EFFICIENT POSE-BASED HUMAN ACTIVITY RECOGNITION THROUGH SKELETAL FEATURE ENGINEERING AND DIMENSIONALITY REDUCTION

Abstract

Precision agriculture is now an important tool for smart and automated farming that can increase agricultural productivity. Weed infestation is one of the many problems in agriculture which needs attention, as it not only impacts on the growth of crops, but also lowers the crop quality and the use of chemical herbicides. The old methods of weed control involve heavy labour, a long time and are not efficient in large field. Thus, the need for automated crop and weed classification systems based on Artificial Intelligence (AI) and Deep Learning (DL) has grown to be crucial in sustainable and smart farming applications. This research compares several Convolutional Neural Network (CNN) based deep learning architectures for classification of wheat crop and weeds in real agricultural field. The collected self-generated data set comprised 2,190 RGB images of wheat crops and weeds obtained from agricultural fields located at Haryana, India, was used for training and evaluation. This study compares the performance of different deep learning models such as InceptionV3, U-Net, GoogleNet, VGG19, Xception, ResNet50V2, EfficientNetB0, MobileNetV2, VGG16 and DenseNet121. To enhance the generalization of the model and prevent overfitting, image resizing, normalization, and data augmentation (rotation, flipping, and zooming) were applied to the dataset. The proposed architecture compares the various models based on various perfor mance metrics such as accuracy, precision, recall, and f1 score to determine the most effective architecture for intelligent weed detection. Experimental results showed that the Xception model outperformed all models with accuracy of 85%, precision of 86%, recall of 83%, and F1-score of 84%. The main reasons for the superior performance of Xception are its efficient depthwise separable convolution mechanism and its robust feature extraction capability, which allowed to discriminate between similar-looking crop and weed patterns under complex agricultural conditions with high accuracy. Stable and competitive classification performance was observed for DenseNet121, MobileNetV2 and VGG19 while ResNet50V2 and VGG16 had comparatively poor performance because of their inability to process the changes in the field, illumination and overlapping vegetation pattern. The findings of this research can be helpful in utilizing advanced CNN architecture for crop and weed classification and recognition. The developed framework can help in designing intelligent weed detection systems that can be integrated with drones and automated farming to improve efficiency of farming practices. Besides, this research also contributes in reducing the use of chemical fertilizers and human labor by promoting sustainable farming techniques.