 91.279 2.685 10.680 RGB-NIR-RE 0.987 0.547 0.767 0.0147 0.142 91.281 2.685 10.704 5 WeedBananaCOD RGB 0.990 0.724 0.857 0.0248 0.103 91.292 2.700 15.188 RGB-NIR 0.990 0.753 0.872 0.0300 0.107 91.294 2.700 15.221 WeedMap RGB 0.986 0.532 0.759 0.0168 0.132 91.292 2.700 10.679 RGB-NIR 0.986 0.540 0.763 0.0169 0.141 91.294 2.700 10.703 RGB-NIR-RE 0.986 0.543 0.765 0.0173 0.147 91.269 2.700 10.726 7 WeedBananaCOD RGB 0.988 0.682 0.835 0.0253 0.105 91.315 2.724 15.235 RGB-NIR 0.991 0.736 0.863 0.0309 0.109 91.317 2.724 15.269 WeedMap RGB 0.985 0.527 0.756 0.0176 0.132 91.315 2.724 10.712 RGB-NIR 0.986 0.535 0.761 0.0178 0.144 91.317 2.724 10.736 RGB-NIR-RE 0.986 0.533 0.760 0.0174 0.147 91.319 2.724 10.759 The ablation study on the FPN feature dimension (Table 5) shows that a dimension of 128 represents the optimal balance between performance and efficiency. A non-linear behavior is observed, where performance drops when increasing to 160, followed by a slight recovery at 192 without reaching the previous peak. This phenomenon suggests that increasing the dimensionality beyond 128 introduces structural redundancy that initially hinders convergence and dilutes encoder signals. Although a capacity of 192 allows the model to reorganize part of this information, the exponential increase in GFLOPs (from 15.1 to 28.4 for WeedBananaCOD) and latency confirms that higher dimensions saturate the decoder without providing meaningful improvements in data representational quality. Table 5: Ablation study of the FPN feature dimension. Feature dim. Dataset Modality IoU↑BG_BG IoU↑W_W mIoU↑ Inference time (s) Training time (h) Total params. (M) Trainable params. (M) GFLOPS 96 WeedBananaCOD RGB 0.990 0.717 0.853 0.0222 0.093 90.763 2.172 10.448 RGB-NIR 0.991 0.739 0.865 0.0292 0.095 90.765 2.172 10.482 WeedMap RGB 0.986 0.524 0.755 0.0130 0.126 90.763 2.172 7.346 RGB-NIR 0.986 0.530 0.758 0.0132 0.133 90.765 2.172 7.370 RGB-NIR-RE 0.984 0.537 0.761 0.0135 0.142 90.767 2.172 7.393 128 WeedBananaCOD RGB 0.991 0.733 0.862 0.0227 0.099 91.277 2.685 15.156 RGB-NIR 0.991 0.762 0.877 0.0298 0.105 91.279 2.685 15.190 WeedMap RGB 0.985 0.534 0.760 0.0141 0.128 91.277 2.685 10.657 RGB-NIR 0.986 0.545 0.766 0.0145 0.134 91.279 2.685 10.680 RGB-NIR-RE 0.987 0.547 0.767 0.0147 0.142 91.281 2.685 10.704 160 WeedBananaCOD RGB 0.989 0.719 0.854 0.0252 0.105 91.919 3.328 21.172 RGB-NIR 0.990 0.748 0.869 0.0308 0.108 91.921 3.328 21.205 WeedMap RGB 0.986 0.526 0.756 0.0170 0.132 91.919 3.328 14.886 RGB-NIR 0.983 0.535 0.759 0.0171 0.143 91.920 3.328 14.910 RGB-NIR-RE 0.986 0.541 0.764 0.0177 0.151 91.923 3.328 14.934 192 WeedBananaCOD RGB 0.990 0.731 0.861 0.0260 0.107 92.691 4.099 28.494 RGB-NIR 0.991 0.750 0.871 0.0316 0.110 92.693 4.099 28.528 WeedMap RGB 0.983 0.527 0.755 0.0170 0.131 92.691 4.099 20.035 RGB-NIR 0.986 0.542 0.764 0.0173 0.147 92.693 4.099 20.059 RGB-NIR-RE 0.985 0.545 0.765 0.0179 0.153 92.695 4.099 20.082 The ablation of refine depth in Table 6 shows that integrating two refinement blocks achieves the optimal balance between noise suppression in the fused FPN features and preservation of spatial detail. Exceeding this threshold leads to performance degradation. This is likely due to excessive smoothing blocks introducing an over-smoothing effect that blurs fine details, saturating the model’s capacity and leading to overfitting without improving representational quality, despite the increase in parameters and the associated rise in computational cost. Table 6: Ablation of the decoder refinement depth. Refine depth Dataset Modality IoU↑BG_BG IoU↑W_W mIoU↑ Inference time (s) Training time (h) Total params. (M) Trainable params. (M) GFLOPS 0 WeedBananaCOD RGB 0.990 0.716 0.853 0.0219 0.093 90.981 2.390 10.312 RGB-NIR 0.990 0.731 0.860 0.0226 0.094 90.983 2.390 10.346 WeedMap RGB 0.985 0.509 0.747 0.0140 0.131 90.981 2.390 7.251 RGB-NIR 0.986 0.521 0.753 0.0144 0.135 90.983 2.390 7.274 RGB-NIR-RE 0.986 0.530 0.758 0.0146 0.135 90.985 2.390 7.298 1 WeedBananaCOD RGB 0.990 0.713 0.852 0.0219 0.095 91.129 2.537 12.734 RGB-NIR 0.991 0.741 0.866 0.0229 0.097 91.131 2.537 12.768 WeedMap RGB 0.984 0.520 0.752 0.0145 0.133 91.129 2.537 8.954 RGB-NIR 0.985 0.527 0.756 0.0146 0.133 91.131 2.537 8.977 RGB-NIR-RE 0.986 0.539 0.762 0.0147 0.139 91.133 2.537 9.001 2 WeedBananaCOD RGB 0.991 0.733 0.862 0.0227 0.099 91.277 2.685 15.156 RGB-NIR 0.991 0.762 0.877 0.0298 0.105 91.279 2.685 15.190 WeedMap RGB 0.985 0.534 0.760 0.0141 0.128 91.277 2.685 10.657 RGB-NIR 0.986 0.545 0.766 0.0145 0.134 91.279 2.685 10.680 RGB-NIR-RE 0.987 0.547 0.767 0.0147 0.142 91.281 2.685 10.704 3 WeedBananaCOD RGB 0.990 0.725 0.857 0.0251 0.105 91.424 2.833 17.579 RGB-NIR 0.991 0.753 0.872 0.0278 0.107 91.426 2.833 17.612 WeedMap RGB 0.984 0.504 0.744 0.0150 0.131 91.424 2.833 12.360 RGB-NIR 0.986 0.540 0.763 0.0154 0.141 91.426 2.833 12.383 RGB-NIR-RE 0.986 0.543 0.764 0.0157 0.145 91.428 2.833 12.407 4 WeedBananaCOD RGB 0.989 0.693 0.841 0.0249 0.106 91.572 2.981 20.001 RGB-NIR 0.990 0.737 0.864 0.0263 0.107 91.574 2.981 20.034 WeedMap RGB 0.985 0.519 0.752 0.0153 0.138 91.572 2.981 14.063 RGB-NIR 0.986 0.531 0.759 0.0153 0.142 91.574 2.981 14.087 RGB-NIR-RE 0.987 0.536 0.762 0.0161 0.147 91.576 2.981 14.110 5 WeedBananaCOD RGB 0.990 0.714 0.852 0.0268 0.108 91.720 3.128 22.423 RGB-NIR 0.990 0.719 0.855 0.0276 0.110 91.722 3.128 22.457 WeedMap RGB 0.986 0.527 0.756 0.0161 0.140 91.720 3.128 15.766 RGB-NIR 0.986 0.532 0.759 0.0165 0.149 91.722 3.128 15.790 RGB-NIR-RE 0.985 0.535 0.760 0.0167 0.146 91.724 3.128 15.813 The ablation shown in Table 7 demonstrates that integrating the FCB block is the determining factor for achieving optimal performance. The results indicate that the w/FCB configuration achieves the highest mIoU across all datasets and spectral modalities, significantly outperforming both the baseline model (w/o FCB) and the alternative using CBAM. When the FCB is removed, a marked performance drop is observed, indicating that the frozen backbone alone lacks the necessary adaptation capacity for multispectral segmentation. In contrast, replacing the FCB with the CBAM attention block does not provide meaningful benefits; although CBAM presents fewer theoretical parameters and GFLOPs, this does not translate into real efficiency, resulting in noticeably higher training and inference times. This behavior confirms that CBAM operations are more costly in practice, while the FCB achieves the best balance between robustness, lower latency, and prediction accuracy. Table 7: Comparison of the proposed FCB with other alternatives. Component Dataset Modality IoU↑BG_BG IoU↑W_W mIoU↑ Inference time (s) Training time (h) Total params. (M) Trainable params. (M) GFLOPS w/o FCB WeedBananaCOD RGB 0.988 0.664 0.826 0.0223 0.093 89.871 1.280 14.053 RGB-NIR 0.989 0.693 0.841 0.0250 0.094 89.874 1.280 14.087 WeedMap RGB 0.978 0.514 0.746 0.0131 0.121 89.871 1.280 9.881 RGB-NIR 0.981 0.519 0.750 0.0134 0.125 89.874 1.280 9.905 RGB-NIR-RE 0.982 0.521 0.752 0.0134 0.130 89.876 1.280 9.928 w/CBAM WeedBananaCOD RGB 0.989 0.710 0.850 0.0382 0.140 90.048 1.457 14.056 RGB-NIR 0.988 0.693 0.840 0.0421 0.144 90.050 1.457 14.089 WeedMap RGB 0.978 0.521 0.750 0.0282 0.188 90.048 1.457 9.883 RGB-NIR 0.983 0.518 0.751 0.0286 0.194 90.050 1.457 9.906 RGB-NIR-RE 0.985 0.523 0.754 0.0292 0.202 90.052 1.457 9.930 w/FCB WeedBananaCOD RGB 0.991 0.733 0.862 0.0227 0.099 91.277 2.685 15.156 RGB-NIR 0.991 0.762 0.877 0.0298 0.105 91.279 2.685 15.190 WeedMap RGB 0.985 0.534 0.760 0.0141 0.128 91.277 2.685 10.657 RGB-NIR 0.986 0.545 0.766 0.0145 0.134 91.279 2.685 10.680 RGB-NIR-RE 0.987 0.547 0.767 0.0147 0.142 91.281 2.685 10.704 The performance of all FCBNet variants is presented in Table 8. The results show a consistent improvement in performance as the encoder scale increases, across both RGB and multispectral configurations in both datasets. In terms of efficiency, the expected increase in latency and training time is observed when scaling the architecture; however, the number of trainable parameters remains notably low due to the freezing strategy. In the Tiny variant, the number of trainable parameters is reduced from 30.6M to only 2.01M, representing a 93.4% reduction, while the Large variant achieves a massive reduction of 97.7%. This allows even the most complex model to be trained in approximately 0.2 hours, validating FCBNet as a flexible model whose structural efficiency enables the deployment of both lightweight and robust models under limited computational resources. Table 8: Performance comparison of FCBNet across different ConvNeXt encoders. Encoder Dataset Modality IoU↑BG_BG IoU↑W_W mIoU↑ Inference time (s) Training time (h) Total params. (M) Trainable params. (M) GFLOPS ConvNeXt-tiny WeedBananaCOD RGB 0.990 0.717 0.853 0.0139 0.060 30.604 2.015 13.151 RGB-NIR 0.991 0.740 0.865 0.0149 0.062 30.605 2.015 13.176 WeedMap RGB 0.985 0.537 0.761 0.0095 0.083 30.604 2.015 9.247 RGB-NIR 0.986 0.540 0.763 0.0103 0.091 30.605 2.015 9.264 RGB-NIR-RE 0.986 0.543 0.764 0.0107 0.098 30.607 2.015 9.282 ConvNeXt-small WeedBananaCOD RGB 0.987 0.658 0.822 0.0178 0.075 52.238 2.015 13.540 RGB-NIR 0.989 0.703 0.846 0.0200 0.080 52.240 2.015 13.565 WeedMap RGB 0.985 0.540 0.762 0.0132 0.107 52.238 2.015 9.520 RGB-NIR 0.986 0.544 0.765 0.0144 0.113 52.240 2.015 9.538 RGB-NIR-RE 0.987 0.545 0.766 0.0133 0.118 52.241 2.015 9.556 ConvNeXt-base WeedBananaCOD RGB 0.991 0.733 0.862 0.0227 0.099 91.277 2.685 15.156 RGB-NIR 0.991 0.762 0.877 0.0298 0.105 91.279 2.685 15.190 WeedMap RGB 0.985 0.534 0.760 0.0141 0.128 91.277 2.685 10.657 RGB-NIR 0.986 0.545 0.766 0.0145 0.134 91.279 2.685 10.680 RGB-NIR-RE 0.987 0.547 0.767 0.0147 0.142 91.281 2.685 10.704 ConvNeXt-large WeedBananaCOD RGB 0.991 0.746 0.868 0.0360 0.169 202.322 4.555 19.504 RGB-NIR 0.992 0.771 0.881 0.0381 0.171 202.325 4.555 19.554 WeedMap RGB 0.987 0.546 0.766 0.0252 0.201 202.322 4.555 13.714 RGB-NIR 0.986 0.548 0.767 0.0251 0.211 202.325 4.555 13.749 RGB-NIR-RE 0.987 0.551 0.769 0.0257 0.215 202.329 4.555 13.784 Finally, Table 9 presents the comparison between the proposed architecture and several state of the art methods. The results show that FCBNet-large achieves the highest performance across both datasets and all spectral modalities, suggesting a superior ability to adapt and effectively leverage information from multiple channels. It is also notable that, in the WeedMap dataset, all evaluated models exhibit a lower performance ceiling compared to WeedBananaCOD; however, FCBNet consistently outperforms its competitors even under these data constraints. On the other hand, although the FCBNet-tiny variant is slightly outperformed in mIoU by models such as DeepLabV3+ or SegFormer in certain scenarios, such as the WeedBananaCOD dataset, the performance gap is minimal compared to the substantial efficiency gains. While these models require training times exceeding 0.12 and 0.31 hours respectively and exhibit significantly higher GFLOPs, FCBNet-tiny achieves competitive performance with only 0.06 hours of training. Overall, these results validate the superiority of the proposed approach, establishing it as a highly efficient solution that drastically reduces computational requirements without compromising accuracy compared to the other heavy methods in the literature. Table 9: FCBNet compared to other approaches from the literature. We have selected the best and lightest FCBNet model for comparison. Encoder Dataset Modality IoU↑BG_BG IoU↑W_W mIoU↑ Inference time (s) Training time (h) Total params. (M) Trainable params. (M) GFLOPS FCBNet-tiny WeedBananaCOD RGB 0.990 0.717 0.853 0.0139 0.060 30.604 2.015 13.151 RGB-NIR 0.991 0.740 0.865 0.0149 0.062 30.605 2.015 13.176 WeedMap RGB 0.985 0.537 0.761 0.0095 0.083 30.604 2.015 9.247 RGB-NIR 0.986 0.540 0.763 0.0103 0.091 30.605 2.015 9.264 RGB-NIR-RE 0.986 0.543 0.764 0.0107 0.098 30.607 2.015 9.282 FCBNet-large WeedBananaCOD RGB 0.991 0.746 0.868 0.0360 0.169 202.322 4.555 19.504 RGB-NIR 0.992 0.771 0.881 0.0381 0.171 202.325 4.555 19.554 WeedMap RGB 0.987 0.546 0.766 0.0252 0.201 202.322 4.555 13.714 RGB-NIR 0.986 0.548 0.767 0.0251 0.211 202.325 4.555 13.749 RGB-NIR-RE 0.987 0.551 0.769 0.0257 0.215 202.329 4.555 13.784 WeedSense [33] WeedBananaCOD RGB 0.984 0.596 0.790 0.0223 0.089 12.721 12.721 12.731 RGB-NIR 0.986 0.625 0.805 0.0225 0.093 12.723 12.723 12.778 WeedMap RGB 0.979 0.471 0.725 0.0115 0.138 12.721 12.721 8.952 RGB-NIR 0.984 0.515 0.749 0.0190 0.151 12.723 12.723 8.985 RGB-NIR-RE 0.984 0.519 0.752 0.0196 0.166 12.724 12.724 9.018 DeepLabV3+ [3] WeedBananaCOD RGB 0.988 0.721 0.855 0.0270 0.123 26.678 26.678 36.762 RGB-NIR 0.991 0.725 0.858 0.0390 0.125 26.681 26.681 36.968 WeedMap RGB 0.983 0.509 0.746 0.0190 0.161 26.678 26.678 25.849 RGB-NIR 0.984 0.515 0.749 0.0190 0.170 26.681 26.681 25.993 RGB-NIR-RE 0.986 0.534 0.760 0.0200 0.189 26.684 26.684 26.137 U-Net [29] WeedBananaCOD RGB 0.990 0.715 0.851 0.0243 0.097 32.521 32.521 42.771 RGB-NIR 0.990 0.720 0.855 0.0257 0.100 32.524 32.524 42.976 WeedMap RGB 0.975 0.427 0.701 0.0109 0.119 32.521 32.521 30.073 RGB-NIR 0.981 0.508 0.745 0.0112 0.126 32.524 32.524 30.218 RGB-NIR-RE 0.986 0.550 0.768 0.0117 0.131 32.528 32.528 30.362 SegFormer [41] WeedBananaCOD RGB 0.989 0.695 0.842 0.0203 0.310 84.595 84.595 49.733 RGB-NIR 0.990 0.738 0.864 0.0295 0.312 84.598 84.598 49.784 WeedMap RGB 0.986 0.506 0.746 0.0177 0.420 84.595 84.595 34.116 RGB-NIR 0.985 0.532 0.758 0.0180 0.421 84.598 84.598 34.153 RGB-NIR-RE 0.985 0.535 0.760 0.0189 0.421 84.601 84.601 34.189 SK-U-Net [25] WeedBananaCOD RGB 0.986 0.641 0.813 0.0231 0.170 34.444 34.444 44.745 RGB-NIR 0.990 0.705 0.847 0.0238 0.170 34.447 34.447 44.950 WeedMap RGB 0.983 0.532 0.758 0.0140 0.160 34.444 34.444 31.461 RGB-NIR 0.982 0.539 0.761 0.0164 0.170 34.447 34.447 31.606 RGB-NIR-RE 0.984 0.539 0.762 0.0140 0.178 34.450 34.450 31.751 To complement the quantitative analysis, Figs. 6(a) and 6(b) present a visual comparison of inference on the WeedBananaCOD and WeedMap datasets, respectively, using their full modalities. In WeedBananaCOD, FCBNet achieves more precise instance delineation and a significant reduction in false detections (rows b and d), producing results that more closely match the ground truth compared to the other models. In WeedMap, despite the sparse distribution and small size of weed instances, the proposed architecture also demonstrates a greater ability to identify small objects with a lower error rate. In both cases, the qualitative results confirm that the FCBNet correction mechanism better preserves fine details and improves the spatial coherence of the segmentation. (a) WeedBananaCOD (RGB-NIR). (b) WeedMap (RGB-NIR-RE). Figure 6: Qualitative comparison between FCBNet and other models from the state-of-the-art. Zoom in for better visualization. 5 Limitations and Future work Despite the evaluation across multiple datasets, spectral modalities, and ablation tests, we acknowledge limitations that define future research directions. First, the evaluation can be extended to other datasets containing additional scenarios, spectral bands, and classes (crops or different weed species), as this study focused on binary segmentation. Furthermore, evaluating FCBNet on specialized hardware and embedded systems, such as Raspberry Pi or UAV platforms, would provide valuable complementary validation of its behaviour under real deployment conditions. Finally, exploring structural modifications, including the use of dual encoders for multispectral processing or the integration of alternative decoder designs, represents a natural research direction derived from this work. 6 Conclusions We introduce FCBNet, an efficient architecture for weed detection. It employs a frozen ConvNeXt backbone integrated with the proposed FCB block, which combines efficient convolutions to refine feature representations at minimal computational cost, along with a lightweight FPN decoder. Extensive evaluations on the WeedBananaCOD and WeedMap datasets demonstrate that FCBNet outperforms established models such as U-Net, SK-U-Net, SegFormer, and WeedSense in both mIoU and operational efficiency. Specifically, the freezing strategy enables a reduction of trainable parameters >>93%, achieving training times of only 0.06 to 0.2 hours without compromising performance. References [1] D. Chen, Y. Lu, Z. Li, and S. Young (2022) Performance evaluation of deep transfer learning on multi-class identification of common weed species in cotton production systems. Comput. Electron. Agric. 198, p. 107091. Cited by: §1. [2] D. Chen, X. Qi, Y. Zheng, Y. Lu, Y. Huang, and Z. Li (2024) Synthetic data augmentation by diffusion probabilistic models to enhance weed recognition. Comput. Electron. Agric. 216, p. 108517. Cited by: §1. [3] L. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam (2018) Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, Cited by: §4.2, Table 9. [4] P. Chen, S. Liu, H. Zhao, and J. Jia (2021) Distilling knowledge via knowledge review. In CVPR, Cited by: §1. [5] F. Dang, D. Chen, Y. Lu, and Z. Li (2023) YOLOWeeds: a novel benchmark of yolo object detectors for multi-class weed detection in cotton production systems. Comput. Electron. Agric. 205, p. 107655. Cited by: §1. [6] B. Deng, Y. Lu, and J. Xu (2024) Weed database development: an updated survey of public weed datasets and cross-season weed detection adaptation. Ecol. Inform. 81, p. 102546. Cited by: §1. [7] M. El Hafyani, A. Saddik, M. Hssaisoune, A. Labbaci, A. Tairi, F. Abdelfadel, and L. Bouchaou (2025) Weeds detection in a citrus orchard using multispectral uav data and machine learning algorithms: a case study from souss-massa basin, morocco. Remote Sens. Environ. 38, p. 101553. Cited by: §1. [8] J. Gou, B. Yu, S. Maybank, and D. Tao (2021) Knowledge distillation: a survey. Int. J. Comput. Vis. 129, p. 1789–1819. Cited by: §1. [9] Z. Guo, Y. Xue, C. Wang, Y. Geng, R. Lu, H. Li, D. Sun, Z. Lou, T. Chen, J. Shi, and L. Quan (2025) Efficient weed segmentation in maize fields: a semi-supervised approach for precision weed management with reduced annotation overhead. Comput. Electron. Agric. 229, p. 109707. Cited by: §2. [10] S. K. Gupta, S. K. Yadav, S. K. Soni, U. Shanker, and P. K. Singh (2023) Multiclass weed identification using semantic segmentation: an automated approach for precision agriculture. Ecol. Inform. 78, p. 102366. Cited by: §2. [11] J. B. Hopson, A. Flaus, C. J. McGinnity, R. Neji, A. J. Reader, and A. Hammers (2024) Deep convolutional backbone comparison for automated pet image quality assessment. IEEE TRPMS 8 (8), p. 893–901. Cited by: §1. [12] Md. J. Karim, Md. Nahiduzzaman, M. Ahsan, and J. Haider (2024) Development of an early detection and automatic targeting system for cotton weeds using an improved lightweight yolov8 architecture on an edge device. Knowl.-Based Syst. 300, p. 112204. Cited by: §1. [13] D. Karimi, S. K. Warfield, and A. Gholipour (2021) Transfer learning in medical image segmentation: new insights from analysis of the dynamics of model parameters and learned representations. Artif. Intell. Med. 116, p. 102078. Cited by: §1. [14] A. Kaushal, O. Almurshed, A. Alabbas, N. Auluck, and O. Rana (2023) An edge-cloud infrastructure for weed detection in precision agriculture. In IEEE IDASC/PiCom/CBDCom/CyberSciTech, Cited by: §1. [15] J. Liao, M. Chen, K. Zhang, H. Zhou, Y. Zou, W. Xiong, S. Zhang, F. Kuang, and D. Zhu (2024) SC-net: a new strip convolutional network model for rice seedling and weed segmentation in paddy field. Comput. Electron. Agric. 220, p. 108862. Cited by: §1, §2. [16] T. Lin, P. Dollar, R. Girshick, K. He, B. Hariharan, and S. Belongie (2017) Feature pyramid networks for object detection. In CVPR, Cited by: §1. [17] Z. Liu, H. Mao, C. Wu, C. Feichtenhofer, T. Darrell, and S. Xie (2022) A convnet for the 2020s. In CVPR, Cited by: §1. [18] R. U. Modi, M. Kancheti, A. Subeesh, C. Raj, A. K. Singh, N. S. Chandel, A. S. Dhimate, M. K. Singh, and S. Singh (2023) An automated weed identification framework for sugarcane crop: a deep learning approach. Crop Prot. 173, p. 106360. Cited by: §1. [19] N. S. Naik and H. K. Chaubey (2024) Weed detection and classification in sesame crops using region-based convolution neural networks. Neural Comput. Appl. 36, p. 18961–18977. Cited by: §1. [20] S. Niu, Y. Liu, J. Wang, and H. Song (2020) A decade survey of transfer learning (2010–2020). IEEE TAI 1 (2), p. 151–166. Cited by: §1. [21] D. G. Pai, R. Kamath, and M. Balachandra (2024) Deep learning techniques for weed detection in agricultural environments: a comprehensive review. IEEE Access 12, p. 113193–113214. Cited by: §1, §1, §1. [22] H. Peng, Z. Li, Z. Zhou, and Y. Shao (2022) Weed detection in paddy field using an improved retinanet network. Comput. Electron. Agric. 199, p. 107179. Cited by: §1. [23] R. Raja, W. Su, D. C. Slaughter, and S. A. Fennimore (2025) Real-time precision crop identification in high weed-density environments for robotic weed control using spectral fluorescence imaging in celery. Comput. Electron. Agric. 231, p. 110022. Cited by: §1, §1. [24] L. T. Ramos and A. D. Sappa (2024) Multispectral semantic segmentation for land cover classification: an overview. IEEE JSTARS 17, p. 14295–14336. Cited by: §1. [25] L. T. Ramos and A. D. Sappa (2025) Leveraging u-net and selective feature extraction for land cover classification using remote sensing imagery. Sci. Rep. 15 (1), p. 784. Cited by: §4.2, Table 9. [26] L. T. Ramos and A. D. Sappa (2026) Multi-encoder convnext network with smooth attentional feature fusion for multispectral semantic segmentation. arXiv preprint arXiv:2602.10137. Cited by: Figure 3, Figure 3. [27] N. Razfar, J. True, R. Bassiouny, V. Venkatesh, and R. Kashef (2022) Weed detection in soybean crops using custom lightweight deep learning models. J. Agric. Food Res. 8, p. 100308. Cited by: §1, §2. [28] M. U. Rehman, H. Eesaar, Z. Abbas, L. Seneviratne, I. Hussain, and K. T. Chong (2024) Advanced drone-based weed detection using feature-enriched deep learning approach. Knowl.-Based Syst. 305, p. 112655. Cited by: §1, §1. [29] O. Ronneberger, P. Fischer, and T. Brox (2015) U-net: convolutional networks for biomedical image segmentation. In MICCAI, Cited by: §4.2, Table 9. [30] I. Sa, M. Popović, R. Khanna, Z. Chen, P. Lottes, F. Liebisch, J. Nieto, C. Stachniss, A. Walter, and R. Siegwart (2018) WeedMap: a large-scale semantic weed mapping framework using aerial multispectral imaging and deep neural network for precision farming. Remote Sens. 10 (9). Cited by: §4.1, §4.1. [31] H. M. Sahin, T. Miftahushudur, B. Grieve, and H. Yin (2023) Segmentation of weeds and crops using multispectral imaging and crf-enhanced u-net. Comput. Electron. Agric. 211, p. 107956. Cited by: §1, §2. [32] B. B. Sapkota, S. Popescu, N. Rajan, R. G. Leon, C. Reberg-Horton, S. Mirsky, and M. V. Bagavathiannan (2022) Use of synthetic images for training a deep learning model for weed detection and biomass estimation in cotton. Sci. Rep. 12 (1), p. 19580. Cited by: §1. [33] T. T. Sarker, K. R. Ahmed, T. Islam, C. B. Rankrape, and K. Gage (2025) WeedSense: multi-task learning for weed segmentation, height estimation, and growth stage classification. In ICCVW, Cited by: §4.2, Table 9. [34] A. Sharma, V. Kumar, and L. Longchamps (2024) Comparative performance of yolov8, yolov9, yolov10, yolov11 and faster r-cnn models for detection of multiple weed species. Smart Agric. Technol. 9, p. 100648. Cited by: §1. [35] V. Singh, D. Singh, and H. Kumar (2024) Efficient application of deep neural networks for identifying small and multiple weed patches using drone images. IEEE Access 12, p. 71982–71996. Cited by: §1, §2. [36] J. Su, D. Yi, M. Coombes, C. Liu, X. Zhai, K. McDonald-Maier, and W. Chen (2022) Spectral analysis and mapping of blackgrass weed by leveraging machine learning and uav multispectral imagery. Comput. Electron. Agric. 192, p. 106621. Cited by: §1. [37] B. Tang, J. Zhou, C. Zhao, Y. Pan, Y. Lu, C. Liu, K. Ma, X. Sun, R. Zhang, and X. Gu (2025) Using uav-based multispectral images and cgs-yolo algorithm to distinguish maize seeding from weed. Artif. Intell. Agric. 15 (2), p. 162–181. Cited by: §1, §2. [38] H. O. Velesaca, A. Mero, H. Villegas, and A. D. Sappa (2026) Unveiling the hidden: early detection of invasive vegetation in crops with uav multispectral imaging. Smart Agric. Technol., p. 101875. Cited by: §4.1, §4.1. [39] L. Wang and K. Yoon (2022) Knowledge distillation and student-teacher learning for visual intelligence: a review and new outlooks. IEEE TPAMI 44 (6), p. 3048–3068. Cited by: §1. [40] P. Wimmer, J. Mehnert, and A. P. Condurache (2023) Dimensionality reduced training by pruning and freezing parts of a deep neural network: a survey. Artif. Intell. Rev. 56 (12), p. 14257–14295. Cited by: §1. [41] E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo (2021) SegFormer: simple and efficient design for semantic segmentation with transformers. In NeurIPS, Cited by: §4.2, Table 9. [42] K. Xu, P. Yuen, Q. Xie, Y. Zhu, W. Cao, and J. Ni (2023) WeedsNet: a dual attention network with rgb-d image for weed detection in natural wheat field. Precis. Agric. 25 (1), p. 460–485. Cited by: §1. [43] H. Zhang, Z. Jiang, G. Zheng, and X. Yao (2023) Semantic segmentation of high-resolution remote sensing images with improved u-net based on transfer learning. Int. J. Comput. Intell. Syst. 16 (181), p. 1–11. Cited by: §1. [44] M. Zhang, H. Singh, L. Chok, and R. Chunara (2022) Segmenting across places: the need for fair transfer learning with satellite imagery. In CVPRW, Cited by: §1. [45] F. Zhuang, Z. Qi, K. Duan, D. Xi, Y. Zhu, H. Zhu, H. Xiong, and Q. He (2021) A comprehensive survey on transfer learning. Proc. IEEE 109 (1), p. 43–76. Cited by: §1. [46] R. Zoetmulder, E. Gavves, M. Caan, and H. Marquering (2022) Domain- and task-specific transfer learning for medical segmentation tasks. Comput. Methods Programs Biomed. 214, p. 106539. Cited by: §1. [47] K. Zou, X. Chen, Y. Wang, C. Zhang, and F. Zhang (2021) A modified u-net with a specific data argumentation method for semantic segmentation of weed images in the field. Comput. Electron. Agric. 187, p. 106242. Cited by: §2.