Design of a real-time, embedded weed-detection and spray-control system

An optical weed sensor and a real-time, embedded weed-detection and spray-control system were developed and evaluated in this study.; Spectral reflectance characteristics of various plant species and soil were analyzed in order to select feature wavelengths that maximized the contrasts between major object categories—weed leaves, weed stems, crop leaves, crop stems, and soil—for weed detection. Relative color indices insensitive to illumination variation were developed at these feature wavelengths. Several calibration models for differentiating weeds against crops and soil were developed using statistical methods of partial least squares and discriminant analysis. The best classification model achieved classification rates of 98.3%, 98.7%, and 64.3% for wheat, bare soil, and weeds, respectively. The classification rates achieved for the validation data set were 83.1%, 79.5%, and 62.5%, respectively.; Based on the classification model, an optical sensor was designed and tested. The effective sensing area of the sensor was determined through a laboratory test. The sensor was tested on different weed densities. When multiple weeds shared the sensor's effective sensing area with soil, a classification rate of 65% was achieved for weeds. The classification rate fell below 50% for single weeds. However, under field conditions, the sensor successfully detected weeds at densities of 0.5 plants/dm2 or above with a classification rate of higher than 96.9%.; Two optical weed sensors and their control modules, a central-control module, a GPS device, and a spray-control module were successfully integrated into a realtime, embedded system. The system components were networked using a Controller Area Network. The system was tested extensively in two wheat fields. With good training, the system generally reached weed-detection accuracies of over 80%. The addition of a light-blocking screen and artificial lights facilitated the use of the system under variable light conditions, including night operations. Classification models trained with multiple weed species improved the classification accuracy. Classification accuracy also was affected by the position of the sensor relative to the training sample during training. At the current design, the total cost for hardware of the system with two weed sensors is about {dollar}2,500.*; *This dissertation includes a CD that is compound (contains both a paper copy and a CD as part of the dissertation). The CD requires the following applications: Windows MediaPlayer or RealPlayer.