Assessing the Accuracy of Multiple Classification Algorithms for Crop Classification using Landsat ‐ 8 and Sentinel ‐ 2 Data

: The launch of Sentinel ‐ 2A and B satellites has boosted the development of many applications that could benefit from the fine resolution of the supplied information, both in time and in space. Crop classification is a necessary task for efficient land management. We evaluated the benefits of combining Landsat ‐ 8 and Sentinel ‐ 2A information for irrigated crop classification. We also assessed the robustness and efficiency of 22 nonparametric classification algorithms for classifying irrigated crops in a semiarid region in the southeast of Spain. A parcel ‐ based approach was proposed calculating the mean normalized difference vegetation index(NDVI) of each plot and the standard deviation to generate a calibration ‐ testing set of data. More than 2000 visited plots for 12 different crops along the study site were utilized as ground truth. Ensemble classifiers were the most robust algorithms but not the most efficient because of their low prediction rate. Nearest neighbor methods and support vector machines have the best balance between robustness and efficiency as methods for classification. Although the F1 score is close to 90%, some misclassifications were found for spring crops (e.g., barley, wheat and peas). However, crops with quite similar cycles could be differentiated, such as purple garlic and white garlic, showing the powerfulness of the developed tool.


Introduction
The use of remotely sensed data to perform crop classification is a complex task. However,it is a useful tool for planning and management of many agriculture's activities like irrigation, among many others [1]. Moreover, the precise and time appropriate information on crop type and surface provided by the remote sensing data are considered the key to forecast the crop production [2]. Therefore, crop classification is the basis for many environmental and socioeconomic applications [3]. Several studies have applied machine learning techniques to classify crop types, for instance support vector machines [4][5][6], neural network [6][7][8], maximum likelihood [9], random forest [4,7] or decision trees [6].
Some applications of crop classification using Earth observation (EO) techniques are: (1) monitoring irrigated land, primarily in semiarid and arid regions [10]; (2) crop monitoring for proper funds allocation, such as subsidies for the Common Agriculture Policy [11]; (3) detection of seasonal crop abandonment [12] and (4) changes in cropping patterns along the years [13].Further, there are many other applications primarily focused on water and land management and governance.
Different sources of information are currently available from both open access and private sources, which empower the capabilities of EO techniques for crop classification tasks. The launch of Sentinel-2A and B satellites has boosted the development of many applications that could benefit from the fine resolution of the supplied information. However, it is still necessary to analyze the contribution of this fine information in vegetation classification [14]. Additionally, the interoperability between Sentinel-2A (MSI) and Landsat-8 (OLI), as the most commonly used remote sensing information sources, must be evaluated [15]. Some efforts have been made in the evaluation of the Sentinel-2A products capabilities for vegetation classification [14]. This preliminary analysis concluded that the main bands that contribute to proper crop classification are located in the red edge and the ones with fewer contributions in the near infrared. The series of spectral bands of Sentinel-2 in the red-edge and shortwave infrared (SWIR; Band 5, Band 6, Band 7,Band 10, Band 11 and Band 12) are exceptionally valuable particularly for the investigation of agriculture and vegetation mapping [14]. It is easy to provide a series of indexes based on the use of these different bands, which are related to the presence of chlorophyll, useful for the discrimination and classification of the type of covering present on a territory [16]. Given that the temporal orbit cycle of Sentinel-2 was especially designed to perform combined observation with Landsat-8 [17], an increase in the crop monitoring mission became possible. Consequently, one of the objective of this manuscript is to evaluate the use the multi-temporal Sentinel-2A and Landsat-8 data to take advantage from their interoperability [15,18]

for crop classification.
Accessible and open remote sensing information has led to applications dealing with multitemporal data, which restricts the results to the end of the crop season. To analyze the effect of the number of dates with the aim of early crop classification during the crop season, [19] evaluated a number of different images along the crop season to determine the earliest crop classification and the improvement in the overall accuracy for the different dates. These authors concluded that using the available information up to mid-July could result in high accuracy (86%) in crop classification, reaching 92% at the end of the season. Additionally, they used the enhanced vegetation index EVI instead of the normalized difference vegetation index NDVI to avoid environmental distortion.
Vegetation indices (VIs) derived from reflectance data acquired from optical sensors have been used over a wide range of scenarios to assess variations in the physiological states and biophysical properties of vegetation [20,21], particularly, the normalized difference vegetation index NDVI [22], which is one of the most used indices in remote sensing studies [23]. NDVI measurements incorporate observations of different plants and can be used to present descriptive characteristics of phenological stages [24].
A great effort has been made in the selection of proper algorithms for crop classification, from the use of traditional parametric algorithms, such as maximum likelihood, to nonparametric algorithms, such as neural networks, support vector machines, decision trees and random forest among many others [3]. Table 1 summarizes some experiences in crop classification with different classification algorithms, sensors, number of classified crops and locations, which complement those referenced in [3]. This study evaluates the contribution of Sentinel-2A information to crop classification over the use of Landsat-8 images in a semiarid region located in the Southeast of Spain. For the same area and data, a classification by using only Landsat-8 images, only Sentinel-2A and both sources of information was evaluated. We also compared 22 classification algorithms, all of them nonparametric, to determine which one is the most robust and efficient. Additionally, we analyzed the interoperability between Sentinel-2A and Landsat-8 and its influence on the classification results.

Overview of the Methodology
To assess the potential of Landsat-8 and Sentinel-2A information for the improvement of crop classification we followed the classical main steps (Figure 1) for crop classification tasks described by [3]:  Data collection. This step covers the selection of remotely sensed and ground truth data. In this study, we selected Landsat-8 and Sentinel-2A images that cover the case study area during 2016. Concerning the data preprocessing, we did not perform any atmospheric correction as we used reflectance at the top of the atmosphere (ToA) to calculate VIs. Crop classification is based on the temporal pattern of VIs and not on its absolute value.Data collection step also comprises collecting the ground truth data from the field visits, which greatly influences the accuracy of the classification procedure.
 Data preparation, which consistsofgenerating VIs, combining VIs from different sources of information (Sentinel-2A and Landsat-8, in this case) and performing statistics of the VIs for each ground truth plot. In this case, selected VI was NDVI, because it is the most widespread used VI and very accurate in the monitoring of the crop phenology.To combine NDVI values from different sources (Sentinel-2A and Landsat-8) a comparison of the mean values of NDVI [32] for each plot and each source of information (after removing border effect and other artifacts) was made. Then, Landsat-8 NDVI values were corrected using the obtained linear relationship between both sources of information.To train the classification algorithm, a wide database of ground truth data was acquired, being one of the main strengths of this study. To ensure high quality input data, ground truth was selected to consider plots with more than 1 ha, eliminate border effect (using a buffer of 30 m) and calculating the mean and standard deviation of the VI for each plot. These values represent the pattern of the vegetation in the plot and, therefore, return the information to perform crop classification.  Classification process, which was carried out using field visits during 2016 of 2032 plots for 12 crops. We calibrated and evaluated the performance of 22 nonparametric algorithms.  Quality assessment and selection of the best classification algorithm. With the 30% of the ground truth data, a quality assessment was done based in the generation of the confusion matrix and classical performance indicators of the overall accuracy, producer's accuracy, user's accuracy and the F1 score [33]. The evaluation of selection of the best classification algorithm was performed based on these performance indicators.

The Case Study
The proposed methodology was carried out in the hydrological unit (H.U.) 08.129 ( Figure 2) with a total surface area of 7200 km 2 . This extension can be considered a wide area that can provide trustful results compared with other studies performed in smaller areas [34][35][36][37][38]. The H.U. is located in the south of the north temperate zone, although it presents a continental nature due to its mean elevation (700 ma.s.l.) and distance from sea. Farmland is the most common type of land use in the aquifer. The most limiting factor for farming is the weather. This area is classified as semiarid (aridity index (AI) 0.26) [39]. Annual reference evapotranspiration values (ETo) are from 1165 mm year −1 in the central area of the aquifer to more than 1300 mm year −1 in the northwest and southeast. Agroclimatic stations showed precipitation values from 336 to 413 mm year −1 with a maximum value of 82 mm in summertime. The analysis of thermal characteristics shows variations from 19.3 to 20.8 °C for annual mean daily maximum temperature and from 6.3 to 6.6 °C for annual mean daily minimum temperatures [40]. These constraints determine the following four groups of crops with different culture systems [41]: (1) cold weather crops that are sown in the autumn or early winter and are harvested at the end of spring or at the beginning of summer;(2) warm season crops for which the growing cycle develops in the summer;(3) rain-fed crops that are limited by the rainfall regimes and (4) irrigated crops. As an example, the crop distribution in 2012 included 14.2% wheat, 14.0% barley, 7.1% maize, 5.8% woody crops, 5.6% opium, 4.9% garlic, 4.4% alfalfa and 3.9% onion, garlic, pea, double crops and other vegetables [42]. The size of the irrigated farms is also important. Farms with more than 100 ha of irrigated area account for 32% of the total area. A significant portion of the total water is consumed by large farms, which are generally equipped with high performance irrigation systems. In addition, collective irrigated farms that were created by public initiative and organized into Water Users Associations (WUAs) are included within the group of farms with a surface area of more than 100 ha. The main irrigation methods include permanent solid set systems (39.2%), center pivot systems (37.8%), drip irrigation systems (17.7%), surface irrigation systems (3.6%) and portable sprinkler systems (1.7%) [43]. After significant efforts were made during the modernization process, farmers demanded management systems and tools to obtain the potential water and energy use efficiency that can be provided by these irrigation systems. In this sense, irrigation has become essential for rural development and maintenance.

Ground Truth Data
The field visits to obtain the ground truth were performed by the Confederación Hidrográfica del Júcar, Spain (www.chj.es) during the irrigation season of 2016. Theyvisited 6341 plots that covered 28,963 ha. After the spatial analysis of the plots, which eliminated plots with a size area less than 1 ha, the number of used plots was reduced to 3111 (24,208 ha). Finally, after the selection of the crops of interest for this study (Table 2), the number of utilized plots was 2032, representing an area of 17,281 ha, which covers 15% of the irrigated area in the Mancha Oriental Aquifer.The preparation of ground truth data wasconducted using QGIS software.
A selection of the main crops (12) established in the area was analyzed. The double crops were not included, although they are relatively abundant, because the field visits were performed for each plot only once a year. These are easily detectable plots due to the special pattern of the NDVI during the irrigation season. Thus, it will not have a negative effect on the results.
To illustrate the NDVI pattern of some studied crops during the period march-July, wepresented in Figure 3 the NDVI temporal behavior of six cropsfrom the selected twelve.

Remote Sensing Information and Processing
Using a parcel-based approach [44],the average NDVI value and standard deviation for each plot were calculated. Only plots larger than 1 ha were utilized to ensure that there was a sufficient number of pixels inside each plot (3pixels×3 pixels for Landsat-8). To select pixels that were completely inside the plot and thus avoid a border effect, a buffer of 30 m inside the plot was implemented. With all the pixels inside the plot that fulfill these conditions, average and standard deviation values of the NDVI were obtained. To determine the threshold value of standard deviation an analysis of these values for two representative dates (May 6th and August 1st) was analyzed. Standard deviation values of all the plots for both dates were sorted and plotted. The values reached the asymptote for an approximate value of 0.1 in both cases, establishing this value as threshold. The observed variations at the plot levelcould be explained by the fact that the farmer did not grow the whole plot due to the effect of some pest or weed or because of improper management. These plots were eliminated to ensure a proper ground truth.
The images detailed in Table 3 were utilized in this case study. Additionally, Table 3 shows the percentage of cloud cover of each image. Cloud cover is a key aspect to consider when assessing crop classification with satellite-based remote sensing techniques [45]. The criteria for selecting a maximum cloud cover percentage is heterogeneous, from 10% [46] to a case in which cloud cover percentage was not considered [47]. Among those criteria, we found references that considered 20% [19], 25% [29,48] and even 60% [13]. In this context of nonstandardized criteria, we selected 40%, which resulted in the elimination of only three images from the set. This is a key issue that should be investigated in areas where cloud cover is more intense (humid or sub-humid regions). To avoid this effect, SAR (Synthetic Aperture Radar) information can contribute in areas were the presence of clouds limits the applicability of the proposed methodology [49].

Analysis of Interoperability between Landsat-8 and Sentinel 2
Although all images were utilized to carry out crop classification, to perform the interoperability analysis between Landsat-8 and Sentinel-2A, only three dates when both satellites covered the case study in the same day were utilized (24 May, 12 August and 31 October). The first date was coincided with maximum vegetation cover of the spring crops, the second date with maximum vegetation cover of summer crops and the last date with senescence of summer crops and orchards.The area included bare soils and crop cover to evaluate the widest range of NDVI values. Figure 4 shows the difference between the NDVI Landsat-8 and Sentinel-2A for the three dates analyzed for a spring crop (two pivots on the east) and a summer crop (two pivots on the west) as an example. On 24 May the spring crops was close to full cover, on 12 August the summer crop was also at this stage, and on 31 October all the pivots were without crops. It can be seen that the difference in NDVI values among both sources of information (Landsat-8 and Sentinel-2A) were higher (blue color) when vegetation is in the low development stage or not vegetating than when the crop was close to full cover. Thus, both sources could be interoperable in the case of the high presence of vegetation, but there is a mismatch among NDVI values of both satellites when there is low crop cover.
To perform the analysis of interoperability, a grid of points separated 30 m in the analyzed area was generated. In each point of this mesh, the Landsat-8 NDVI value was calculated. In the case of Sentinel-2A, the bands of the nine neighboring pixels were averaged and the NDVI value was calculated from these mean bands. Once the data to evaluate the interoperability between the two sources was obtained, a statistical analysis was performed by calculating the coefficient of determination (R2), the root mean square error (RMSE), and the relative error (RE). The normality and homoscedasticity of the residuals were also analyzed. A calibration equation for the area was obtained to interoperate with both sources of data. There is an underestimation of NDVI values obtained with Sentinel-2A compared with those obtained with Landsat-8 for all the dates and the whole data set. However, the dispersion of the data is low, which means that there is a good correlation between both sources of data ( Table 4).
The difference between both sources of data fits a normal distribution and the residuals are homoscedastic. The main statistics of the relations between both datasets are shown in Table 4.

Classification Methods
A set of 22 classification algorithms wasevaluated comprising of decision trees, discriminant analysis, support vector machines, nearest neighbor and ensemble classifiers. These algorithms are the most frequent algorithms used in the literature review summarized in the Introduction Section. Table 5 describes the main characteristics of each method. The ClassificationLearner application of Matlab ® was utilized with the aim of calibrating and validating the different algorithms. The Classification Learner app, can be found in the latest versions of the Statistics and Machine Learning toolbox of Matlab. This application enables automated training for the set of supervised machine learning classifiers incorporated in the toolbox. Besides, trained classifiers can be exported to the Matlab workspace. With Matlab software, the app supports a total of 22 classifier types, organized in five major classification algorithms: decision trees, discriminant analysis, support vector machines, nearest neighbor andensemble classifiers. Additionally, the Classification Learner app offers built-in validation schemes that point out the predictive accuracy of the trained model. The objective of this manuscript was to define the best performing classifier group in crop classification. Thus, the default hyperparameters for each classifier were utilized. In future works, refinement of these hyperparameters will be performed. A detailed description of each of the classifier groups can be found in the basic literature treating about machine learning algorithms.  (Table 3) and the output information is a label corresponding to each crop type. Crossvalidation with 5 folds was carried out for each algorithm in the calibration process to avoid overfitting problems.

Evaluation of Classification Accuracy
A confusion matrix and classical performance indicators of the overall accuracy, producer's accuracy and user's accuracy were obtained. The kappa statistic, which is traditionally used to test the classification accuracy, presents several limitations in accuracy testing [50]. Additionally, the different number of plots for each crop can make kappa values misinterpret the obtained results. Thus, weighted F1 score statistic was used for classification algorithm selection [51] and this accuracy metrics was calculated using aPython library called Scikit-Learn.Once the most accurate algorithm for each dataset has been selected, the remainder indicators, together with the confusion matrix,wereanalyzed to determine the efficiency in the classification for each crop and group of crops.

Evaluation of Classification Methods for Crop Classification
An analysis of the available images that cover the case study for Landsat-8 and Sentinel-2A was performed. The results of using only Landsat-8 information, only Sentinel-2A information and a combination of both sources of information wereobtained for the evaluated classification methods. A comparison between the 22 classification methods with the different combinations of available information is shown in Table 6.
The method with the highestF1 score value for all of the instances of available information wasthe subspace ensemble method, with nearest neighbor learners (M21). Sentinel-2A information slightly improved the F1 score value of the classification compared with using only Landsat-8 information (from 0.86 to 0.87). The use of Sentinel-2A or Landsat-8 independently returned almost the same results. Additionally, ensuring interoperability between both sources of information didnot provide a remarkableimprovement in the classification results. This fact might be due to the similarities between Landsat-8 and Sentinel-2A for high NDVI values, as stated in the interoperability analysis above performed. Additionally, to perform a proper crop classification,onlyNDVI patterns wererequested while absolute NDVI values werenot required. Thus, improving NDVI values with the interoperability procedure didnot contribute to improve crop classification. Among the groups of classification methods, decision tress, discriminant analysis, support vector machines, nearest neighbor classifiers and ensemble classifiers, the second-best classification method for combined information from Landsat-8 and Sentinel-2A wasthe support vector machine with quadratic kernel (M7) with a F1 score = 0.88, and the fine KNN (k-nearest neighbor), (M12) with F1 score = 0.88. Decision trees, which are commonly used in classification, returned poor results, with a F1 score = 0.76 for the best method (complex tree). The discriminant analysis wasnot adequate for classification issues resulting in F1 score values ranging between 0.76 and 0.66. The comparison between the prediction speed (expressed in observation per second) and the classification (F1 Score) is illustrated in Figure 6.

Selection and Evaluation of the Best Method for Crop Classification
To select the best classification method,one should consider the maximum efficiency (maximum F1 score) or a combination of high prediction speed and high accuracy (not the highest) shown in Figure 6. Thus, the subspace ensemble method with nearest neighbor learners (M21) as the most accurate method and the nearest neighbor classifier with fine KNN (M12) as the best agreement or combination between accuracy and prediction speed were evaluated. It is important to mention, that definitely the classifier with the highest accuracy (M21 in this case) had the highest priority to be selected as the best method for crop classification. To evaluate the results of the methods, the confusion matrix and the producer's and user's accuracies (PA and UA) were calculated (Tables 7  and 8).  Crops with UA of 100% means that all the testing plots with this crop were properly classified, which in this case was poppy. Crops with a UA higher than 95% werealmond trees (96.8%), with confusion with vineyard and maize (95.9%) andconfusion with onion and sunflower (there wasone plot of barley that it wasprobably due to mistakes in the field data). Purple garlic (UA=93.1%) was differentiated from white garlic (93.3)even when their phenological cycle was close. Purple garlic plots were classified in one plot as white garlic and in another plot as peas. Barley (UA=82.3%) was confused with wheat for 13 plots, white garlic for four plots, and peas for six plots, with these crops having similar cycles.
When analyzing the results of the M12 classification method, which has a relatively high prediction speed with a high accuracy, the UA and PA werealmost the same as M21. Only slight changes in onion, sunflower, vineyard, peas (one plot misclassified), purple garlic and almond trees (two plots misclassified) were observed. Additionally, some other plots were better classified, such as barley and wheat (one plot improved). Thus, the combined analysis of F1 score value and the confusion matrix wasdeterminant to select the most appropriate algorithm.  In the case of sunflower, there wereonly 41 plots for ground truth, being one of the crops with a lower number of plots visited. This fact led to a worse calibration and testing of the algorithms, with UA values of 69.2% (the lowest value). Additionally, this crop had different types of management depending upon the farmer. Sowing dates can range from early April to late July, which could lead to a misclassification. The 2016 irrigation season presented a cold spring in the area, which delayed the sowing date of the onion. This fact could have led the algorithm to confuse the sunflower with the onion primarily because of the similarities in the initial stages of these crops. This fact would also justify the high PA value for onion (90%), because the main problem detected is that some onion plots were classified as sunflower. Additionally, sunflower is not usually managed to produce the maximum yield to reduce irrigation and fertilization costs, making the development of the sunflower closer to that of the onion.
Another crop with a low number of visited plots was peas, with 98 visited plots and a UA= 70%. For peas, the objectivized yield and growing cycles are heterogeneous. We couldfind proteaginous peas with quite different cycles than the green peas grown for industry. This heterogeneity of growing cycles and the climatological conditions of 2016, which presented severe frost in spring, justify the confusion between peas, barley and wheat.
It deserves highlighting the high UA for poppy (100%), which presents a similar cycle to barley and wheat. This could be due to the fact that is a crop totally controlled by the pharmacy industry and they homogenize the varieties, cycles, and management, leading to the proper classification. Maize and almonds have a UA higher than 95% without any tendency towards misclassification, probably due to mistakes during the field visits works.

Discussion
In this paper, we aimed to test a total of 22 nonparametric classification algorithms, organized in five major classification algorithms: decision trees, discriminant analysis, support vector machines, nearest neighbor and ensemble classifiers.Results showed that among these 22 classifiers the subspace ensemble method with nearest neighbor learners (M21) and the nearest neighbor classifier with fine KNN (M12) werethe best performing methods based on two criteria: high accuracy of the F1 score and high prediction speed. However, as mentioned previously in the ResultsSection, accuracy wasthe most important criteria. Indeed previous research has indicatedthat the integration of two or more classifiers, which is the case of (M21), improve the classification accuracy compared to the use of a single classifier [31,52,53]. Focusing on the results of the other classifiers, we noticed that the decision trees compared to the rest of the algorithms did not provide the most accurate results (the average F1 score value obtained for decision trees in our analysis was equal to 0.68, while the best classifier obtained a F1 score equal to 0.87), but did provide the highest prediction speed ((22000, 20000 and 20000 observation/second respectively for M2, M1 and M3 compared with the best classifier (M21) that resulted in 710 observations/second). In the literature, decision trees classifiers has been widely used in classification process [3] for the just reason that it is known as the speediest classifiers that can be trained [54].According to [53,55], SVM has been considered as a powerful technique for classification tasks, but unexpectedly in this work the coarse Gaussian SVM classifiers (M11) performed less than ensemble classifiers such as (M1, M19 and M21).
Generally the use of only Sentinel-2A data provideda slightly higher F1 score than the use of Landsat-8, and,the difference is barely noticeable ranging between 1%and3%. However for some classifier we observed that the classification with Sentinel-2A data yielded better results (F1 score equal or higher than 4%) than with Landsat-8, such as, (M2,M4,M1,M13, M14, M16 and M20). Regarding the synergic data use, this methodology brought only slight improvement when the following classifiers were tested (M1, M4, M6, M7, M8, M10,M12,M18, M19, M20, M21 and M22), but for the rest of them we observed an F1 score equal (M3 and M11) or even lower (M1,M2, M5, M9, M13, M14, M15, M16 and M17) than the results obtained with only usingSentinel-2A data. Thus, with Sentinel-2A data the majority of the classifiers presented a slightly higher F1 score and when we used both data sources we did not notice any outstanding improvement in the classification results. Consequently, it can be concluded that it is not worthy to include Landsat-8 data especially sinceSentinel-2A has higher spatial and temporal resolution.
Additionally, this study evaluated the interoperability between Sentinel-2A and Landsat-8 through the use of ToA NDVI values. Atmospheric corrections arenot necessary in this application, avoiding an intensive workload. Indeed, in the work conducted by [31], the classification was performed using Sentinel-2 (Level-1C) and Landsat-8 (Level-1T) and the results showed an overall accuracy ranging between 87% and 92%, when classifying crops individually (15 classes) and when crops were grouped based on phenology, prior to classification (seven classes), respectively. Additionally, for the classification process developed by [14], the authors used Sentinel 2 (ToA) product. We obtained adequate interoperability results. This suitability can be explained by the fact that the NDVI normalizes the information obtained from both sensors by reducing the divergence between their spectral bands. However, underestimation of NDVI values was observed when Sentinel-2A was compared with Landsat-8. First of all, at the beginning of the crop cycles when NDVI values are low (at the start of the growing season the soil was not yet covered by the vegetation), but also when NDVI values were high (when the vegetation cover reach the maximum). This matter could have a greater impact on other applications, such as the estimation of water volumes from NDVI values but for classification application, we just need the tendency of the NDVI values along time.

Conclusions
The combined use of Sentinel-2A and Landsat-8 information didnot greatly improve crop classification results compared with using only Landsat-8or only Sentinel-2A. Incorporating interoperability between Landsat-8 and Sentinel-2A information only slightly improved classification. However, in areas where cloud cover is more frequent, having a higher temporal resolution would lead to a better performance of Sentinel-2A over Landsat-8. In addition, the higher spatial resolution of Sentinel-2A compared with Landsat-8 permits classifying smaller plots because the buffer to be utilized would be 10 m instead of 30 m, as is used for Landsat-8 to avoid a border effect.
Between the classification algorithms utilized, the most robust algorithm wasthe subspace ensemble method, with nearest neighbor learners (M21) followed by (M12) the nearest neighbor classifier with fine KNN as the method with the best balance between accuracy and processing time. F1 score values can give a general overview of method performance, but it is necessary to evaluate the confusion matrix with an agronomic perspective to conclude the best method to perform crop classification.
The main conclusion wasobtained with the spring crops, the crop cycles of whichereclose in time and of which variability wasdependent on the sowing date and climatic conditions. Woody crops and summer crops wereeasier to distinguish in semiarid regions because of their difference in the growing cycles.
The use of optical remotes sensing data can be limited by the presence of clouds that introduce missing values in the dataset affecting the capturing of the seasonality of the vegetation cover. For this reason, for future studies we will focus on merging radar with optical data especially with the availability of Sentinel 1 data, which have the same spatial resolution of Sentinel 2. Indeed, several studies ensures that the synergic use of optical and radar data improve the classification OA [56][57][58]. Similarly the use of the data of this work along with radar data can be used to extract valuable phenological parameters to map the crop type behavior of the study area as performed in [59].