The study employed rice and barnyard grass specimens cultivated indoors under meticulously controlled conditions. These conditions ensured a temperature range of 25-28°C, humidity levels ranging between 40-70%, and a regulated 12-hour exposure to light. Only specimens of impeccable quality, devoid of any pest infestations or diseases, were chosen for leaf hyperspectral data collection. Spectral reflectance measurements of the leaves were conducted using a ground truth spectrometer (Field Spec Pro FR2500 spectrometer). This device boasted a wavelength range spanning from 350 to 2500 nm, with a sampling interval of 1.4 nm and a spectral resolution of 3 nm within 350-1000 nm range, and a sampling interval of 2.0 nm with a spectral resolution of 10 nm within 1000-2500 nm range. Leaf blade measurements were meticulously performed using leaf clamps within the canopy position of plants. Prior to each measurement, adjustments to the spectrometer parameters were meticulously carried out using the RS3 software, and spectral calibration against a whiteboard was conducted. To ensure robustness and accuracy, five consecutive samples were collected for each measurement, and the average value was calculated. A total of 660 samples were collected across three batches, comprising 313 rice and 347 barnyard grass specimens. The training set included 495 samples, incorporating data from the first two batches, with 243 rice and 252 barnyard grass samples. The test set, derived from the third batch, consisted of 70 rice and 95 barnyard grass specimens.

The dataset amassed in this investigation comprises spectral band reflectance data, with each sample encapsulating 2151 band reflectance values ( Figure 3A ). To fortify the reliability of the spectral data for subsequent analysis, four preprocessing methodologies were employed: standard normal variate transformation (SNV) (Bi et al., 2016), moving average (MA) (Timothy and Anuradha, 2023), savitzky golay smoothing (SG) (Luo et al., 2005), and mean centering (MC) (Ijomah, 2019). These methods served to mitigate data noise and rectify wavelength shifts, thus enhancing the overall quality and integrity of the spectral data. SNV, for instance, functions to rectify scale differences such as tilts or spikes, thereby accentuating the dynamic components of the spectral data. MA, on the other hand, operates by smoothing the dataset through point averaging within a window, effectively mitigating noise influence and suppressing periodic noise. SG, renowned for its efficacy in managing spikes and data variations, smooths the signal by attenuating high-frequency noise components. Lastly, MC plays a pivotal role in dataset rectification, eradicating biases stemming from measurement disparities, inherent variations, or substantial value discrepancies. The original spectral reflectance data and preprocessed data are shown in Figure 2 .

Consider a comprehensive breakdown of the preprocessing steps: employing a window length denoted as “w” and a sliding step designated as “s”, the preprocessed hyperspectral data, comprising “m” bands, is meticulously partitioned into (m-w)/s+1 segments. Notably, the final subsequence necessitates an extension in reverse to address any potential length discrepancies. Following this segmentation process, the divided subsequences undergo transformation into a two-dimensional matrix, as depicted in Figure 3B . This transformation sets the stage for the extraction of local band correlation features through the utilization of a sophisticated deep convolutional network.

Following the subsequence transformation, each sample is epitomized by an n×w two-dimensional matrix, wherein n signifies the number of subsequences and w denotes the window length. This matrix stands as the foundational input layer of DeepBGS. Subsequently, the tensor of the n×w matrix traverses through the initial convolutional layer, endowed with rectified linear unit (ReLU) activation and a batch normalization (BN) layer. This layer, housing 16 filters of size 3×4 with a stride of 1, amplifies the dimensionality of the initial single-channel data, thus enhancing network expressiveness. The data then progresses through three consecutive feature extraction convolutional modules, each composed of two 3×3 convolutional layers, an attention mechanism layer, and a 2×2 max pooling layer. Notably, the attention mechanism layer harnesses a 3×3 convolutional kernel from the convolutional block attention module (CBAM), bolstering network feature expression and local region expression by integrating both channel and spatial attention. Meanwhile, the maximum pooling layer down samples the feature layer with a step size of 2×2. These three convolutional modules contain 32, 64, and 128 filters, respectively. Subsequent to this, the data is downscaled using two convolutional layers, each containing 64 filters, and then merged into a two-layer long-short term memory (LSTM) module with 32 hidden units, and two ReLU-based fully connected layers, featuring 32 and 64 output units, respectively, are incorporated. Regularization is implemented with a dropout probability of 50%, fortifying model generalization. Finally, the output layer employs the Softmax function to yield the probability of a sample belonging to rice. The parameters of DeepBGS are optimized using the Adam Optimizer, with a learning rate of 5e-6, a weight decay parameter of 0.0001, and training epochs and batch sizes set at 150 and 11, respectively. The structure of DeepBGS is shown in Figure 3C .

Four traditional machine learning methods, including decision tree (DT) (Song and Lu, 2015), random forest (RF) (Breiman, 2001), SVM (Hearst et al., 1998) and extreme gradient boosting (XGBoost) (Osman et al., 2021) classifiers were used alongside four deep learning models—DeepBGS-NoLSTM, ResNet, VGG and multi-layer perceptrons (MLP) (Riedmiller, 1994) — to evaluate the performance of DeepBGS model. The DT model was implemented using the R package part 4.1-15, and the tree was pruned by the optimal C parameter value with the least cross-validated error. For the RF model, The R package random Forest 4.6-14 was utilized, with parameters ntree and mtry set to 500 and No .  of features , respectively. The radial basis function (RBF) based-SVM model was executed using the Library for SVM, available at https://www.csie.ntu.edu.tw/%7Ecjlin/libsvm/index.html, A 10-fold cross-validation (10 CV) approach on the training set was used to optimize penalty parameters C (C∈[2^-5, 2¹⁵]) and g parameter g (g∈[2^-15, 2³]). XGBoost was implemented based on the Python sklearn library with default parameters. The MLP consists of three fully connected layers, where each hidden layer followed by a non-linear transformation employing the ReLU activation function. The output parameters of the initial two hidden layers were configured to have 1024 units, while the output layer employed the softmax activation function. DeepBGS-NoLSTM was akin to DeepBGS but the LSTM module was replaced by full connectivity layer. ResNet and VGG adopted 18-layer and 10-layer network architectures, respectively, which were implemented using the PyTorch framework.

The primary challenge inherent in the analysis of multivariate hyperspectral data analysis lies in the necessity to represent a considerable number of wavebands, which significantly elevates the dimensionality of the data. Therefore, performing hyperspectral data dimension reduction becomes essential to mitigate data redundancy and extract valuable knowledge (Sarić et al., 2022). In this study, three types of feature engineering methods were explored, including feature dimensionality reduction, feature selection, and vegetation indices, to enhance the model’s prediction performance with high-dimensional bands.

Two techniques, principal component analysis (PCA) (Wold et al., 1987) and t-distributed stochastic neighbor embedding (t-SNE) (Belkina et al., 2019), were employed for feature dimensionality reduction. The determination of final dimensions was grounded on the principle of achieving optimal 10-fold cross-validated accuracy within the training set. Implementation of both PCA and t-SNE was facilitated through the Python sklearn library. Successive projections algorithm (SPA, https://gitee.com/aBugsLife/SPA) (Araújo et al., 2001) and consecutive adaptive reweighted sampling (Li et al., 2009) (CARS, https://gitee.com/aBugsLife/CARS) were utilized for the selection of important hyperspectral spectral bands. SPA employs a method that initially selects variables with the lowest covariance and redundancy while maximizing the projection vector in vector space. Subsequently, the final retained bands are determined based on the optimal principle of 10-fold cross-validated accuracy within the training set. Conversely, CARS utilizes a different approach. It began by retaining points with larger absolute weights of the regression coefficients in the partial least squares regression (PLS) model through adaptive reweighted sampling (ARS). It then iteratively builds PLS models based on the new subset, selecting wavelengths with the smallest root mean square error (RMSE) of cross-validation for the PLS model as the characteristic wavelengths after several calculations.

Ultimately, we diminished the dimensionality of the original data features by extracting spectral vegetation indices. A total of types of vegetation indices were extracted, including normalized difference vegetation index (NDVI), ratio vegetation index (RVI), triangular vegetation index (TVI), photochemical reflectance index (PRI), and normalized pigment chlorophyll ratio index (NPCI). These indices were constructed using the following formula:

The bands r ₁ and r ₂ correspond to distinct regions of the light spectrum, with r ₁ encompassing the near-infrared region and r ₂ encompassing the infrared region. Specifically, r ₁ comprises wave lengths of 724nm, 738nm, 750nm, 764nm, 776nm, 790nm, 802nm, and 814nm, while r₂ consists of 601nm, 605nm, 614nm, 627nm, 6336nm, 644nm, 652nm, 660nm, 669nm, and 677nm. Leveraging these bands, we derived 80 features for the NDVI and 80 features for the RVI. Additionally, we formulated a TVI feature utilizing wavelengths 750nm, 550nm, and 650nm, a PRI feature using wave-lengths 531nm and 570nm, and an NPCI feature using wavelengths 680nm and 430nm. In total, 163 vegetation index features were extracted from the hyperspectral data.

Accuracy (ACC), area under the curve (AUC), and matthews correlation coefficient (MCC) are utilized as evaluation metrics to assess the performance of the model predictions. These metrics are defined as follows:

Where TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively. The receiver operating characteristic (ROC) curve illustrates the true-positive rate against the false-positive rate (1 - specificity) for various thresholds. The AUC of the ROC curve is analyzed to provide a comprehensive metric for evaluating prediction methods. Higher values of ACC, MCC, and AUC indicate better prediction ability.

We initially assessed the performance of traditional machine learning models in aiding the hyper-spectral differentiation between barnyard grass and rice during the seedling stage. Additionally, we investigated the impact of various data preprocessing methods on model performance. Hyperspectral data, subjected to diverse various preprocessing techniques, constituted the input data for the traditional machine learning models, and the prediction results for independent test set of each model are illustrated in Figure 4 . Among the four traditional machine learning models, XGBoost exhibited the most predictive performance, with average ACC, MCC, and AUC values reaching 0.9261, 0.8499, and 0.9579, respectively. Following XGBoost, the SVM model emerged as the second best, achieving average ACC, MCC, and AUC of 0.9224, 0.8415, and 0.9565, respectively. In contrast, the DT model displayed the lowest prediction accuracy, with average ACC, MCC, and AUC metrics of only 0.8982, 0.7920, and 0.8920, respectively. Compared to the original data, all four types of preprocessing methods effectively improved the model prediction performance, with SNV exhibiting the most notable improvement. SNV yielded average ACC, MCC, and AUC values of 0.9212, 0.8420, and 0.9653, respectively. Overall, irrespective of preprocessing or the choice of machine learning algorithms, the ACC and MCC of the hyperspectral-based machine learning model consistently surpassed 0.88 and 0.76, respectively, in distinguishing between barnyard grass and rice during the seedling stage. These findings underscore the potential of hyperspectral data in distinguishing between barnyard grass and rice at the seedling stage.

To enhance the predictive accuracy of our models, we employed three distinct feature engineering strategies aimed at reducing its dimensionality: feature dimensionality reduction, feature selection, and vegetation index conversion. We applied spectral data after SNV pre-processing for both feature dimensionality reduction and feature selection, while vegetation index conversion was conducted based on the original spectral reflectance data. For PCA and t-SNE, we determined the optimal retained dimensionality by evaluating 10-fold cross-validation ACC using XGBoost on training set. PCA retained thirty principal components, while t-SNE retained twelve features. The SPA band selection process and the CARS band selection process were visualized in Figure 5 . As the number of characteristic bands increases, the RMSE value decreases and levels off. When the RMSE stabilizes and reaches its optimum, the value is 0.2462. At this point, the SPA retained 16 bands, comprising wavelengths such as 678nm, 457nm, 693nm, 1410nm, 2182nm, 1951nm, 354nm, 2481nm, 376nm, 2483nm, 381nm, 374nm, 1619nm, 2485nm,2486 and 2488nm. In Figure 5A , the number of bands gradually decreases and stabilizes, while Figure 5B shows a decreasing trend followed by an increase. At a sampling number of 21, the RMSECV reaches its minimum, leading to the selection of the bands obtained from the 21^th sampling as the characteristic wavelengths, totaling 108 wavelengths. Additionally, we derived 163 vegetation index features through band conversion.

Drawing from important features retained by various feature engineering strategies, we constructed traditional machine learning models and presented the results of independent tests in Table 1 . Utilizing vegetation index features, the average ACC and MCC of the four machine learning models were 0.9152 and 0.8311, respectively, notably lower than those of the prediction model based on the full band, where the average ACC and MCC were 0.9212 and 0.8420, respectively. Following PCA downscaling, the model prediction accuracy showed slight improvement Compared to the full band except for the DT model, registering an average ACC and MCC of 0.9353 and 0.8712, yet, the DT model experiences a slight reduction, primarily observed in the decline of ACC and MCC to 0.8242 and 0.6583. However, t-SNE, SPA, and CARS feature downscaling methods exhibited enhanced model prediction performance compared to full-band modeling. Among these, the t-SNE feature dimensionality reduction method demonstrated the most significant improvement, with an average ACC and MCC of 0.9409 and 0.8800, respectively. The SVM model based on the retained features of SPA exhibited the best prediction outcomes, attaining an ACC and MCC of 0.9455 and 0.8886, respectively. These findings indicate the effectiveness of feature dimensionality reduction methods, particularly t-SNE, in enhancing the performance of traditional machine learning models.

Transforming linear hyperspectral reflectance data into 2D matrix data through subsequence division is crucial for extracting spectral global and local correlation features by DeepBGS. Therefore, our initial focus lies in optimizing the subsequence window length and sliding step length to select appropriate parameters for the DeepBGS model. Through grid optimization grounded in 10-fold cross-validation of the training set, we explored various combinations of window lengths and sliding step lengths within the range of 50-300, with a step length of 50. Previous results have shown that the models achieve the best prediction performance following the preprocessing of original data using SNV. Hence, DeepBGS modeling exclusively relies on SNV preprocessed data, and the 10-fold cross-segmentation of the training set remains consistent across all combinations of window length and sliding step length. The model prediction outcomes across different combinations of window length and sliding step length were visualized in Figure 6 . Notably, when the window length is set to 200 and the sliding step length to 150, the DeepBGS model attained its highest training performance, with ACC, MCC, and AUC reaching 0.9979, 0.9958, and 0.9999, respectively.

Based on the optimal combination of window length and sliding step (w = 200, s = 150), we further compared the performance of deep convolutional networks for extracting sequence matrix features. As illustrated in Table 2 , Figure 7 , all five deep learning models demonstrate robust capabilities in distinguishing seedling barnyard grass from rice, with ACC surpassing 0.93 and MCC exceeding 0.87. Remarkably, the DeepBGS model emerges as the top performer, achieving a flawless differentiation rate of 98.18%. All 95 barnyard grass samples were correctly classified, with only 3 out of 70 rice samples misclassified as barnyard grass ( Table 3 ). In contrast, the DeepBGS-NoLSTM model, which excludes the LSTM module, experiences a slight reduction in independent test accuracy, primarily observed in the decline of ACC and MCC to 0.9697 and 0.9379. The VGG model, without the batch normalization layer, exhibited a more pronounced decrease in prediction performance, with ACC and MCC plummeting to only 0.9515 and 0.9014, respectively, comparable to traditional machine learning models. And the MLP module attained ACC and MCC scores of 0.9394 and 0.8779. Meanwhile, the Resnet18 model, which featured an 18-layer network, attained ACC and MCC scores of 0.9636 and 0.9256, respectively, underscoring its efficacy in distinguishing between barnyard grass and rice at the seedling stage. However, augmenting the network depth to 50 layers resulted in diminished prediction accuracy for the Resnet50 model, with ACC and MCC declining to 0.9455 and 0.8886, respectively.