1. Introduction
Late blight, caused by
Phytophthora infestans, is a major disease of potato crop, with a strong negative impact on tuber yield and quality [
1,
2,
3]. The pathogen is regarded as a threat to global food security because worldwide losses due to late blight are estimated to exceed annually
$5 billion [
4,
5].
P. infestans is described as a lower water oomycete and infects the potato crop through the tuber and soil during cool and wet weather. Infection of shoots can be caused by mycelium growing from the tuber into the developing shoot or through sporangia and zoospores formed on the tuber surface under wet conditions [
6]. Then, the potential risk of disease development depends in part on the aerial transport of
P. infestans sporangia to potato fields from neighbor’s infection fields [
7,
8,
9,
10]. Depending on host susceptibility and environmental conditions, the first symptoms can be visible 3–4 days after infection [
1]. Night temperatures of 10–16 °C accompanied by light rain, fog or heavy dew and followed by days of 13–16 °C with high relative humidity are ideal conditions for late blight infection and development [
11]. The first symptoms are followed by the production of new sporangia and the infection cycle is repeated as many times as the weather conditions allow the viability of the released sporangia [
6,
8,
12,
13].
Potato crops contributed to alleviating world hunger and potatoes remain one of the agricultural resources needed in line with the zero-hunger sustainable development goal. However, the crop management strategies increasingly require the application of phytochemicals, which may cause undesirable effects on the environment and on human health, with an increasing production costs for growers. For effective and ecofriendly management of potato late blight, scientific and technical efforts must be made to understand how the disease progresses and how this progression can be slowed down. One of the important issues in late blight management is to forecast when, where and how abundant airborne inoculum will be, to prevent the onset of the epidemic. The airborne inoculum of the pathogen appears to have a significant impact on the disease epidemic. However, the prediction of airborne spores of plant pathogens is difficult because they are influenced by a plethora of factors (temperature, relative humidity, leaf wetness, wind, phenological stage) [
7,
10,
12,
14,
15]. Hence, the efforts towards understanding and predicting airborne sporangia of
P. infestans based on multiple factors that condition its development, with a significant impact on the management of late blight, with less fungicide applications are sought.
In recent years, the agricultural sector was able to adopt the main technological innovations relying on artificial intelligence (AI), artificial neural networks (NN) and machine learning (ML). The goal is to digitize itself and increase the autonomy of many processes by making better data-driven decisions, reducing the workload, inputs and increase the quality of the final product [
16,
17,
18,
19,
20]. The multi-view spectral information from unmanned aerial vehicles (UAV) based color-infrared images combined with machine learning algorithms was used to improve the estimation of nitrogen nutrition status in winter wheat and optimize the fertilization [
17]. Classification methods and clustering trough image analyses such as neural networks (CNN) were used to simulate the humans’ decision-making process. CNNs were shown to have great potential for fine classification problems using an image of the same object from different views [
18,
19]. Decision trees, support vector machines or k-means together with information from foliage of the crop were used in precision agriculture and the effective detection, identification and quantification of plant diseases [
21,
22]. In the case of potato crop, ML algorithms were recently applied for monitoring diseases through image-based techniques [
20,
23,
24,
25,
26,
27]. Sugiura et al. [
24] proposed a phenotyping system for mapping late blight on potato crop by analyzing pixel change between consecutive images. The assessment of late blight severity in potato by acquiring high resolution multispectral images with a low-cost camera and ML algorithms was also reported [
25]. More recently, the early detection and severity assessment of late blight in potato crops by multispectral imagine were evaluated [
26,
27]. However, these studies focused on the detection and identification of disease after the onset of the infection process. The preceding step is the early detection of inoculum in the environment of the crop able to cause first late blight symptoms on the potato canopy. Aerobiology is an excellent discipline for this purpose, allowing real time knowledge of sporangia in the potato atmosphere [
9,
10,
14,
15,
28,
29,
30].
There were some aerobiological studies that focused on understanding the influence of climatic factors on the dynamics of spores in the atmosphere of the potato crop using different multivariate statistical techniques and ML algorithms [
31,
32,
33]. However, there were fewer studies trying to predict
P. infestans sporangia levels in the environment crop [
10,
15]. Furthermore, despite the great available scientific information on late blight, few studies focused on the specific value of airborne sporangia concentration as a monitoring tool for late blight control [
10,
14,
15,
29,
34,
35]. In this sense, with the purpose of estimating the late blight risk during the early stages of potato crop development, ML algorithms were applied. The goals of the present study were: (i) to assess the concentration of
P. infestans in each phenological stage of potato crop in northwest Spain; (ii) to derive a simple binary classification model for predicting the days exceeding the aerobiological risk level of pathogen; and (iii) to validate ML algorithms as a tool for forecasting late blight outbreaks.
4. Discussion
The understanding of different factors that support the aerial dispersal of
P. infestans for the correct prediction of late blight epidemics is crucial [
7,
8,
10,
14,
15,
34,
44]. However, experimental studies under field conditions with the aim of predicting sporangia are limited. In this sense, the present study is one of the few that predicted airborne sporangia using the value of the inoculum quantified in the crop environment and at the same time, considering climatic factors. Several factors (e.g., wind, temperature, solar radiation, rain) could be used to forecast the outbreak of late blight [
45]. This study focused on the main weather factors (T, RH, LW and wind) and derived variables (IP, SR and escape) with inoculum quantity to predict daily sporangia risk level that cause outbreaks of potato late blight. The results of this research support the importance of considering inoculum to make decisions of fungicides applications. This information combined with decision support system for late blight would improve the exactitude of warnings for a correct management of fungal treatments. The first application rate could be predicted before the onset of the disease in the field and consequently, the number of chemical applications would decrease.
According to previous research, linear regression models or neural network can predict with success disease behavior such as mycotoxin secretion in crops such as grapevine, rice and wheat or improve the ability of crop-growth monitoring [
17,
46,
47,
48,
49,
50]. In potato, the publications related with prediction of inoculum of late blight in air using ML algorithms are non-existent. For effective management of potato late blight, efforts must be made to slow the progress of the disease, especially by reducing the primary inoculum. In this sense, ML algorithms were applied because of their usefulness in managing large databases with multiples variables, images or spectra [
17,
18,
19]. The performance of the models was evaluated as a binary classification system, categorizing the daily sporangia into two classes by aerobiological criteria (ARL) according to the daily sporangia concentration greater or less than 10 sporangia/m
3. The C5.0 and RF algorithms were good and robust algorithms to predict days with high detached sporulation of
P. infestans, resulting a ROC of 0.903 and 0.902, respectively. These ML algorithms provided good predictions in other potato pathogen, such as
Alternaria spp. [
37]. The ranking of the variables of importance by RF and C5.0 algorithms took the sporangia variables from previous days as the most influential to develop the prediction models. This suggests a strong influence of previous sporangia counts in predicting sporangia level at present, as also corroborated by the Spearman correlation test. A strong influence of sporangia of previous days in the presence of sporangia of present day was observed. This fact coincides with previous studies on
P. infestans [
10,
15] and
Alternaria [
33,
37,
51,
52] on potato crop, as well as
Botrytis cinerea [
53] and
Uncinula necator [
49] on vineyards. This trend emphasizes the need to continually monitor the airborne spores, as they will be needed to accurately predict future spread of spores.
To predict the behavior of
P. infestans, it is important to know about crop season and geographical area [
7]. Typically, the potato growing seasons in northwest Spain run from the beginning of May to the end of September. During this period, the temperature is the key weather factor for the growth of the potatoes. Throughout the growing season, spring rains are common and contribute to the development of the crop during the first few weeks. However, temperatures and high humidity also prove to be suitable for the presence of certain diseases for potato crop, such is the case of late blight. The production, dissemination and germination of
P. infestans sporangia, as well as their penetration into host tissues are particularly influenced by mild temperatures and high relative humidity [
1,
7,
8,
10,
14,
15,
34,
44]. According to the Spearman correlation test, the coefficients between T and the sporangia presence were significantly negative. On the contrary, RH and LW variables showed a significantly positive relationship with the sporangia presence. Both algorithms agreed with IP and SR as the variables of higher importance. It is known that optimum temperatures for late blight epidemics are between 16 and 23 °C [
6]. The IP variable combines the effect of T and RH under one variable, considering optimum values to late blight infection of 10–24 °C and RH > 88% [
10]. In the studied area, mean temperatures between 16–21 °C were repeated during the whole period. In addition, rains and high humidity in the first months of crop development can explain higher sporangia concentration during foliar development stage. Under optimal temperature range to late blight epidemics, the lack of rain and dry weather could decrease the infection process and sporulation, as our results showed, with a decrease in rainy days and RH in flowering and senescence stages. Furthermore, it was shown that temperatures above 28 °C negatively affect sporangia production [
54,
55]. Thus, this could explain the lowest sporangia concentration trapped during senescence of five studied crop seasons.
The results agree with previous studies that showed changing dynamics of the
P. infestans inoculum concentration during the potato growing season [
9,
10,
14,
28,
34]. However, these studies focused on the prediction of late blight risk solely based on weather factors, while assuming the constant presence of inoculum [
55,
56,
57]. Although the infection of potato plants with
P. infestans is highly dependent on weather conditions, late blight epidemics cannot be explained exclusively by weather data [
7,
54]. Aerobiological information can be useful to avoid false alarms of infection risk when climatic conditions are favorable, but if there is no presence of inoculum in the air, there is no risk [
34]. The variability in the sporangia concentrations during each phenological stage confirm that the presence of aerial inoculum in any place is not an unlimited factor. Thus, the climatic conditions, the local topography, the associated fungal host and the phenological state of the plant condition affect this variation [
9,
10,
14,
15,
30]. Despite the scarcity of studies that considered quantification of the airborne inoculum, the present results agree with researchers that support its presence as essential to know the real late blight pressure in a particular area and to predict new reinfections [
9,
10,
13,
14,
15,
29,
35,
54]. However, for the success of the proposed methodology, several key factors are necessary: having several years of study, meteorological stations on the plot itself or in the vicinity of the monitored area, and specialized personnel for the extraction of aerobiological and meteorological data, as well as the correct treatment of algorithms with complex statistical methods. This type of research has great practical utility in the agricultural sector because it allows farmers to detect the infection before it manifests itself. Consequently, these professionals achieve greater effectiveness in the application of chemical treatments, and reduce investment in preventive treatments. The search for more environmentally sustainable agricultural solutions and tools to minimize the impact of these changing fungal diseases in recent years will have an impact on the economic value of the final product placed on the market. At the same time, it will favor food security and human health in the world.