Abstract: Abstract: Weeds are located in the seedling line zone area near the root zone of rice. It is inevitable for the weeds to compete with the rice for the water and nutrients, leading to the reduction of rice yield and quality. An accurate assessment of weed density can greatly contribute to the weed control and ecological benefits. However, the commonly-used visual perceptions cannot assess the weed density under the background disturbance and leaf shading of rice plants. In this study, a novel tactile perception was proposed to accurately evaluate the weed density grade in the seedling line zone of the paddy fields. Firstly, the polyvinylidene fluoride (PVDF) sensor was wrapped with a mixture of two properties of liquid silica gel. A tactile sensor was then obtained to produce the voltage changes after solidification using micro deformation. As such, the tactile sensor was also applied to evaluate the weed density in the seedling line zone. Secondly, an accurate extraction was performed on the joint features in the time domain and mel-frequency cepstral coefficients in the tactile data. The joint features were then input into the principal component analysis. The optimal feature dimensions were determined to construct the feature vector of weed density assessment. A comparison of multiple models was conducted to optimize the classification and assessment of the weed density model. Classical machine learning models were selected, such as the support vector machines (SVM) with the high generalization capability, K-nearest neighbors for classification and regression, random forests (RF) with the a high resistance to overfitting, and neural networks (NN) to implement the complex learning tasks. More importantly, the newly developed random subspace K-nearest neighbor model was selected using multi-classifier fusion and Random Sparse Sampling Ensemble Learning model with the adaptive balancing sample class function of random under-sampling integrated learning models. The training set, validation set, and test set were randomly selected during this time. The average value was taken as the accuracy of weed density level after assessment. Finally, a weed density level assessment model was constructed using the NN. The effectiveness of the joint features was also verified to extract from the tactile signals, and then to compare the accuracy of the assessment of different models. The best feature dimensions of the different models were selected to compare the assessment accuracy of weed density levels in the different dimensions. Furthermore, the tactile sensor speed and contact height were selected to verify the accuracy and stability of weed density assessment. As such, the experimental factors were taken as the most influence on the assessment. Meanwhile, the regulation was clarified for the effect of tactile sensor speed and contact height on the performance of weed density assessment after experiments. Specifically, the barnyard grass seeds were randomly scattered by hand within 5 days of rice transplanting, according to the natural growth pattern of rice and grass in the transplanted rice fields. The experimental results showed that the artificial NN model was provided the best evaluation accuracy using the optimal feature dimension. Once the tactile sensor was in the low contact mode with the stalk, the accuracy and stability of weed density assessment decreased as the paddy machine travel speed increased. When the tactile sensor was in the high contact mode with the leaf canopy, the weed density assessment accuracy increased and stability decreased, as the paddy machine travel speed increased. Consequently, the best overall performance was found at the travel speed was 0.5 m/s under the low stalk contact. The accuracy of weed density assessment was 85.33% at this time. The coefficient of variation was 2.73% for the accuracy of weed density assessment. A systematic optimization can be expected to determine the weed control operation mode and application system parameters, in order to further investigate the weed data set expansion and tactile feature extraction. The finding can also provide a more practical for the quantitative application and decision making.