Fig 1.
Progress of disease with time.
The appearance of disease symptoms with time on leaves of a diseased plant.
Fig 2.
Block diagram for the proposed algorithm.
Block diagram for the proposed algorithm for detection of diseased plants. The algorithm applies the transformation τ to align the colour image with the thermal image. IL and IR represent visible light images from left and right cameras, IT represents the thermal image.
Fig 3.
Registration of thermal and visible light images.
Thermal image (a), visible light image (b), Overlay of thermal image on visible light image after registration (c). The blue shade represents lower temperature and the red shade represents high temperature values. It can be observed that high temperature stem regions in thermal image faithfully follow the stem regions in the visible light image. Similarly low temperature leaf regions overlap the leaf regions in the visible image.
Fig 4.
Overview of the proposed multi-resolution semi-global matching approach, where ↓ n and ↑ n denote down sampling and up sampling of the image by a factor of n.
Fig 5.
Colour image registered with thermal image (a); Colour image obtained after the background removal (b).
Fig 6.
Identification of potential diseased areas.
Identification of diseased areas in Fig 5 (b) by classification of feature vector at each pixel.
Fig 7.
Classification using local feature set.
Scatterplot of the standard deviation (σp1 & σp2) of data, corresponding to potentially diseased pixels for healthy and diseased plants (day13), along the 1st and 2nd principal components respectively.
Table 1.
Extraction of global feature set.
p-values of the separation power of selected feature set for day 5 to day 13 after inoculation computed using ANOVA.
Fig 8.
Comparison of disparity estimation results.
Disparity estimation results of methods in Section 2.2 on the stereo plant image. The colour bar on the right shows disparity values in pixels.
Fig 9.
Classification results using local feature set.
Average accuracy, sensitivity, specificity and positive predictive value (PPV) results of diseased plant detection using classification of local feature set.
Fig 10.
Comparison of classification results using different combination of features.
Average accuracy of detection algorithm using different combination of global features. a, b, c & d show diseased plant detection results using classification of colour only, colour + thermal, colour + depth and colour + thermal + depth features respectively. Combining colour information with thermal or depth slightly increases the accuracy of the classifier, however combining colour information with thermal and depth improves the accuracy to approximately 70% on day5.
Fig 11.
Classification results using global feature set.
Average accuracy, sensitivity, specificity and positive predictive value (PPV) results of diseased plant detection using classification of global feature set.
Fig 12.
Identification of plants which captured the disease through natural transmission.
Projection of features on 1st and 2nd principal component after performing PCA. The projection shows feature values corresponding to some of the plants which were not inoculated with any disease, occur in disease regions. A couple of these plants are marked as p4 & p47 and are shown in Fig 13.
Fig 13.
Examples which captured the disease through natural transmission.
The plants (a) p47 & (b) p4 shown for illustrative purpose, the plants were not inoculated with any disease but later showed symptoms of the disease. These plants were successfully captured by our novel feature set.
Fig 14.
Comparison of classification results by marking the plants as diseased which captured the disease through natural transmission.
Average accuracy results of diseased plant detection using classification of a) local feature set, b) global feature set, c) local feature set with p4 & p47 marked as diseased and d) global feature set with p4 & p47 marked as diseased.
Fig 15.
Temperature vs Humidity plots as recorded from day5 to day13 of the experiment.