Recent advancements in hyperspectral imaging (HSI) for early disease detection have shown promising results, yet there is a lack of HSI data representing the responses of plants at different disease stages as measured by in planta pathogen density. To address this question, we used bacterial leaf spot of tomato as a model and collected hyperspectral images and in planta pathogen populations for seven days. Machine learning models were trained using data from leaf-level full spectra, vegetation index (VI), and pixel-level full spectra at four disease stages. The results suggest that HSI can detect tomato foliar disease at pre-symptomatic stages and differentiate bacterial disease spots from abiotic leaf spots. Using VI data as features for machine learning improved overall classification performance by 26-37% as compared to the direct use of raw data. Critical wavelength bands and VIs varied across disease progression stages, suggesting that pre-symptomatic disease detection relied more on changes in leaf water content and plant defense responses rather than changes in leaf pigments or internal structure, which are crucial during symptomatic stages. In conclusion, this study reveals the potential benefits of leaf structure segmentation and VI group pattern analysis in using HSI for the early detection of leaf diseases.
Article
Hyperspectral Imaging Analysis for the Early Detection of Tomato Bacterial Leaf Spot Disease
https://doi.org/10.21203/rs.3.rs-3778695/v1
This work is licensed under a CC BY 4.0 License
You are reading this latest preprint version
Over the past few years, climate change has brought many challenges to global crop production and food security, including a growing concern about emerging crop diseases1-3. Plant pathogens can spread through shipments of stock plants between production facilities, and from production facilities to growers or to retailers and consumers4-6. Failing to prevent the introduction and spread of plant diseases can cause up to 40% yield loss in agricultural production7. A central challenge in preventing accidental pathogen dissemination and disease outbreaks is the difficulty in detecting many plant diseases at their early stages8,9 As a result, asymptomatic plants can spread undetected pathogens when being shipped from one location to another. Due to these challenges, early disease detection has become a key focus in plant disease surveillance.
Traditionally, disease detection and diagnosis rely on methods such as microscopic observation, culturing, immunological assays, and molecular tests1,10, however, many conventional methods are not efficient in tackling emerging crop diseases1,10,11. There is a pressing need for new tools that offer rapid, sensitive, and cost-efficient plant disease detection and diagnosis12,13, and imaging-based remote sensing techniques11,14-16 appear to be one promising solution which can provide high-throughput and non-invasive disease detection17-20, facilitating the automation of crop disease detection in both field and controlled environments12,13,21. Major imaging-based disease detection approaches include hyperspectral imaging (HSI), multispectral imaging (MSI), RGB imaging, thermal imaging, and fluorescence imaging12,13,17.
Hyperspectral and multispectral imaging have been applied using satellites, airplanes, unmanned aerial vehicles, ground vehicles, and lab devices19,22, providing imaging data at various scales. In these studies, early detection was achieved through recording and analyzing the spatial and spectral patterns of plant pixels at different wavelength bands13,19. For example, in-field detection of Alternaria solani in potato16, Candidatus Liberibacter solanacearum in carrot23, grapevine vein-clearing virus in grapevines24, and multiple pathogens in winter wheat25 have been achieved using HSI or MSI. The accuracy in those studies ranged from 60% to 93%. Studies conducted in controlled environments also showed promising results. A previous study showed that five different foliar diseases of wheat can be detected by using HSI under greenhouse conditions, with an average accuracy over 90%14. Similarly, HSI was used to achieve the detection of Phytophthora infestans and Alternaria solani on potato26, as well as the detection of Sclerotium rolfsii on peanut27, with an accuracy higher than 80%. Pre-symptomatic detection was achieved with an accuracy higher than 60-80%26,28. Previous studies revealed multiple bands in VIS (around 500 nm, 550 nm, and 690 nm), Far-red (around 750 nm), and SWIR (around 1890 nm, 2,300 nm, and 2,380 nm) regions that are informative to differentiate between healthy and diseased crops 16,23-26,28,29.
Although the above-mentioned studies revealed the potential of applying spectral imaging to crop disease detection, especially at early stages, there are some factors that limit the information provided by the spectral data. Some utilized cameras that only collect data from a small spatial area14,26,27,29. Some used multi-spectral (MSI) cameras that collect data from a few wavelength bands25,30. In a few studies that focused on pre-symptomatic detection, the absence of pathogen population in asymptomatic plant samples was not validated24,27, or was not confirmed for every sample used for imaging31. Additionally, there is a lack of information on environmental/plant or abiotic factors that might influence plant reflectance and confound disease detection.
In this study, we used bacterial leaf spot (BLS) of tomato caused by Xanthomonas perforans32-34 as a model system to characterize changes in leaf reflectance from before inoculation through disease onset to the advanced stage of disease progression (Figure 1). Tomato has been widely used as a model system in HSI-based early disease detection of fungal35-37, viral38, and bacterial leaf diseases39. However, there is a lack of HSI data on disease progress specific to tomato BLS or information on differentiating abiotic spots from bacterial spots. In order to differentiate uninoculated from infected tomato leaves at early and late stages of disease progression in this controlled environment study, we collected HSI data at relatively high spectral and spatial resolution, along with pathogen population data to capture disease progression dynamics in planta. Our hypothesis was that HSI combined with machine learning would be capable of early detection of BLS on tomato seedlings, which could be useful in monitoring tomato transplants grown in greenhouses so that infected seedlings can be removed to prevent outbreaks in tomato production fields40.
The objectives of this study included: 1) to examine the influence of background, leaf size, and leaf structure on tomato leaf reflectance, 2) to compare the differences between uninoculated and infected tomato leaves at pre-symptomatic and symptomatic stages during BLS disease progression, 3) to identify key hyperspectral and VI signatures by analyzing leaf reflectance data at pixel and leaf level with multiple machine learning algorithms, 4) to compare the performance of classification models generated from this study, and correlate the model performance with disease development inferred from measurements of pathogen population size. We anticipate that the outcome of this study can provide practical recommendations on applying HSI to achieve early detection of BLS in tomato transplant production facilities, which may also benefit adapting HSI to early disease detection in other controlled crop production environments.
Disease Progression and Bacterial Growth
All tomato plants inoculated with X. perforans showed initial symptoms on day 4 (4dai, Figure 2). Initial symptoms appeared as wilting on leaf margins and small water-soaking lesions. The initial disease incidence was approximately 18%, and the disease severity index was 7%. Disease continued to progress from 4 dai to 7dai, with leaf spots becoming bigger and darker (brown or dark) whileleaves wilted, curled, and decayed. The culture (isolation) results showed that X. perforans was absent before inoculation in both inoculated and uninoculated groups. However, X. perforans reached 105 CFU/ cm2 leaf area in inoculated plants, merely two hours after inoculation (Figure 2). Bacterial populations continued to grow from 0 dai to 3 dai, even though no visible leaf symptoms were observed. The growth curve indicated that bacterial growth was exponential throughout the experiments (7 days) and peaked on day 6 or 7 at 108 CFU/ cm2 leaf area. All uninoculated plants remained healthy-looking and free of X. perforans. Although a few uninoculated plants showed white paper texture spots on some (around 15%) leaves, the culture isolation results suggested there was no pathogen present in those leaves. Therefore, those white spots were labeled as abiotic spots.
Scan precision and impacts of background, leaf size, leaf structure on tomato leaf reflectance
LDA results suggested reflectance data of the same leaves were similar between scans. Healthy leaf data showed clear separation from diseased leaf data (Figure 3A). Linear Discriminant 1 (LD1) value was 0.85, indicating major differences between healthy and diseased groups, while LD2 value was 0.09, suggesting close similarity between VISNIR data from two scans on the same leaves (similar trend was observed in SWIR data). There was a clear separation between data collected on the white background reference panel (Figure 3B, green and pink) and data collected on the black background (Figure 3B, purple and beige), while the healthy (uninoculated, green and purple) and diseased leaf data (pink and beige) distinctly separated. The reflectance data did not seem to be strongly affected by leaf size (Figure 3C). The data points from large leaves and small leaves were close (LD2=0.00) compared to the distance between leaf pixels and background pixels (LD1=1.00, Figure 3C). In contrast, pixels were clearly grouped by different leaf structures based on the reflectance data (Figure 3D). Especially, the pixels at leaf edges (apex/tip and margin) were far from the pixels at the inner areas (midrib, veins, and interveinal areas).
Comparison of Models Trained with leaf-level full spectra data
To evaluate the performance of classification models trained with leaf-level full spectra, the testing accuracy and weighted-averaged F1 scores were compared for four stages during disease progression. Model performance varied between disease progression stages and between VISNIR and SWIR wavelength ranges (Table S1). Model performance increased from the pre-symptomatic (beginning and early) to the post-symptomatic (mid and late) stages. Among LDA, SVM, KNN, RF, GB, XGB, and MLP, the LDA models generally had the best performance for VISNIR and SWIR data across all four stages (Table S1).
At 0 dai, model performance was relatively low, with RF having highest accuracy (0.38) for VISNIR and LDA having highest accuracy (0.63) for SWIR. When we performed a binary classification using uninoculated (data selected from both two groups of plants before inoculation and the uninoculated group 2 hours after inoculation) versus inoculated plants (2 hours after inoculation), the VISNIR testing accuracy (0.61±0.05) and the F1 score (0.47±0.02) were much higher. The same trend was observed with the SWIR data across all models (testing accuracy: 0.57±0.10, F1 score: 0.45±0.05). At the early stage, LDA models performed the best for both VISNIR (0.55) and SWIR data (0.64). At mid stage, model performance generally increased by 0.1 up to 0.74. This level of accuracy and F1 score showed that some of the classes were not separable, for example, in Figure 4A, uninoculated and inoculated leaves were considered as two separate groups, however, both samples included health leaves and the LDA plot showed that these data points could not be separated.
LDA was able to separate inoculated from uninoculated leaves as early as 2 hours after inoculation while no difference was found before inoculation, with either VISNIR or SWIR data (Figure 4A). At the early, pre-symptomatic stage (1 dai to 3 dai), VISNIR data revealed a difference between inoculated and uninoculated leaves in LDA, but SWIR data failed to do so (Figure 4B). At the disease onset stage (4 dai), LDA showed clear separation of data points, (Figure 4C) for VISNIR data. SWIR data did not show a clear differentiation for 4 dai. However, the performance of cross-validation (F1 score and accuracy in Table S1) was not high. This means the model was not able to generalize to the leave-out-testing data. Both VISNIR and SWIR showed a difference between inoculated and uninoculated leaves across 5 dai (Figure 4D), 6 dai, and 7 dai (Figure 4D). In all these analyses, the first linear discriminant (LD1) accounts for most (pre-symptomatic, 0dai-3dai: 59% - 89%; post-symptomatic, 4 dai to 7dai: 68% - 82%) of the between-class variance (Figure 4A-D).
In terms of hyperspectral signatures, inoculated samples were generally associated with lower reflectance intensity in the 800-1000 nm NIR region and 1400-1500 nm SWIR region (Figure S1). As disease progressed, the NIR reflectance difference between uninoculated and inoculated samples became larger. However, the SWIR reflectance difference became smaller at the pre-symptomatic stage, and the reflectance intensity of inoculated samples continued to increase and became higher than that of uninoculated samples. To understand which wavelengths played important roles, we calculated the gini importance (Table S2). At 0 dai of disease progression, multiple wavelength bands in the red edge/far-red (740-750 nm, and 750-780 nm), near-infrared (1000-1025 nm), and short-wave infrared (around 1400 nm, 1700 nm, 1529 nm, 1380 nm, 1007 nm, etc.) regions were important in differentiating infected from uninoculated tomato leaves (Figure 5A). At the early stage (1-3 dai, pre-symptomatic), wavelength bands in far-red (around 743nm, 712nm, and 733 nm) and short-wave infrared (around 1700nm, 1600nm, 1400 nm, etc.) were still important, and several bands in the purple-blue regions (around 393 nm, 404nm, 425nm, and 478 nm) became important (Figure 5B). At mid stage (4-5 dai, symptomatic), wavelength bands in far-red (around 690 nm and 750 nm) and near-infrared (750-780 nm, around 850 nm, 900-930nm, around 1000 nm, 1050-1100 nm, etc.) stood out (Figure 5C). At the late stage (6-7 dai), the important bands shifted to the red region (600-630 nm) and short-wave infrared bands around 1400 nm showed higher importance than other wavelengths (Figure 5D).
Comparison of Models Trained with leaf-level VI data
Instead of using all spectral bands, we tested whether using specific VI can improve machine learning prediction. Using the VIs extracted from the spectral data (collected from 0 dai to 7 dai), all machine learning models showed good performance across all four stages during disease progression (Table S3). The accuracy and F1 scores were 1.00, expect that MLP had an accuracy of 0.88 for the “beginning”, 0.67 for the “mid”, and 0.74 for the “late” stage (Table S3). The weighted-averaged F1 scores were the same as the accuracy scores. These results indicate that calculating Vegetation Indices (VIs) and using them as features for machine learning significantly improved overall performance in classifying diseases compared to using raw data directly.
VIs (Table S4) that were important in differentiating infected tomato leaves varied between different disease progression stages (Figure 6). At the beginning (0 dai and 2 hours ai), CTR2 was the most important VI, followed by mNDVI, MSI, NDVI, NDVI705, SIPI, PRI, and WI (see discussion). At the early stage (1-3 dai), LLSI, NDWI-hyperion, MSI, PSSR, CTR2, and NDVI were relatively more important than the other VIs. After the onset of visible disease symptoms, NDVI, PRI, and PSSR had higher importance scores than other VIs. At the mid stage, BRI2, CTR2 and SIPI were also important, and mNDVI705 and LLSI had higher importance scores than other VIs at the late stage, in addition to NDVI, PRI, and PSSR. These results align with the known effects of those VIs and revealed plant physiological and biochemical function specific VI group patterns that have not been discussed before in research.
Comparison of Models Trained with pixel-level full spectra data
In the pixel-level full spectra classification models, pixels sampled from 7 dai leaf images were classified into nine classes, including green areas on the edge of uninoculated healthy leaves (GH-e), green areas in the interveinal areas of uninoculated healthy leaves (GH-iv), bacterial spots on the edge of infected leaves (BS-e), bacterial spots in the interveinal areas of infected leaves (BS-iv), abiotic spots on the edge of uninoculated leaves (AS-e), abiotic spots in the interveinal areas of uninoculated leaves (AS-iv), green areas on the edge of symptomatic tomato leaves (GS-e), green areas in the interveinal areas of symptomatic tomato leaves (GS-iv), and background (bg) pixels. Overall, RF models showed better performance than others, with VISNIR (0.63±0.13) and SWIR (0.57±0.16) data (Figure 7A). Although the SVM model (0.60±0.18) trained with SWIR data was slightly better than RF, yet the difference was not statistically significant (Figure 7A). Even though the performance of LDA varied among repetitions (Figure 7A), LDA plots showed interesting results (Figure 7B): Background pixels were distinctly different from leaf pixels in both VISNIR and SWIR ranges. Abiotic spots formed two adjacent clusters (edge and interveinal pixels), which were far from bacterial spots, green healthy leaf areas, and green areas on symptomatic leaves. Bacterial spots formed another two adjacent clusters (edge and interveinal pixels), which were far from other classes in the VISNIR plot, but close to green healthy leaf edge pixels in the SWIR plot. Pixels from GH and GS were inseparable in the SWIR LDA plot. However, GH-e and GS-e formed two adjacent clusters while GH-iv and GS-iv formed two adjacent clusters, which were clearly separated from other groups in the VISNIR LDA plot.
The confusion matrix of the best classification models showed that most confusion occurred between the edge and interveinal classes within the same type of leaf samples (Figure 7 C-D). With VISNIR data (Figure 7C), it was possible to differentiate GH completely from BS-e/iv, AS-e/iv, and GS-iv. BS were 100% differentiated from all other groups except that there was 22% confusion between AS-e and BS-e. GS were 100% differentiated from GH and BS groups, but there was less than 20% confusion between GS and AS-iv. The classification of AS pixels was less successful, there was 22% confusion between AS-e and GH-e, 11% between AS-e and BS-e, and 25% between AS-iv and BS-iv. Background pixels were separated from any other groups. With SWIR data (Figure 7D), GH-iv, BS-e, AS-e, and GS-iv were 100% differentiated from all other leaf types. The classification of GH-iv and BS-iv was not as successful, due to multiple cross-group confusion. The gini importance results from RF models suggested that wavelength bands in 550-600 nm, 690 nm, and 1400-1450 nm were important in classifying the above-mentioned leaf pixels (Figure S2). Strong positive correlation (Table S5) was found between VISNIR and SWIR model (trained with leaf-level full spectra data) accuracy (r=0.93, P=0.07), between VISNIR accuracy and bacterial population (r =0.96, P =0.04), as well as between SWIR accuracy and bacterial population (r =0.99, P =0.01).
Prediction from Leaf and Living Plant Images
The prediction results suggested that VISNIR-RF was relatively accurate in detecting BLS on detached diseased leaves, although some pixels on uninoculated green healthy leaves were misclassified as “green areas on symptomatic leaves” (Figure 8). The classification of abiotic spots was generally accurate (Figure 8). The prediction of SWIR pixels on detached leaves was less successful as the majority of bacterial spot interveinal pixels were misclassified as abiotic spots (Figure 8). The whole-plant-image prediction was able to capture the presence of abiotic spots on healthy tomato plants with both VISNIR (Figure S3A) and SWIR models (Figure S3B) and also classified some bacterial spots correctly (Figure S3C-D). However, the incidence of false positives was high as a large numbers of healthy leaf pixels and a small portion of background pixels were misclassified as bacterial spots (Figure S3A-D).
Due to its potential application in high-throughput crop disease detection, hyperspectral remote sensing has received a lot of attention. In particular, the possibility of applying HSI to perform early disease detection, i.e., detecting crop diseases before any visible symptoms are present has been explored. However, many early published results/data were acquired with hand-held point spectrometers that only provided limited spatial information on plants, or multi-spectral imagers that only provided limited spectral information. In addition to that, oftentimes, there was a lack of validated data representing the responses of plants at different disease progress stages. The advancement in HSI instrument development and machine learning techniques made it possible for us to collect and analyze data at a different level. We utilized leaf/plant-level hyperspectral images and established pathogen population data collected from 7 consecutive days to analyze the differences between plants with BLS disease and healthy plants. The results suggested that HSI is capable of detecting BLS on tomato leaves at pre-symptomatic stages and differentiating BLS spots from biotic spots. Through machine learning image analysis, important wavelength bands and VIs were revealed. We found that wavelength bands and VIs that were important in classification (differentiating between infected and uninfected leaves) varied among disease progression stages. What’s interesting is that, during the 7-day disease progression, the change in which wavelength bands were important was consistent with that of VI patterns, which in turn was highly correlated with the growth of bacterial populations in tomato leaves and in line with the physiological changes occurring in tomato leaves.
In this study, both uninoculated control and inoculated groups (tomato plants) were treated in exactly the same way before inoculation. Neither VISNIR nor SWIR full spectra showed any difference between the control and inoculated group. However, both VISNIR and SWIR data detected clear differences between the control and inoculated tomato leaves merely 2 hours after inoculation. Wavelength bands around 740-750 nm and 1400 nm weighed more in differentiating these two groups at this stage. Previous researchers reported that 750 nm (far-red) and nearby bands in far-red and near-infrared regions are associated with disease activities in various crops including tomato28,41,42. Previous research suggested that far-red is associated with tomato defense hormone-mediated responses (Courbier et al. 2020), which have a big impact on tomato morphology, development, photosynthetic performance, and fruit production (Tan et al. 2022; Kalaitzoglou et al. 2019). The other important band, 1400 nm, is known to be associated with water and carbon dioxide absorption43,44. It is known that Xanthomonas pathogens often cause water soaking on tomato leaves by drawing water into the apoplast (the intercellular space in plant tissues) from the leaf surface45,46, and that the water-soaking effect allows other bacteria to enter plants within 30 minutes. Considering that bacterial populations in tomato leaves already reached 105 CFU/cm2 leaf area at 2hr ai, it is highly likely that the changes at 1400 nm that we observed at 2hr ai were caused by X. perforans infection. During the pre-symptomatic stage (1-3dai), VISNIR data continued to support the differentiation between healthy and infected plants with the most important bands around 743 nm (far-red). SWIR data were less supportive of such a differentiation, possibly because at this stage the pathogen was already established in the leaf tissues and no longer needed to draw a large amount of water from the leaf surface.
The onset of disease symptoms occurred on inoculated tomato plants at 4 dai. At this point, RF, GB, and XGB models outperformed the LDA model using VISNIR data, yet uninoculated and inoculated groups formed two adjacent clusters on the LDA plot. The separation between these two clusters was less clear on the SWIR LDA plot. However, both VISNIR and SWIR data showed distinct differences between inoculated and inoculated from 5 dai to 7dai. Compared to the pre-symptomatic stages, important wavelength bands shifted towards the near-infrared region (750-1200 nm) during the symptom development (mid, 4 -5 dai) stage. Since at this stage X. perforans had already started to damage leaf structures and affected photosynthesis, it is not surprising to see a hyperspectral pattern different from those of the pre-symptomatic stages. At the mid stage, multiple bands including the ones around 765nm, 755 nm, 850 nm, 920-930 nm, 1057 nm, 1080-1110 nm, and 1200 nm had big impacts on the classification. Previous studies showed that reflectance data at those bands are correlated with tomato defense hormone-mediated responses47, leaf pigment contents and leaf structures48. At the late, i.e., the advanced BLS disease progression stage, bands between 600nm and 640 nm and bands around 1404 nm were of great importance in classification. That could be due to water loss and reduced leaf pigment resulting from the severe necrosis of infected leaves.
VI pattern changes across BLS disease progression stages were consistent with the hyperspectral band pattern changes we observed with full spectra analysis. Previous publications reported that the importance of individual VIs changed as plant disease progressed (Pane et al. 2022; Lu et al. 2018). In this study, in addition to identifying the most informative individual VIs, we treated the entire selection of VIs as supergroup and observed its pattern at different stages during disease progression. The VIs we selected can be categorized into four groups focusing on leaf pigment and photosynthesis, stress response, leaf structure, and leaf water content. This way, we were able to compare the physiological changes of inoculated leaves and the change of VI patterns. This study provided strong evidence supporting our hypothesis that early (pre-symptomatic) BLS detection relies more on tomato leaf water content and disease hormone-mediated plant response monitoring rather than leaf pigment or leaf structure observations. As disease progresses to more advanced stages, the importance of leaf pigment and structure monitoring increases. Therefore, we suggest that future studies should include various groups of VIs instead of focusing on individual VIs ranking on the top in the feature importance table. Based on the dynamic change of VI patterns during disease progression observed in this study, we hypothesize that VI patterns can be engineered with time series data to create new features that might be more informative in disease detection.
In this study, we investigated the influence of potential confounding factors including scan precision, background materials, leaf size, and leaf structure. The results showed background materials and leaf structure have great impacts on tomato leaf reflectance data. In fact, through analyzing the confusion matrices of all the classification models, we found abiotic spots, bacterial spots, healthy tissues, and healthy-looking tissues on symptomatic leaves were generally separable, however, edge pixels tend to cause more cross-group confusion than interveinal pixels. Therefore, we recommend segmenting different anatomical parts of leaves before analyzing the impact of biotic and abiotic factors on leaves.
We also compared average accuracy of classification models. The best full-spectra classification models generated in this study were able to detect pre-symptomatic BLS infection with an averaged accuracy of 0.63. The mean accuracy increased up to 0.74 at symptomatic stages. With the pixel-level full spectra data collected at 7 dai, most model accuracy scores exceeded 0.80, with an average around 0.60. We note that those models were trained to differentiate multiple classes, and the confusion matrices suggested most confusions came from within-group misclassification. In fact, by reducing the number of classes in the classification models, we were able to achieve much higher accuracy and F1 scores with the models trained for the “beginning” stage classification. Most VI classification models had nearly 100% accuracy across the four BLS disease progression stages. This is possibly due to the fact that the data set size was small compared to our pixel data set size. Another possible reason is that our VI selection included most of the important bands identified from the full-spectra classification models. VI data may have contained more relevant features compared to the full spectra data, which resulted in better classification performance than full-spectra models, suggesting the importance of feature selection.
The prediction results obtained from detached leaf images were relatively accurate, although some pixels of uninoculated green healthy leaves were misclassified as green areas on symptomatic leaves by the VISNIR model, and some bacterial spot interveinal and other pixels were misclassified as abiotic spots by the SWIR models. Abiotic spots, bacterial spots, and most healthy leaves were classified correctly. Whole-plant-image classification had relatively poor performance, although it was able to capture the presence of abiotic spots on healthy tomato plants and to detect bacterial spots on infected plants. Applying our models on living-plant images was difficult, as our training data were acquired from detached leaves laying on a flat surface. The fact that different leaves on living plants have different leaf angles and distances to the camera can heavily reduce the accuracy of our classification models. To tackle this problem, a large amount of validated disease-related HSI data with depth information (RGBD; 3D reconstruction) are needed, and deep learning approaches domain adaptation transfer learning could be considered 49,50.
In conclusion, the results from this study demonstrate the possibility of using HSI and machine learning to detect BLS at pre-symptomatic stages. Using both leaf-level and pixel-level data, with full spectra and VI analyses, we were able to match the physiological leaf changes and pathogen growth during disease progression with the HSI patterns we observed. Our findings about the significance of VI group patterns and leaf structure segmentation can help design more effective data collection and analysis methods. Overall, the outcome may benefit future HSI-related crop disease research studies and applications.
Plant and Inoculum Preparation
Tomato (Solanum lycopersicum) seeds (c.v. Early Girl, Mountain Valley See Co.) were seeded in a 48-cell tray and grown under greenhouse conditions between 25-38°C for three weeks. The tomato seedlings germinated seven to ten days after seeding and they were fertilized (Osmocote Smart-Release, ICL) one time two weeks after seeding. Plants were irrigated for two minutes twice a day at eight o’clock in the morning and five o’clock in the afternoon. Three-week-old tomato plants were used for the inoculation experiments. The plants were sprayed with sterilized water and bagged for 24 hours before inoculation and another 24 hours after inoculation. All institutional guidelines at Virginia Tech for handling tomato plants were followed. The tomato plants used in this research were obtained from a licensed seed company and do not contain transgenic materials. For non-transgenic plants, no additional permission is required for handling these tomato plants according to Virginia Tech biosafety guidelines.
Inoculum Preparation
A Xanthomonas perforans strain isolated previously from a tomato field in Virginia was grown in a shaker incubator overnight in R2 broth (Teknova) at 28°C and 200 rpm. Two-hundred-micro-liter aliquots of X. perforans. liquid cultures were spread on R2A medium amended with 20 mg/L of the antibiotic rifampicin (Sigma-Aldrich). The survivors grown on 20 mg/L rifampicin-R2A medium were transferred to 50 mg/L rifampicin-R2 agar medium, and this process was repeated on 100, 150, and 200 mg/L rifampicin-R2 agar medium. The survivors grown on 200 mg/L rifampicin-R2A medium were cultured on 200 mg/L rifampicin-R2A agar medium again to confirm their resistance to rifampicin. Six final survivors were selected for pathogenicity tests on three-week-old healthy tomato plants, and the one that showed the highest virulence was used for the inoculation experiments. The inoculum consisted in bacteria resuspended in 10 mM MgSO4 at a concentration of 1-5 x 10^7 CFU/ml.
Inoculation and Experimental Design
Tomato plants were inoculated by dipping the plants (including stems and leaves) into a sterilized plastic cup filled with 300 ml X. perforans inoculum mixed with 0.02% Silwet, for a minute. The control plants were dipped in a sterilized plastic cup filled with 300 ml 10 mM MgSO4 mixed with 0.02% Silwet for a minute (Hert et al. 2009). The inoculation experiments were performed twice in this study. In the first experiment, there were eight experimental units (4-cell inserts) and each contained two tomato plants. Units 1, 2, 5 and 6 contained uninoculated control plants, while units 3, 4, 7, and 8 contained plants inoculated with X. perforans. The second experiment employed a similar design, except that there were 10 experimental units, with units 1-5 uninoculated and units 6-10 inoculated.
Data Collection
Data were collected nine times in each experiment: before inoculation (bi), 2 hours after inoculation (0 days after inoculation, dai), every 24 hours from 1 to 7 dai. Three types of data were collected, including disease rating, bacterial population size measurements, and hyperspectral image data (Figure 1). Disease-related data were collected based on the visual observation of tomato plants in each experimental unit. Then one to six leaves were removed from each experimental unit. After rating the disease severity of detached leaves, we used a hyperspectral benchtop imaging system (Resonon PIKA-L and PIKA-IR) to capture the hyperspectral images of each detached leaf. Then the leaves were processed to estimate the bacterial population size in the leaf tissue.
Disease Rating
Two types of disease-related data were collected in this study to characterize the disease progress of BLS. First, we collected disease incidence data by estimating the percentage of leaves showing bacterial leaf spot in each experimental unit. Then we randomly selected five tomato leaves and rated the disease severity on each leaf (Chester 1950). We used a 0-5 scale to rate the severity data, where 0 indicates healthy, 1 indicates spot coverage between 0 and 10%, 2 indicates spot coverage between 10% and 30 %, 3 indicates spot coverage between 30% and 50%, without apparent dead tissues, 4 indicates spot coverage of more than 50% and some of the diseased tissue is dark and curled or decayed, and 5 indicates completely dry and curled or decayed tissue with numerous bacterial spots. Then a disease severity index was calculated as:
where NRi indicates the number of seedlings showing the corresponding disease level i; i ranges from 0 to 5 (Chester 1950). After rating the disease incidence and severity in each experimental unit, one to six leaves were removed from each experimental unit, and disease severity was rated for each detached leaf with the same 0-5 scale.
Hyperspectral Image Acquisition
Hyperspectral images of the detached leaves were collected using a RESONON benchtop system with Pika L (visible & near infrared: 400-1000 nm) and Pika NIR320 (shortwave infrared: 900-1700 nm). All leaf samples were scanned with both Pika L and Pika NIR320 cameras to collect hyperspectral image data ranging from 400 to 1700 nm. Before scanning, a camera (Pika L or Pika NIR320) was mounted on a tower, and the leaf samples were laid on a stage below the camera and four high-intensity halogen spotlights (Figure S4). The system was set to reflectance mode, where the system measured the absolute reflectance of the leaf samples by applying dark correction, and response correction using a reference (calibration) tile provide by RESONON. The environmental lights were turned off, with only the imaging halogen lights on. A piece of black cotton fabric or a white calibration tile was placed on the stage, and the leaf samples were laid on the top of the background materials. During data acquisition, the stage moved in auto speed mode, and the camera scanned multiple lines simultaneously and translated into three-dimensional image cubes containing two-dimensional spatial and corresponding spectral data. Each scan yielded a bil (Band Interleaved by Line) data cube and a header file, which contained either 561 spatial pixel information (5.86 mm/pixel) at 300 VIS-NIR (385.63-1027.18 nm) wavelength bands or 164 spatial pixel information (30mm/pixel) at 168 NIR (890.68-1719.81 nm) wavelength bands.
Bacteria Growth Curve
After the hyperspectral data from the detached leaves were acquired, each leaf was processed to estimate the bacterial pathogen population size (CFU per cm2 of leaf area). To begin with, five (1 cm diameter) round leaf discs were randomly subsampled from a leaf sample. The subsamples were sterilized with 75% ethanol for 45s, followed by a triple rinse with sterilized distilled water (SDW), and then sterilized with 10% bleach (0.5 % sodium hypochlorite) solution for 30s, followed by triple rinse with SDW. Five surface-sterilized leaf discs were placed in a 1.5 ml Eppendorf tube and crushed by a disposable pellet pestle (Fisher brand) with 500 mL SDW. The leaf sap was collected as the fine leaf tissues were evenly suspended in the SDW, and it went through serial dilutions to obtain 10-1–10-8 of the original concentration. Two 20 mL aliquots of each diluted leaf sap were evenly spread on 200 mg/L rifampicin-R2A medium. The cultures were stored at 28 °C for two days and the bacterial colonies grown from each leaf sap concentration were counted and averaged. Finally, the original concentration of each leaf sap sample was calculated and the bacterial pathogen population size per cm2 of leaf area was estimated for each detached leaf sample.
Potentially Confounding Factors
In order to investigate the influence of potential confounding factors (including background materials and colors, scan precision, leaf size, and leaf structure) on leaf reflectance, additional leaves and corresponding hyperspectral images were collected.
The image data of 18 tomato leaves collected using dark fabric as background were used to test scan precision. The sample leaves included 8 leaves (4 uninoculated and 4 inoculated) from the first inoculation experiment and 10 leaves (5 uninoculated and 5 inoculated) from the second experiment. To test this hypothesis, each leaf was scanned twice at different position on the same stage and background. Leaf-level full spectra reflectance data of each leaf was compared between two scans, along with those of 18 randomly selected background pixels from each scan, using principal component analysis (PCA) and linear discriminant analysis (LDA) with scikit-learn in Python 64.
All leaf samples were collected with two background materials and colors: a piece of black cotton fabric, and a white calibration tile, placed on the stage. Therefore, each leaf yielded two sets of data: VISNIR (400-1000 nm) and SWIR (900-1700 nm) images with black cloth, and VISNIR and SWIR images with white tile. The image data of 16 tomato leaves were used to test this hypothesis, which included eight uninoculated and eight inoculated leaves collected at the end of each experiment. Leaf-level, spatially averaged full spectra reflectance data of each tomato leaf were compared across background materials, along with those of 16 randomly selected background pixels, using PCA and LDA.
The image data of 18 uninoculated tomato leaves collected with dark fabric background were used to test the leaf size hypothesis. Two leaves were collected from each of nine uninoculated plants, with a length of over 2.0 inches (big) or less than 1.5 inches (small). Therefore, the leaves collected from those nine uninoculated plants were labeled as “big” or “small” and plant group ID. Leaf-level full spectra data of each leaf were compared between and among nine groups, along with those of 18 randomly selected background pixels, using PCA and LDA.
The image data of nine uninoculated tomato leaves collected with dark fabric background were used to test this hypothesis. Two leaves were collected from each of nine uninoculated plants. Five pixels were randomly selected from five structure areas from each leaf, including apex, margin, midrib, veins, and interveinal leaf tissues. The full spectra data of each pixel were compared between leaf structures, along with those of 45 randomly selected background pixels, using PCA and LDA.
Hyperspectral Image Analysis
In order to differentiate between infected tomato leaves and uninoculated healthy leaves and observe the physiological changes on tomato leaves during disease progression, HIS data collected at nine time points were analyzed with machine learning methods (Figure 1). HSI data were analyzed at three different levels, including VI (Vegetation Index), pixel, and whole image levels. Eight algorithms including LDA, Supportive Vector Machines (SVMs), K-nearest neighbors (KNN), Random Forest (RF), Gradient Boosting Machines (GBM), Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGB or XG Boost) were used with scikit-learn 64 and xgboost 65 in Python to perform feature selection and classify tomato leaf samples based on HSI data. During image data analysis, leaf samples from two classes (uninoculated healthy and infected) from nine data collection time points were used for ML training and testing, with 10-fold cross-validation. Then accuracy and F1 scores were compared to evaluate the performance of each model.
HSI images were cropped with Spectronon software (Spectronon Pro, Resonon, Bozeman, MT) to retain the leaf and minimum background pixels. The leaves were labeled by treatment, uninoculated or inoculated, and by data collection time points. The time points were further categorized into four stages: beginning (bi & 2hr ai), early (pre-symptomatic: 1-3 dai), mid (4-5 dai), and late (6-7 dai) stages. Thirty-six images containing 228 leaves were used for leaf-level analysis. Each leaf generated a mean spectral data set that averaged the pixel reflectance intensity at every wavelength band. There were 36 VISNIR and 36 SWIR leaf-level data from ‘beginning’, 54 VISNIR and 54 SWIR leaf-level data from ‘early’, 46 VISNIR and 46 SWIR leaf-level data from ‘mid’, and 92 VISNIR and 92 SWIR leaf-level data from ‘late’. There were four classes included in the ‘beginning’ dataset: uninoculated group bi, inoculated group bi, uninoculated 2hr ai, and inoculated group 2hr ai. There were six classes in the ‘early’ dataset, including uninoculated 1 dai, inoculated 1 dai, uninoculated 2 dai, inoculated 2 dai, uninoculated 3 dai, and inoculated 3 dai. There were four classes in the ‘mid’ dataset, including uninoculated 4 dai, inoculated 4 dai, uninoculated 5 dai, and inoculated 5 dai. There were four classes in the ‘late’ dataset, including uninoculated 6 dai, inoculated 6 dai, uninoculated 7 dai, and inoculated 7 dai.
As for pixel-level analysis, nine classes of 72 pixels (648 VISNIR pixels and 648 SWIR) were randomly selected from 27 cropped image files containing 72 leaves collected at the end (7dai) of each experiment, which included class 0: green areas on the edge of uninoculated healthy leaves (GH-e), class 1: green areas in the interveinal areas on uninoculated healthy leaves (GH-iv), class 2: bacterial spots on the edge of infected leaves (BS-e), class 3: bacterial spots in the interveinal areas on infected leaves (BS-iv), class 4: abiotic spots on the edge of uninfected leaves (AS-e), class 5: abiotic spots in the interveinal areas on uninfected leaves (AS-iv), class 6: green areas on the edge of symptomatic tomato leaves (GS-e), class 7: green areas in the interveinal areas on symptomatic tomato leaves (GS-iv),and class 8: background (bg). The presence of abiotic spots and bacterial spots was verified by visual observation and culture isolation. Pixels were selected manually to ensure the same amount of leaf margin pixels and interveinal pixels were selected from each leaf, with the number ranging from four to six. The spectra data were saved in txt files using Spectronon Pro, and the raw data files were converted to csv files before data analysis.
For leaf-level full spectra analysis, the first and last wavelength bands were removed to exclude potential noises. Therefore, VISNIR spectra data files contained 298 bands (387.65 nm to 1024.90 nm, with an approximate 2nm interval) and SWIR spectra data files contained 166 bands (895.52 nm to 1714.71 nm, with an approximate 5nm interval). Data were normalized to fit a 0-1 scale. After data preprocessing, each pixel in each class contained 298 bands in VISNIR files and 166 bands in its SWIR files. LDA, SVM, KNN, RF, GBM, XGBoost, and MLP were employed to train the classification models. Hyperparameters (Figure S5) for LDA, SVM, KNN, RF, and MLP were tuned with Grid Search, and those for GBM and XGBoost were tuned with Randomized Search 64. The best combinations of hyperparameters were retained to train the classification model, with stratified 3-fold cross-validation repeated three times. The choice of 3-fold cross validation was made because of the small sample size. The accuracy and weighted F1 scores of each model were compared, and the best model was applied to predict leaf health on whole hyperspectral images. Important wavelength bands were extracted based on gini importance (mean decrease impurity) from corresponding RF, GB, or XGB models, whichever showed the best performance.
In leaf-level VI analyses, 14 vegetation indices (VIs, Table S4) that might be related to tomato disease stress were extracted from the pre-processed HSI data 24,30,31,36,56. The above-mentioned classifiers were employed to train the classification models, and the performance was evaluated as described above. As for pixel-level full spectra analysis, the process was similar, except that 10-fold cross-validation was used instead of 3-fold, as the pixel data set size was bigger. The accuracy of classification models was compared using ANOVA with statsmodels 51 followed by post-hoc analysis with scikit-posthocs 52. The testing accuracy scores generated from classification models trained with leaf-level full spectra data were correlated with bacterial population in tomato leaf tissues averaged across two experiments during BLS disease progress. Pearson correlation coefficient and p-value were calculated with scipy.stats 63.
The best classification models were applied to healthy and diseased whole leaf and living plant images collected at 7dai. The RF model trained with 7dai VISNIR data and the SVM model trained with the SWIR model were applied to whole leaf images and living plants. The raw hyperspectral images were processed, and the reflectance data were extracted using spectral python (SPy) in Python 66. The predicted results were compared with visual observations and bacterial population sizes (the ground truth data).
Acknowledgments
We thank Jeff Burr (Virginia Tech) for providing greenhouse tech support for this study. This work was supported by U.S. Department of Agriculture (2021-67021-34037).
Data Availability
The raw hyperspectral image data have been uploaded to Ag Data Commons. https://data.nal.usda.gov/dataset/early-detection-bacterial-spot-disease-tomato-hyperspectral-imaging.
- Abdulridha, Jaafar, Yiannis Ampatzidis, Sri Charan Kakarla, and Pamela Roberts. 2020. "Detection of target spot and bacterial spot diseases in tomato using UAV-based and benchtop-based hyperspectral imaging techniques." Precision Agriculture21: 955-978.
- Buja, I. et al. Advances in plant disease detection and monitoring: From traditional assays to in-field diagnostics. Sensors21, 2129 (2021).
- Ristaino, J. B. et al. The persistent threat of emerging plant disease pandemics to global food security. Proceedings of the National Academy of Sciences118, e2022239118 (2021).
- Burdon, J. J. & Zhan, J. Climate change and disease in plant communities. PLoS biology18, e3000949 (2020).
- Gergerich, R. C. et al. Safeguarding fruit crops in the age of agricultural globalization. Plant disease99, 176-187 (2015).
- Liebhold, A. M., Brockerhoff, E. G., Garrett, L. J., Parke, J. L. & Britton, K. O. Live plant imports: the major pathway for forest insect and pathogen invasions of the US. Frontiers in Ecology and the Environment10, 135-143 (2012).
- Waage, J. & Mumford, J. D. Agricultural biosecurity. Philosophical Transactions of the Royal Society B: Biological Sciences363, 863-876 (2008).
- Savary, S., Ficke, A., Aubertot, J.-N. & Hollier, C. Vol. 4 519-537 (Springer, 2012).
- Martinelli, F. et al. Advanced methods of plant disease detection. A review. Agronomy for Sustainable Development35, 1-25 (2015).
- Riley, M. B., Williamson, M. R. & Maloy, O. Plant disease diagnosis. The Plant Health Instructor10 (2002).
- Dyussembayev, K. et al. Biosensor technologies for early detection and quantification of plant pathogens. Frontiers in Chemistry, 144 (2021).
- Kumar, R., Pathak, S., Prakash, N., Priya, U. & Ghatak, A. in Diagnostics of Plant Diseases (IntechOpen, 2021).
- Oerke, E.-C. Remote sensing of diseases. Annual review of phytopathology58, 225-252 (2020).
- Mahlein, A.-K. Plant disease detection by imaging sensors–parallels and specific demands for precision agriculture and plant phenotyping. Plant disease100, 241-251 (2016).
- Bohnenkamp, D., Behmann, J., Paulus, S., Steiner, U. & Mahlein, A.-K. A hyperspectral library of foliar diseases of wheat. Phytopathology®111, 1583-1593 (2021).
- Velásquez, D. et al. A Cyber-Physical Data Collection System Integrating Remote Sensing and Wireless Sensor Networks for Coffee Leaf Rust Diagnosis. Sensors21, 5474 (2021).
- Van De Vijver, R. et al. In-field detection of Alternaria solani in potato crops using hyperspectral imaging. Computers and Electronics in Agriculture168, 105106 (2020).
- Bhupathi, P. & Sevugan, P. Application of hyperspectral remote sensing technology for plant disease forecasting: An applied review. Annals of the Romanian Society for Cell Biology25, 4555-4566 (2021).
- Zilberman, A., Asher, J. B., Kopeika, N. S. & Reshef, Y. in Remote Sensing for Agriculture, Ecosystems, and Hydrology XXI. 484-492 (SPIE).
- Teke, M., Deveci, H. S., Haliloğlu, O., Gürbüz, S. Z. & Sakarya, U. in 2013 6th international conference on recent advances in space technologies (RAST). 171-176 (IEEE).
- Mahlein, A.-K., Oerke, E.-C., Steiner, U. & Dehne, H.-W. Recent advances in sensing plant diseases for precision crop protection. European Journal of Plant Pathology133, 197-209 (2012).
- Bock, C. H., Barbedo, J. G., Del Ponte, E. M., Bohnenkamp, D. & Mahlein, A.-K. From visual estimates to fully automated sensor-based measurements of plant disease severity: status and challenges for improving accuracy. Phytopathology Research2, 1-30 (2020).
- Lu, B., Dao, P. D., Liu, J., He, Y. & Shang, J. Recent advances of hyperspectral imaging technology and applications in agriculture. Remote Sensing12, 2659 (2020).
- Cubero, S., Marco-Noales, E., Aleixos, N., Barbé, S. & Blasco, J. Robhortic: A field robot to detect pests and diseases in horticultural crops by proximal sensing. Agriculture10, 276 (2020).
- Nguyen, C. et al. Early detection of plant viral disease using hyperspectral imaging and deep learning. Sensors21, 742 (2021).
- Bebronne, R. et al. In-field proximal sensing of septoria tritici blotch, stripe rust and brown rust in winter wheat by means of reflectance and textural features from multispectral imagery. Biosystems Engineering197, 257-269 (2020).
- Gold, K. M. et al. Hyperspectral measurements enable pre-symptomatic detection and differentiation of contrasting physiological effects of late blight and early blight in potato. Remote Sensing12, 286 (2020).
- Wei, X., Johnson, M. A., Langston Jr, D. B., Mehl, H. L. & Li, S. Identifying optimal wavelengths as disease signatures using hyperspectral sensor and machine learning. Remote Sensing13, 2833 (2021).
- Yeh, Y.-H. et al. Strawberry foliar anthracnose assessment by hyperspectral imaging. Computers and Electronics in Agriculture122, 1-9 (2016).
- Gold, K. M., Townsend, P. A., Larson, E. R., Herrmann, I. & Gevens, A. J. Contact reflectance spectroscopy for rapid, accurate, and nondestructive Phytophthora infestans clonal lineage discrimination. Phytopathology110, 851-862 (2020).
- Veys, C. et al. Multispectral imaging for presymptomatic analysis of light leaf spot in oilseed rape. Plant methods15, 1-12 (2019).
- Pane, C., Manganiello, G., Nicastro, N. & Carotenuto, F. Early detection of wild rocket tracheofusariosis using hyperspectral image-based machine learning. Remote Sensing14, 84 (2021).
- Abrahamian, P., Klein-Gordon, J. M., Jones, J. B. & Vallad, G. E. Epidemiology, diversity, and management of bacterial spot of tomato caused by Xanthomonas perforans. Applied microbiology and biotechnology105, 6143-6158 (2021).
- Osdaghi, E., Taghavi, S. M., Hamzehzarghani, H., Fazliarab, A. & Lamichhane, J. R. Monitoring the occurrence of tomato bacterial spot and range of the causal agent Xanthomonas perforans in Iran. Plant Pathology66, 990-1002 (2017).
- Strayer-Scherer, A. et al. Advanced copper composites against copper-tolerant Xanthomonas perforans and tomato bacterial spot. Phytopathology108, 196-205 (2018).
- Fahrentrapp, J., Ria, F., Geilhausen, M. & Panassiti, B. Detection of gray mold leaf infections prior to visual symptom appearance using a five-band multispectral sensor. Frontiers in plant science10, 628 (2019).
- Lu, J., Ehsani, R., Shi, Y., de Castro, A. I. & Wang, S. Detection of multi-tomato leaf diseases (late blight, target and bacterial spots) in different stages by using a spectral-based sensor. Scientific reports8, 1-11 (2018).
- Xie, C., Shao, Y., Li, X. & He, Y. Detection of early blight and late blight diseases on tomato leaves using hyperspectral imaging. Scientific reports5, 1-11 (2015).
- Wang, D. et al. Early detection of tomato spotted wilt virus by hyperspectral imaging and outlier removal auxiliary classifier generative adversarial nets (OR-AC-GAN). Scientific reports9, 1-14 (2019).
- Jones, C., Jones, J. & Lee, W. Diagnosis of bacterial spot of tomato using spectral signatures. Computers and Electronics in Agriculture74, 329-335 (2010).
- Abrahamian, P., Sharma, A., Jones, J. B. & Vallad, G. E. Dynamics and spread of bacterial spot epidemics in tomato transplants grown for field production. Plant Disease105, 566-575 (2021).
- Zhang, M. & Qin, Z. in IGARSS 2004. 2004 IEEE International Geoscience and Remote Sensing Symposium. 4091-4094 (IEEE).
- Hahn, F. Actual pathogen detection: sensors and algorithms-A review. Algorithms2, 301-338 (2009).
- Prieto-Blanco, A., North, P. R., Fox, N. & Barnsley, M. J. in Global Developments in Environmental Earth Orbservation from Space. Proceedings of the 25th EARSeL Symposium. Porto: Millpress Science Publishers. 137-144.
- Marble, C. B. et al. Z-scan measurements of water from 1150 to 1400 nm. Optics Letters43, 4196-4199 (2018).
- Schwartz, A. R., Morbitzer, R., Lahaye, T. & Staskawicz, B. J. TALE-induced bHLH transcription factors that activate a pectate lyase contribute to water soaking in bacterial spot of tomato. Proceedings of the National Academy of Sciences114, E897-E903 (2017).
- Schornack, S., Minsavage, G. V., Stall, R. E., Jones, J. B. & Lahaye, T. Characterization of AvrHah1, a novel AvrBs3‐like effector from Xanthomonas gardneri with virulence and avirulence activity. New Phytologist179, 546-556 (2008).
- Courbier, S. et al. Far-red light promotes Botrytis cinerea disease development in tomato leaves via jasmonate-dependent modulation of soluble sugars. Plant Cell Environ43, 2769-2781 (2020). https://doi.org:10.1111/pce.13870
- Liu, L. Y., Huang, W. J., Pu, R. L. & Wang, J. H. Detection of Internal Leaf Structure Deterioration Using a New Spectral Ratio Index in the Near-Infrared Shoulder Region. J Integr Agr13, 760-769 (2014). https://doi.org:10.1016/S2095-3119(13)60385-8
- Yan, K., Guo, X., Ji, Z. & Zhou, X. Deep Transfer Learning for Cross-Species Plant Disease Diagnosis Adapting Mixed Subdomains. IEEE/ACM Transactions on Computational Biology and Bioinformatics (2021).
- Kouw, W. M. & Loog, M. An introduction to domain adaptation and transfer learning. arXiv preprint arXiv:1812.11806 (2018).
- Seabold, S. & Perktold, J. in Proceedings of the 9th Python in Science Conference. 10.25080 (Austin, TX).
- Terpilowski, M. A. scikit-posthocs: Pairwise multiple comparison tests in Python. Journal of Open Source Software4, 1169 (2019).
- Huang, S., Tang, L., Hupy, J. P., Wang, Y. & Shao, G. A commentary review on the use of normalized difference vegetation index (NDVI) in the era of popular remote sensing. Journal of Forestry Research32, 1-6 (2021).
- Evangelides, C. & Nobajas, A. Red-Edge Normalised Difference Vegetation Index (NDVI705) from Sentinel-2 imagery to assess post-fire regeneration. Remote Sensing Applications: Society and Environment17, 100283 (2020).
- Garbulsky, M. F., Peñuelas, J., Gamon, J., Inoue, Y. & Filella, I. The photochemical reflectance index (PRI) and the remote sensing of leaf, canopy and ecosystem radiation use efficiencies: A review and meta-analysis. Remote sensing of environment115, 281-297 (2011).
- Kureel, N., Sarup, J., Matin, S., Goswami, S. & Kureel, K. Modelling vegetation health and stress using hypersepctral remote sensing data. Modeling Earth Systems and Environment8, 733-748 (2022).
- Merzlyak, M. N., Gitelson, A. A., Chivkunova, O. B. & Rakitin, V. Y. Non‐destructive optical detection of pigment changes during leaf senescence and fruit ripening. Physiologia plantarum106, 135-141 (1999).
- Blackburn, G. A. Spectral indices for estimating photosynthetic pigment concentrations: a test using senescent tree leaves. International Journal of remote sensing19, 657-675 (1998).
- Claudio, H. C. et al. Monitoring drought effects on vegetation water content and fluxes in chaparral with the 970 nm water band index. Remote Sensing of Environment103, 304-311 (2006).
- de Jong, S. M., Addink, E. A., Hoogenboom, P. & Nijland, W. The spectral response of Buxus sempervirens to different types of environmental stress–A laboratory experiment. ISPRS Journal of Photogrammetry and Remote Sensing74, 56-65 (2012).
- Penuelas, J., Baret, F. & Filella, I. Semi-empirical indices to assess carotenoids/chlorophyll a ratio from leaf spectral reflectance. Photosynthetica31, 221-230 (1995).
- Jin, X. et al. Estimation of leaf water content in winter wheat using grey relational analysis–partial least squares modeling with hyperspectral data. Agronomy Journal105, 1385-1392 (2013).
- Virtanen, P. et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nature methods17, 261-272 (2020).
- Pedregosa, F. et al. Scikit-learn: Machine learning in Python. the Journal of machine Learning research12, 2825-2830 (2011).
- Chen, T. & Guestrin, C. in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 785-794.
- Boggs, T. in http://www.spectralpython.net/ (2022).
No competing interests reported.