Efficient Weed Detection Using CNN with an Autonomous Robot

Abstract

In this work, Artificial intelligence and IoT based robotic network is proposed to optimize the crop growth in the agriculture fields of Sultanate of Oman. Traditionally, weed detection systems use color and texture features from images. Machine learning algorithms then identify the weeds depending on these features. But the major drawback with feature extraction is the loss of originality, quality of the image, and performance issues. To overcome these issues, we propose an easy and efficient weed detection model using deep-learning techniques. In this research, an image comparison model using convolutional neural networks (CNN) was developed for weed detection. Visual studio code with python programs is used for simulating the model. First, to train the CNN model, we collected a sample of 1300 images from various Potato agriculture farms in Sohar with a pixel size of 512 by 512 (512 * 512) and grouped them into two clusters. Among them, Cluster one comprises 737 weed images, while Cluster two comprises 563 non-weed images. However, loading these images takes more time and requires more memory. It will affect the performance of the model. So, we resized the images to 200 by 200 (200 * 200) pixels and stored them in 2-dimensional array as binary values with a seed value 42. The binary values are stored in a memory as zero (0) for non-weed images and one (1) for weed images. This array of values is given input into a CNN using a rectified linear unit as the activation function for convolution and normalization. As a result, each image will be compared with each other and detect the weeds effectively. Nevertheless, 64 iterations of the model are required to improve its efficiency. Second, the model was tested using random images from both clusters, and it successfully identified weeds and non-weeds. At last, we developed an autonomous robot with an ESP32 microcontroller with motors and embedded it with a Raspberry Pi 3B+ with a camera to test the model efficiency in real time. The robot detected the weed and non-weed images with 95.96% accuracy.