Automated Plant Count Using Unsupervised Classification on UAV Acquired Imagery

Precision farming is a farm management technique where in observing, measuring and responding is carried out using the latest technology. Its end output is to preserve the resources used in farming while optimizing on the returns. Previously satellites were used for this purpose and it had a lot of drawbacks of weather anomalies, low resolution and lack of real time data. This entire process can be done using an UAV system. In this study we focus on optimizing the time taken and maintain the accuracy of an industrial process of counting plants. Supervised classification is the most preferred method of classification as there is control over the classes and its well-established accuracy. But the main drawback is time taken in training the classifier. As there are more than hundreds of farms plots the same signature file cannot be used as there will be variation in the lighting conditions and shadows patterns. To solve this, we have used iso-cluster unsupervised classification and grouped the classes into plants and non-plant region. The accuracy stood at 95.4% compared to the accuracy of 97.8% obtained from supervised classification. This was within the 5% inaccuracy limit specified by the client. The major gain was the reduction of the time spent on the process. The supervised classification method took about 35 minutes whereas the unsupervised and grouping method took 10 minutes to complete the process for a 1.5 acres farm plot. This is a reduction of 70% of the time taken which is a very significant when plant counting has to be done for hundreds of plots.