Using multispectral imaging to monitor the physiological condition of crops
The recent technological advances of drones and long-range spectroscopy help us to monitor the physical and physiological conditions of the plants even in their natural habitat or agricultural fields. In this article we demonstrate the capabilities of MicaSense RedEdge-MX through three case studies.
Plants, according to their sessile lifestyle, are exposed to the biotic (pathogens, herbivores) and abiotic (drought, chilling, nutrient deficiency…) aspects of their environment. The extreme and lasting changes of environmental factors cause severe damages to the plants. By definition plant stress is a physiological condition, where the growth, development and reproduction remain out of the optimal acclimation range due to the increased environmental strain, thus plants are not able to exploit their whole genetic potential. Long-lasting stress causes the decrease of fitness or the ability for reproduction. This phenomenon is a multilevel syndrome with well-defined phases and with many physical and physiological manifestations (decreased chlorophyll content, loss of water potential, decreased biomass, yield loss etc.).
Therefore, early recognition of plant stress in the fields is of crucial importance for the elaboration of fast and accurate plant protection strategies.
Using special multispectral sensors, it is now possible to measure relative chlorophyll content, nitrogen status or identify infection focal points. However, it is important to take into consideration that precise analytics beyond a chlorophyll content measurement (for instance nutrient status or pathogen identification) are affected by climatic and geographic factors.
At the AGRON we think that the operation model of a drone service should help the customer from the beginning to the real action on the fields. Our monitoring drone program helps to learn the proper drone controlling and data collection. Data processing and data analytics based on scientifically evaluated crop models helps to interpret the data collected by sensors. Skilled experts, who are linking informatics and field practice, help to work out plant protection strategies.
The AGRON Research Program is focusing on data collection from as much geographic regions as possible for the sake of elaboration of accurate crop models.
The aim of the work is making sense to the use of drones through putting results into practice, thus we are developing a online platform called AGRON Maps for drone data management. The experiments of AGRON took place at the Agricultural Institute, Centre for Agricultural Research, Martonvásár, Hungary.
Statistics of our experiments in 2019:
32 set experiments – 3519 experimental plot – 5010 field sampling
Genotypes: 42 winter wheats – 36 durum wheats – 36 triticales – 36 barleys – 9 maize hybrids
Studying the reaction of varieties to different levels of fertilizers (in case of wheat and maize)
The long-term experiment (since 1980) is focusing on the reaction of species on different level of fertilizers. Treatments are adjusted from the extensive to the intensive agricultural fertilizing methods (0 -> 280 kg/ha, 40 kg/ha steps between treatments). 15 wheat and 9 maize species have been involved in the experiments in every years.
We managed to collect data with SPAD chlorophyll meter and with drone fitted with Micasense RedEdge MX sensor. The sensor was really accurate; thus, the main goal was to calibrate it for chlorophyll measurements, which is essential for long-range spectroscopy-based plant physiology research. The calibration was done on several indexes available in the literature. A representative dataset shows high correlation of SPAD values with NDVI and NDRE indexes counted from sensor data (Figure 1). The structure of the dataset was evaluated and illustrated via principal component analysis (PCA). Each point represents a plot where field and aerial sampling both occurred. The nutrient deficient plots (A) can be separated from optimal and superoptimal fertilizer levels (E and H). This calibration was necessary for accurate data collection and interpretable database construction.
Figure 1 A representative dataset showing the correlation between SPAD values and indexes counted from sensor data (left). Principal component analysis for visualizing data structure (right). (NDVI correlation: 0.761; NDRE correlation: 0.877; A: 0 kg/ha; E: 160 kg/ha; H: 280 kg/ha)
Wheat shows early symptoms of deficiency, however in case of maize it can be detected before ripening. In case of wheat a variety-specific aspect of nutrient reaction can be studied.
Data collection occured at least seven times during the vegetative season of the crops, so crop models could be constructed on nutrient status differences. According to the results it is possible to separate and evaluate the nutrient deficient plots from the optimal and superoptimal fertilizer levels (Figure 2). We managed to elaborate unique look up tables (LUT) or false coloring methods to visualize the differences between the plots. The LUTs are fixed and cannot be changed. This standardization results in informative colors on their own, thus the colors are basic indicators of crop health status in any culture. The collected database allowed us to develop new indexes (e.g. AGRON chlorophyll index), which shows more stability throughout the season and wider range of utility than the classic ones.
Figure 2 Composite RGB (left), AGRON chlorophyll map (middle) and AGRON nitrogen map (right) from the experimental area (winter wheat). Values and color scales are not shown.
Studying the resistance of wheat species against pathogens such as powdery mildew (Blumeria graminis)
These experiments are focusing on evaluation of agronomic value of small grain cereal genotypes bred in Martonvásár. The resistance of varieties and breeding lines are also tested among the characteristics important in cereal production. During the vegetation period 42 winter wheat, 36 spelt wheat, 36 triticale, 36 barley and 36 durum wheat genotypes (released varieties and advanced breeding lines were tested in details). Some plots were treated with fungicide, the others were not. The lack of treatment led to severe infections caused mostly by powdery mildew (Blumeria graminis). In these experiments we could identify infection focal points (hot spots), which plots contained susceptible genotypes. The chlorophyll map shows these resistant (green) and susceptible (yellow or red) varieties or lines (Figure 3).
Figure 3 Chlorophyll map enlarged to the area affected by severe infection (left). The most infected plot is indicated by the mix of yellow and red colors in the middle of the enlarged map. Pictures taken on the field about the most infected plot and plants (middle and right).
Truffle fruitbody spot identification in a young truffle plantation
In Hungary it is in practice to install truffle plantation, which consists of mainly two tree species (hazel and oak). The root of the saplings are inoculated with truffle mycelium. These plantations can be harvested 5-6 years after installation. Truffles can be found only in spots, so one must find it and dig it out to harvest. Monitoring drones make easier to recognise the fruitbody spots in the large plantation areas.
The very first spot of the young plantation appeared under a single hazel in 2018, so we had only one reference sample to find the possible spots in 2019.
The (heavily simplified) null hypothesis for sorting: a symbiotic relationship with mycorrhiza is beneficial for the tree, which can be detected by its better physiological condition compared to other plants without mycorrhiza. This means comparing the multispectral profile constructed and indexes counted from sensor data gives the ability to search for fruitbody spots. We used PCA to evaluate and sort the trees (Figure 4), which showed a group of plants that are similar to the reference sample. These plants are slightly separated from the main point cloud (other plants) and are concentrated in three main spots on the map. In two spots truffles were found. (In these spots only!) We think the third one will be harvested in 2020.
Figure 4 Idetification of truffle fruitbody spot using PCA. The separated group of trees could be mapped into three main fruitbody spots (light blue: hazel; pink: oak). The spots were defined manually.
Improving plant monitoring methods is a key factor in precision agriculture. The big data collected by the sensors are standard and can be used in any culture. However, the analytic methods based on the collected data need to be calibrated and standardized according to the climatic specificity and geographic region.
Due to this fact, local results are of high importance in establishment of a drone service network in Hungary. Crop models help us to determine treatment and protection strategies.
Based on the trends observed in our database we assume, that the identification of nutrient deficiency and pathogen infection will become possible. The progress of artificial intelligence and machine learning will help to analyze sensor data and evaluate the collected information fast and automatically.
Authors: Richárd Turbéki, Tamás Árendás, Gyula Vida, Szabolcs Rudnóy, Kinga Balassa, György Balassa
AGRON Analytics Ltd.
All rights reserved!