Cassava is the staple crop for more than 800 million people around the world and provides a principal source of carbohydrates in tropical regions (FAO, 2017). However, its productivity is seriously curtailed by weed competition mainly within the first three months after planting when it is most vulnerable. Weeds are a major problem in root and tuber crops, causing large yield losses of about 55-75% (Alabi et al., 2022). They also make harvesting tough because the beneficial part of the plant grows underground. In addition, weeds can damage the quality of roots and tubers by directly attacking or feeding them. Since different weeds grow in diverse ways, no single control method works for all circumstances (Onasanya et al., 2021).
       
Weed populations in cassava fields vary widely depending on geographic location. Annual weeds dominate cassava cultivation because of their rapid growth and prolific seed production. Mechanical control methods such as hand pulling and hoeing are widely used to manage these weeds and it is most effective to remove them before seed dispersal (Parven et al., 2025).
       
Precision agriculture offers a paradigm shift by leveraging technology for site-specific crop management. A crucial aspect is the machine-aided identification of weeds, which allows for mechanical weeding (or targeted herbicide sprayers such as robot weeders) to only act where needed (Mohanty et al., 2026). This significantly reduces chemical application, decreasing cost and environmental impact (Sapkota et al., 2023). The recent progressions in DL techniques, especially CNN, have led to major successes on visual recognition tasks, which makes it possible for complex agricultural vision issues such as crop/weed classification under unstructured field conditions (Isinkaye et al., 2024).
       
For weed detection, numerous studies have shown great promise. (Sa et al., 2018) used a Multi-Modal Faster R-CNN to detect weeds in beet fields. (Hasan et al., 2021) analyzed several CNN architectures for detecting weeds and argued in favour of the transfer learning. However, most of the research has studied crops such as corn, soybean and sugar beet. Cassava, with its atypical stem architecture and leaf shapes represents a special case that has been less intensively studied.
       
Recently, CNN-based models have encountered many challenges in detection and classification of weed, particularly in fields of cassava crops. The first major issue is the limited availability of comprehensive, publicly accessible datasets specifically designed for weed identification in cassava-field. The second challenge is the lack of low power, lightweight CNN architectures capable of delivering high accuracy classification of weed under field conditions. Addressing these gaps is essential to support researchers and farmers who may not have advanced technical expertise but require reliable, real-time weed identification tools for cassava cultivation. In this work, we tried to tackle this deficiency by designing dedicated deep learning architecture for cassava farming management and weed identification. This study focuses the customized CNN and specialized datasets, to establish the accuracy and efficient weed classification and identification model in cassava crop fields.
       
Very few research works have been done for the cassava plant crop, which faces the issue of weeds in its cultivation process. Moreover, most advanced approaches use complex deep learning models which cannot be used efficiently in the real world due to their high processing cost and low efficiency. In this research, we introduce a customized approach called the customized convolutional veural vetwork (CCNN), which is efficient for cassava-field weed detection. Our proposed approach performs progressive feature extraction (Mandlik et al., 2026) through its compact network structure, which leads to a reduction in computational cost without sacrificing performance.
       
This work contributes a novel deep learning framework for weed semantic segmentation in cassava crops. The primary objectives are:
1. To generate a high-quality cassava weed image dataset containing cassava plants, broadleaf weeds, grassy weeds and soil classes under diverse field conditions.
2. To design a lightweight customized convolutional neural network (CCNN) for accurate weed classification in cassava cultivation.
3. To evaluate the effectiveness of the proposed CCNN using standard performance metrics, including accuracy, precision, recall, F1-score and ROC analysis.

Fig 1 shows the whole framework of the weed detection system design. The approach used in designing the system comprises five main steps, namely; data collection, image preprocessing, data augmentation, feature extraction and classification through CCNN and model performance evaluation. First, images of cassava fields are collected under different conditions and are divided into four categories, which include: cassava plants, broadleaf weeds, grassy weeds and Sedges. Image preprocessing is done through resizing and normalizing the images obtained, after which data augmentation processes are done to increase model generalization through techniques such as rotation, flipping and scaling. The final stage involves training a CCNN model using the images that have undergone data augmentation, after which model evaluation is done using several performance metrics.


 
Data collection and preprocessing
 
A total of 2,500 images in RGB format were obtained from cassava plant cultivation fields from Sundarapuram, Attur, Salem, Tamil Nadu, India and SRM College of Agricultural Sciences, Baburayanpettai-Chengalpattu. These images were acquired through a digital high-definition camera in different environments based on illumination intensity, weed density, stage of growth of crops and various angles of view. In this way, we obtained a dataset consisting of four semantic classes, namely, cassava plants, broadleaf weeds, grassy weeds and soil. Pixel-wise annotations of these images were created using the LabelMe annotation tool. In order to have effective model training, we have divided the data into training, and testing sets, each consisting of 2,000 and 500 images, respectively. Fig 2 shows sample images with diverse lighting conditions.


       
This dataset contains images of cassava crops and various weed species. After collection, the data undergoes a preparation stage aimed at removing poor-quality samples and ensuring that only clean, high-quality data is retained. This process is essential for selecting and training the most effective model. Fig 3 shows data pre-processing stages.


 
Data augmentation
 
Images are resized, normalized and augmented like rotation, scaling and flipping to improve the model’s robustness. To overcome over-fitting problems and improve model generalization, extensive online data augmentation was applied during training. This included random rotations (±30°), horizontal and vertical flips, brightness and contrast adjustments (±20%) and gaussian blurring. Table 1 shows data augmentation techniques.


 
Model architecture
 
The convolutional neural networks are deep learning architectures that are specifically designed to operate directly on grid-like or image data (Hasan et al., 2021). Convolutional layers apply learnable filters to incoming data, constructing spatial hierarchies based on patterns such as edges, shapes and textures. Each convolution layer has an associated activation function, typically ReLU, which adds non-linearity to the model. Pooling layers, like max pooling, reduce the spatial dimensions of data to control the over-fitting and complexity. CNNs typically end by using fully connected layers that combine the learned features for regression or classification (Chepuri et al., 2025). A major strong point of CNNs is that their shared weights and local connection enable them to be computationally efficient as well as being applicable to high-dimensional data, such as 2D image plane. Fig 4 shows the general layout of Convolution Neural Network and Fig 5 presents the architecture of novel customized CNN employed for cassava weed classification.




 
Convolutional operation
 
During the convolution operation, the filter slides over the input image. A dot product is then taken at each position between the filter and a region of the input it is covering. The resultant value is saved into the output feature map at that (x, y) position. This carries on throughout the whole image. For instance, if the size of filter is 33 and that of input image is 256*256, then it slides over as a result producing an output feature map of 256*256. The output of the convolution operation is a feature map, which is essentially a transformed version of the original output, highlighting the features that the filter was designed to detect such as edges or textures. The stride of ta convolution is the number of row and column steps that are taken when you shift the kernel. Increasing stride speeds up processing but may cause the network to miss small features. Padding adds extra pixels around each input edge, typically zeros, to maintain output size after convolution. Without padding, feature maps shrink after each layer, potentially losing important edge information. Starting with small filters, stride one, padding helps preserve details in early layers. Fig 6 shows a) Convolutional layer, b) Max pooling layer.


 
The CNN’s convolution operation is described as follows:
 

                  Y_i,j = (X * W)_i,j = ∑_m ∑_n X_{i + m, j + n} W_m,n                                ...(1)  

                                                
Where,
X= Input image.
Y (i,j)= Output activation at position (i, j).
W= Filter weights (kernel).
i,j= Indices over the output dimensions.
m, n= Indices over the filter dimensions.
*= Convolution operation.
       
The activation map calculations and output dimension calculations are given below:

                                                                                                                 
Where,
X= Size of the input image.
F= Size of the filter (height or width).
P= Padding applied to the input.
S= Stride of the convolution.
 
Pooling operation
 
After initializing the CNN, a pooling operation is applied to reduce the spatial dimensions of the feature maps. Pooling down-samples the input, thereby decreasing the computational load and reducing the number of learnable parameters in the network. A popular activation function in the DL algorithm adds non-linearity to the model is the rectifier linear unit (ReLU).
 

                ReLU (x) = max (0, x)                          ...(3)

       
Mathematically, ReLU returns the input itself if then is positive; otherwise, it returns zero. ReLU is efficient because it only needs a threshold operation and it makes the network sparser by only activating some of the neurons.

Dropout layer
 
The dropout layer randomly shuts off a fraction of neurons from the previous layer, cutting down the number of active units. This forces the network to learn more robust features instead of relying too much on any single neuron. It is used to deep CNNs, especially when the architecture gets complicated.
 
Flatten layer
 
This layer flattens the multi-dimensional data from convolutional blocks to one-dimensional vector, preparing it for fully connected layers. This one-dimensional vector is suitable as input for fully connected layers, which typically work with flat data for classification or regression tasks.
 
Fully connected layers 
 
The actual classification or prediction happens in these layers. The normal fully connected layers take input from the flatten layer and carry out the classification or prediction task. The output layer has the appropriate function like SoftMax, based on the task, which is usually multi class classification (Pakruddin et al. 2025). We combine all the parts sequentially to build a complete CNN network. Multiple convolution blocks, which will extract features, then a set of fully connected (Dense) Layers, which will use those features to make classification.
 
Algorithm 1: Cassava weed classification using proposed Customized CNN.
Require:
Epochs = 50.
Batch _ size = 16.
Input _ shape = (256, 256, 3).
Classes = 4.
Activation = ReLu, Softmax.
Optimizer = Adam.
       
Function: Cassava_Weed_Detection (epochs, batch_size, input_shape, classes, activation, optimizer).
 
 #1: Data preparation
 
1. Load cassava weed image dataset.
2. Resize images to 256 × 256 × 3.
3. Normalize pixel values.
4. Split datasets into training and validation sets.
train_data, val_data←load_and_preprocess_data (input_ shape).
 
#2: Model definition
 
1. nitialize customized CNN architecture.
2. Add convolutional, pooling and dense layers.
3. Apply ReLU activation in hidden layers.
4. Apply Softmax activation in output layer for multi-class classification.
model¬define_CNN_model (input_shape, classes, activation).
 
#3: Model training
 
1. Compile model using Adam optimizer.
2. Train model with training dataset for 50 epochs using batch size 16.
3. Validate model using validation dataset during training.
trained_model¬ train_model (model, train_data, val_data, epochs, batch_size, optimizer).
 
 #4: Model evaluation
 
1. Evaluate trained model on validation dataset.
2. Compute classification accuracy, loss, precision, recall and F1-score.
evaluate_model (trained_model, val_data).
 
#5: Result interpretation
 
1. Predict weed categories from test images.
2. Analyze performance metrics and classification results.
3. Identify correctly and incorrectly classified weed samples. 
End function.
 
Experimental setup
 
The experimental Parameter setup is shown in Table 2. The weight decay is 0.005. The batch size is set to 16. The initial learning rate is 0.001, which maximum training is fixed at the epoch 50. 


 
Hyperparameter for customized CNN model
 
The hyperparameter values of the customized CNN Model are shown in Table 3.


       
The process of hyperparameter optimization was carried out by conducting experiments. The optimal combination of hyperparameters was found when the learning rate was set at 0.001, batch size was equal to 16, dropout rate was 0.5 and the value of the weight decay coefficient was 0.005. For the model training, the Adam optimizer was chosen owing to fast convergence and adaptive learning.

Performance metrics
 
The classification report is an important performance metric for classification based deep learning algorithms.  A confusion matrix is a table that shows how well a classification model performs by comparing actual (true) and predicted classes. For the sake of completeness of the paper, we recall the mathematical expressions of these metrics in Table 4, where TrPo, TrNe, FaPo and FaNe mentioned in these expressions refer to true positive, true negative, false positive and false negative, respectively.



Customized CNN approach
 
In this study, we first used a CNN model and fine-tuned its parameters to categorise cassava weed images into four classes: Cassava, Broadleaf Weed, Grass Weed and Sedge. We refer to the novel customized CNN Model. The architecture of the novel Customized CNN was carefully designed to enhance its performance.
       
Table 5 presents a comparative summary of the convolutional architectures used in VGG16, ResNet50, VGG19 and the proposed customized CNN model. The VGG16 configuration begins with two convolutional layers, each containing 16 kernels of size 3×3, followed by a 2×2 max-pooling operation with stride 2 and uses a softmax activation in this stage. VGG19 extends this structure by adding four consecutive convolutional layers, each with 8 kernels of size 2×2, also paired with a 2×2 max-pooling layer. ResNet50 incorporates a deeper hierarchical structure with four convolutional blocks. The first block contains 16 kernels, followed by blocks with 32, 64 and 128 kernels, respectively. Each block uses 3×3 kernels and includes max-pooling layers of 2×2 with stride 2, while intermediate layers within each block perform additional convolutions to enhance feature extraction and residual learning. The proposed model follows a similar progressive feature-expansion design but introduces an additional fifth convolutional block with 256 kernels, enabling richer high-level feature representation. Each block begins with a max-pooling layer and stride 2 to reduce spatial dimensions, followed by one or more convolutional layers of 3×3 kernels. This deeper and more structured architecture allows the proposed model to capture both low-level and high-level spatial features more effectively than the baseline models, contributing to its superior performance in subsequent experimental evaluations. As mentioned in Table 6, displays the Customized Convolutional Neural Network Models Performance.





The CNN model being considered in this study shows very high consistency with respect to its classification performance on all four classes, as shown in Table 7. Overall, it shows an accuracy score of 99.56%, which means that almost all of the test samples are correctly classified by the model. Among the four individual classes, the Broadleaf Weed gets a recall of 99.99% with an F1-score of 99.85%, while the Grassy Weed gets a precision score of 99.81%. The Cassava class shows a recall of 99.99%, while the Sedges class gets a recall of 100% with an F1-score of 99.55%.The macro-averaged F1-score of 0.960 and weighted-average F1-score of 0.958 confirm that the model performs consistently well across classes regardless of class imbalance in the support distribution, demonstrating the robustness of the proposed CNN architecture for weed and crop classification tasks.


 
ROC curve
 
The ROC curves for the proposed CCNN model and other competitive deep learning models are shown in Fig 7. The proposed CCNN model produces the highest area under the curve (AUC) value, that is 99.56%. It shows that the model has an excellent discriminatory ability between the weed and crop classes.


       
Fig 8 shows the graph representing training and validation accuracy and loss for 50 epochs. In both cases, the training and validation accuracy increases significantly during the learning process and converges to around 99.56% accuracy, which indicates that the model has learned the features successfully and is optimized well. Moreover, similarity between both curves shows that the model can generalize effectively without overfitting. On the other hand, the training and validation loss of the model decreases significantly within the first few epochs and stabilizes eventually to a very small value. The close alignment between the training and validation loss curves further confirms the model’s stability and robustness. Overall, the results indicate that the model achieves high accuracy with minimal generalization error, demonstrating effective feature learning throughout the training process.


 
Comparison of complexity in network
 
The proposed CNN was compared against the state of the art DL models such as VGG16, ResNet50, VGG19, ResNet-18, MobileNetV2 and Google Net. Designers rarely consider the number of parameters and the associated memory requirements. The two measures provided by this study are memory storage and the number of learnable parameters, as shown in Table 8. The network structure’s complexity is shown by these two measures. When compared to other classical models, our proposed CCNN model has the fewest learnable parameters and memory storage. Surprisingly, CCNN only has 1.09% of VGG16’s learnable parameters. Besides accuracy, efficiency can be considered a crucial aspect to be achieved by the models used in the field of precision agriculture. The model suggested here is composed of 1.50 million parameters and needs about 6.76 MB of memory space, which is quite smaller compared to VGG16, VGG19 and ResNet50. Such features help improve computation performance and allow the model to run much faster. As such, the suggested model can be considered appropriate for use in weed detection on drones and other robotic tools in farming.


 
Comparison of the state of the art
 
To assess the performance of the proposed CNN model, it was compared with several other models, including InceptionV4 (Mishra et al., 2022), as well as lightweight network models such as ShuffleNetV2 (Fan et al., 2024), ResNet50 (Asad et al., 2020), MobileNetV3-small (Zi et al., 2025) and MobileNetV4 (Qin et al., 2024). Based on model identification model size, ability and inference speed, each model’s performance was assessed using the same dataset and experimental setup. Metrics like accuracy, precision, recall and F1-Score were used to evaluate recognition abilities. Latency and FLOPs were the main metrics used to assess inference speed.
       
The findings of the experiments indicate that the developed customized convolutional neural network (CCNN) is effective in recognizing weed plants in cassava fields based on the images. The recognition rate reached 99.56%, surpassing the performance of classical neural network models like VGG16, VGG19 and ResNet50. The increased accuracy can be explained by the advanced architecture of the CCNN, which allows recognizing unique characteristics of cassava plants, broadleaf weeds, grassy weeds and sedges without being computationally expensive.
       
The results of the ROC analysis prove the high discriminatory capability of the proposed framework. The value of the AUC obtained equal to 99.56% confirms that the framework shows a high ability to distinguish between the crop and weed images regardless of the decision threshold. Moreover, the high correspondence of the accuracy graphs for both learning datasets proves good generalization capabilities without significant signs of overfitting.
 
Future directions
 
Future research will focus on extending the proposed CCNN framework toward real-time and large-scale deployment for precision weed management in cassava cultivation. Several promising directions include:
 
Integration with UAV and edge-AI platforms
 
Implementing the model on unmanned aerial vehicles (UAVs) and embedded systems to enable on-site, real-time weed detection and decision-making for autonomous spraying or mechanical removal.
 
Model generalization and transferability
 
Expanding the dataset across diverse agro-ecological zones, soil types and seasonal variations to improve model robustness and ensure reliable performance under varying field conditions.

Multispectral and hyper-spectral data fusion
 
Incorporating multispectral or hyper-spectral imagery to enhance weed-crop discrimination accuracy, particularly during early growth stages and under challenging lighting conditions.

Weed infestation constitutes a critical constraint in cassava cultivation, leading to substantial yield losses if not effectively managed. In this study, a deep learning-based computer vision framework was developed for automated weed detection in cassava fields to support precision weed management. The proposed customized convolutional neural network (CCNN) was trained on a field-specific image dataset comprising cassava crops at varying growth stages and under diverse illumination conditions. Experimental results demonstrated that the CCNN achieved a classification accuracy of 99.56%, outperforming traditional architecture such as VGG16 based on hand-crafted features. Furthermore, the model’s capability to perform pixel-wise segmentation of cassava plants, broadleaf weeds, grass weeds and soil illustrates its robustness and adaptability to heterogeneous field environments.

The present study was supported by Cassava Agricultural fields.
 
Disclaimers
 
The views and conclusions expressed in this article are solely those of the authors and do not necessarily represent the views of their affiliated institutions. The authors are responsible for the accuracy and completeness of the information provided, but do not accept any liability for any direct or indirect loss resulting from the use of this content.

The authors declare that there are no conflicts of interest regarding the publication of this article. No funding or sponsorship influenced the design of the study, data collection, analysis, decision to publish, or preparation of the manuscript.