Pixel classification methods for identifying and quantifying leaf surface injury from digital images

• Many factors such as pollution and disease cause injury on plant leaves.
• We test four approaches for finding injured pixels in digital images of plant leaves.
• A linear discriminant analysis classifier provided the best pixel classification.
• The LDA approach was robust, computationally efficient and had 95% mean accuracy.
• A feature vector based on pixel RGB colour intensities provided good discrimination.

Plants exposed to stress due to pollution, disease or nutrient deficiency often develop visible symptoms on leaves such as spots, colour changes and necrotic regions. Early symptom detection is important for precision agriculture, environmental monitoring using bio-indicators and quality assessment of leafy vegetables. Leaf injury is usually assessed by visual inspection, which is labour-intensive and to a considerable extent subjective. In this study, methods for classifying individual pixels as healthy or injured from images of clover leaves exposed to the air pollutant ozone were tested and compared. RGB images of the leaves were acquired under controlled conditions in a laboratory using a standard digital SLR camera. Different feature vectors were extracted from the images by including different colour and texture (spatial) information. Four approaches to classification were evaluated: (1) Fit to a Pattern Multivariate Image Analysis (FPM) combined with T2 statistics (FPM-T2) or (2) Residual Sum of Squares statistics (FPM-RSS), (3) linear discriminant analysis (LDA) and (4) K-means clustering. The predicted leaf pixel classifications were trained from and compared to manually segmented images to evaluate classification performance. The LDA classifier outperformed the three other approaches in pixel identification with significantly higher accuracy, precision, true positive rate and F-score and significantly lower false positive rate and computation time. A feature vector of single pixel colour channel intensities was sufficient for capturing the information relevant for pixel identification. Including neighbourhood pixel information in the feature vector did not improve performance, but significantly increased the computation time. The LDA classifier was robust with 95% mean accuracy, 83% mean true positive rate and 2% mean false positive rate, indicating that it has potential for real-time applications.