1. Introduction
Crop growth and production are influenced by agro-meteorological, soil and crop variables as well as agronomic practices related to water/nutrient supply and pesticide/herbicide applications. Specific attention is given to the availability of water since it is a limited resource in many arid and semi-arid regions, like the Mediterranean [
1]. In these regions, nonoptimal irrigation is a common cultivation practice and water stress occurs frequently as a main abiotic factor that limits crop growth and yield [
2,
3]. Therefore, the evaluation of water stress and efficient water management is of crucial importance for agricultural production, especially in areas suffering from water scarcity.
In the last few decades, coping with water scarcity has become particularly complex due to climate change and the variability of weather conditions [
4,
5]. The impact of climate change on Mediterranean agriculture is already evident in many areas, especially in arid and semi-arid regions [
6]. Several studies reported evidence of climate change in the last few decades over the Mediterranean and tried to foresee the expected trend and its impact on Mediterranean agriculture in the future [
1,
7,
8,
9]. Frequent droughts, flash floods, heat spells and spring frosts triggered a decline in agricultural production, further depletion of water resources, soil erosion and impoverishment, land abandonment and desertification and have increased pressures on food security and socio-economic development, particularly in marginal rural zones [
10].
Numerous studies have been conducted on the impact of climate change on crop abiotic stresses and production at different scales [
2,
11,
12,
13]. Some authors [
11,
13,
14] applied crop growth models to simulate crop growth dynamics and forecast agricultural production under climate change conditions. A common conclusion is that warm and dry climates adversely affect crop phenology and yield. Hence, the challenge is to optimise the use of resources (water and nutrients) while increasing yields and reducing environmental impacts [
1,
12]. In this context, the application of deficit irrigation [
12,
15], mulching [
16,
17] and the use of bio-stimulants [
18,
19,
20,
21,
22] and other water-saving practices have become frequent with the aim of attenuating the negative impact of abiotic stresses, positively affecting plant physiological processes [
19,
20,
21] and improving soil health [
22].
The adoption of a precision agriculture approach based on the continuous monitoring of weather and hydrological variables, water, nutrients and carbon balance represents the preconditions and priorities for interventions and research. Equally, proactive management tools (e.g., early warning systems and water/nutrient management decision support systems considering remote sensing and weather forecasting data) and on-ground measurements are of primary importance to attenuate the negative impacts of extreme weather events and various abiotic stresses [
1,
4,
12,
23]. The success of such measures is based on the interactive use of certified, innovative technological solutions (i.e., the new generation of sensors, unmanned aerial vehicles, artificial intelligence, the Internet of Things and cloud-based applications) and the adoption of site-specific and resource-optimised management practices and varieties able to respond to adverse environmental conditions and to increase/stabilise yields and water productivity [
2,
4].
Potato (
Solunum tuberosum L.) is a common Solanaceous crop. Moreover, it is one of the largest cultivated food crops in the world and has an important nutritional value [
15]. Globally, potato production is about 388.2 million Mg of fresh yield from about 19.3 million ha [
24]. In the Mediterranean region, it is cultivated on more than 1 million ha, with a production of about 32 million Mg of tubers [
25].
Irrigation is necessary to meet crop water requirements due to erratic and insufficient rainfall for most potato-cultivated areas [
26]. When water availability is limited and evapotranspiration demand is high, potato yield is negatively affected, even if briefly exposed to water stress because of its shallow root system [
27,
28,
29]. Therefore, it is important to adopt the best water management solutions as a function of overall water availability, weather and soil data and the crop response to water stress during the entire growing season. An optimal water supply is particularly relevant during tuber development and bulking since these stages are predominantly affected by water stress [
30,
31]. For the rest of the growing season, regulated deficit irrigation strategies might be a solution [
15,
25,
32]. Hence, the estimation of daily crop evapotranspiration adjusted for water stress (ETc-adj) is of vital importance in water-limited agricultural areas. The knowledge of ETc-adj on a daily basis gives an idea of the effective water uptake from the root zone, which supports the optimisation of irrigation scheduling and the enhancement of crop water productivity.
Hydrology and environmental studies, including agriculture, are characterised by complex processes which include many interactions. For example, crop evapotranspiration (ETc) is influenced by atmospheric and soil conditions, plant/canopy characteristics and applied agronomic measures (irrigation water quantity and quality, the supply of nutrients, plant diseases, pests and weeds management, etc.) [
33,
34]. Today, these interactions can be successfully described by modern mathematical tools, including the application of machine learning methods. In the last few years, various machine learning methods have been tested to estimate both reference and crop ET [
35,
36,
37,
38,
39,
40,
41,
42,
43,
44]. In particular, these methods have been developed to enhance the prediction accuracy for the estimation of ETo with limited data availability and ETc under optimal water supply. For example, Yamaç [
45] examined adaptive boosting (AB), k-nearest neighbour (kNN), random forest (RF) and support vector machine (SVM) methods for modelling sugar beet ETc in Türkiye. The models, considering eight scenarios of climate input data demonstrated their applicability for sugar beet ETc estimation. Saggi and Jain [
46] studied regularisation random forest (RRF) and the fuzzy-genetic (FG) models for estimating maize and wheat ETc in India. They found that the proposed FG and RRF models are suitable for maize and wheat ETc prediction. Chen et al. [
47] investigated temporal convolution network (TCN) models comprising long short-term memory networks (LSTM) and deep neural networks (DNN) for modelling maize ETc under mulched drip irrigation. They highlighted that the TCN models performed well in predicting maize ETc under mulched drip irrigation in China. Feng et al. [
48] analysed the reliability of extreme learning machine (ELM) and generalised regression neural network (GRNN) for maize ETc estimation in China. The models confirmed better performance using meteorological and crop data as input variables. Aghajanloo et al. [
49] applied artificial neural network (ANN), GANN and multivariate nonlinear regression (MNLR) models to predict potato ETc in Iran. The results indicated that all the models used could estimate ETc with the intended level of accuracy. While the aforementioned studies demonstrated the applicability of different machine learning methods to estimate ETc under optimal water supply, there is a lack of studies addressing the estimate of crop evapotranspiration adjusted for water stress (ETc-adj) using machine learning techniques. The present study aims to fill this gap.
Crop evapotranspiration under optimal and water stress conditions is commonly computed using the methodology proposed by the FAO Irrigation and Drainage Paper 56 [
33], which considers the impact of weather through reference evapotranspiration (ETo), crop characteristics through crop coefficient (Kc) and the water stress level through water stress coefficient (Ks). ETo is estimated using the Penman–Monteith (PM) equation as suggested by the FAO [
33]. This method is physically based and has demonstrated its superiority when compared to other empirical methods and equations [
34,
50,
51,
52]. The crop coefficient (Kc) is a variable that encompasses numerous crop characteristics, including crop type and variety, crop growth stage, crop density and height, percentage of ground cover, etc. [
33,
34]. Nevertheless, water stress coefficient (Ks) depends not only on crop sensitivity to water stress but also on soil characteristics, including texture (soil water holding capacity) and effective management depth, which is linked to the rooting system growth. Therefore, estimation of crop evapotranspiration requires knowledge and interaction of weather, crop and soil data and their variability during the crop growing season. In most cases, only some of the above data are available, which affects the estimation of crop ET, especially under water stress.
The aim of the present study was to examine the performance metric of machine learning methods (RF, SVM and AB) for the prediction of potato ETc under both optimal and limited water supply conditions, corresponding to full irrigation (I
100), deficit irrigation with 50% of I100 (I
50) and rainfed cultivation (I
0). Five scenarios of available weather, crop and soil input data were considered. The study aimed to determine the best performance metric for each specific input data scenario for estimation of potato evapotranspiration under optimal conditions (ETc) and under deficit water supply (ETc_adj). Thus, unlike the previous study [
53], which focused on optimal water supply, this study focuses on predicting ETc under water stress and various scenarios using weather, soil and crop data. To the best of the authors’ knowledge, this is one of the first studies to evaluate machine learning methods for prediction of ETc_adj, i.e., crop evapotranspiration under water stress, across different data availability scenarios.
4. Discussion
Considering different data availability, the overall results demonstrated that the performance of models was improved with an increasing number of input variables. Nevertheless, the greatest impact on the simulation results was observed when weather data and crop coefficient values were available. The availability of soil characteristics, root depth and fraction of available water had a minor impact on simulation results. The SVM model showed the lowest predicting accuracy in all input scenarios under three water regimes. Substantial overestimation of crop evapotranspiration was observed, indicating low suitability of SVM for the prediction of crop ET.
The AB model offered the best prediction accuracy for four out of five scenarios of data availability, while the RF model offered the best prediction accuracy for the scenario where weather and crop coefficient data were available under optimal irrigation regimes (I100). In the case of deficit irrigation and 50% of the optimal irrigation supply (I50), the RF model was superior for all scenarios of data availability. Under the rainfed condition, the RF model fed only with climate data had better prediction accuracy than SVM and AB models. However, the AB model had the best performance for the other four scenarios of data availability. In general, the machine learning models demonstrated better prediction of crop evapotranspiration under rainfed (I0) and optimal irrigation regimes (I100) than under deficit irrigation (50% of the optimal irrigation regime).
In general, the assessment of performances of machine learning methods under different irrigation regimes showed that RF, SVM and AB models were able to simulate complex and nonlinear relationships among the climate, soil and crop parameters and to adequately estimate crop evapotranspiration under different water regimes. This can be explained by the ability of selected models to autonomously solve complex and nonlinear problems by gathering datasets from various sources [
70]. The models’ performance improved with an increasing number of input variables [
42]. The greatest positive impact on the models’ performance was observed when the crop coefficient data were added to weather data availability (scenario 2).
Under nonlimited water supply, the soil water balance and, therefore, crop evapotranspiration depend substantially only on the weather variables and specific crop growing stage because there is no water stress and crop parameters (root depth, fraction of readily available water and total available water) are not relevant to support crop growth. As water supply is reduced, the importance of the crop’s root depth and availability of water within the root zone increases because they regulate the crop’s response to water and determine the capability of the crop to use water. In fact, under limited water supply, crop evapotranspiration is reduced as a function of water availability in the root zone and crop-specific sensitivity to water stress.
The assessment of stability of machine learning methods under different irrigation regimes revealed the instability of the RF and AB models as they produced changes in prediction accuracy when new input data were applied under optimal irrigation regimes (I
100). Similar findings were pointed out also by Hassan et al. [
71] who mentioned that the RMSE of the RF model had large differences between training and testing values. For all models, the RMSE of the testing subset was lower than for the training subset only in the case of ETc estimate under optimal water supply. For the other two cases (deficit irrigation and rainfed cultivation), the difference in RMSE between training and testing subsets was only 2–3% and, for some scenarios, it was greater for testing than for training subsets. Therefore, it can be concluded that the performance of models is more stable under water stress conditions than under optimal irrigation supply.