Assessment of Semi-Automated Techniques for Crop Mapping in Chile Based on Global Land Cover Satellite Data

Abstract

In recent years, the Chilean agricultural sector has undergone significant changes, but there is a lack of data that can be used to accurately identify these transformations. A study was conducted to assess the effectiveness of different spatial resolutions used by global land cover products (MODIS, ESA and Dynamic World (DW)), in addition to the demi-automated methods applied to them, for the identification of agricultural areas, using the publicly available agricultural survey for 2021. It was found that lower-spatial-resolution collections consistently underestimated crop areas, while collections with higher spatial resolutions overestimated them. The low-spatial-resolution collection, MODIS, underestimated cropland by 46% in 2021, while moderate-resolution collections, such as ESA and DW, overestimated cropland by 39.1% and 93.8%, respectively. Overall, edge-pixel-filtering and a machine learning semi-automated reclassification methodology improved the accuracy of the original global collections, with differences of only 11% when using the DW collection. While there are limitations in certain regions, the use of global land cover collections and filtering methods as training samples can be valuable in areas where high-resolution data are lacking. Future research should focus on validating and adapting these approaches to ensure their effectiveness in sustainable agriculture and ecosystem conservation on a global scale.

1. Introduction

In recent decades, the Chilean agricultural sector has experienced significant transformations as a result of climate change and economic crises [1]. These changes are expected to have major impacts on farmer productivity and the food security of the population [2]. Therefore, there is a need for data sources that can offer information on the trends and fluctuations in cropland at both the regional and national levels. In order to achieve this, crop cover identification is crucial, yet often lacking.

In Chile, reliable data sources for quantifying designated cropland areas and changes in land cover come from periodic in situ statistical surveys conducted by the Ministry of Agriculture of Chile through agricultural censuses [3]. The most recent surveys were conducted in 2007 and 2021. However, these merged datasets do not provide sufficient spatial details, as they only include area calculations without spatial information. As a result, it is difficult to determine where and when changes occur in agricultural land covers [4]. These limitations make it challenging to analyze trends and changes in land cover using field survey data [4,5,6].

To address this issue, satellite-imagery-based land cover mapping has emerged as a viable alternative globally in recent years [7]. This method facilitates the categorization and description of the Earth’s surface, providing essential data for various management and research applications. These applications include cropland georeferencing, urban planning, flood control, disaster prevention, biodiversity conservation, climate change studies and other Earth system sciences [6,8,9].

Global land cover mapping datasets, which are regularly updated, are crucial for analyzing agricultural trends and other environmental applications [7]. Global land cover data, such as NASA’s Climate Change Initiative Global Land Cover Type product [10], the European Space Agency (ESA) land cover product [11] and the MODIS land cover product [12], have traditionally only been available at a relatively low spatial resolution (typically 500 or 300 m). The coarse resolution of these datasets means that they often provide a generalized representation of land cover, which does not meet the increasingly sophisticated needs of a diverse and growing number of scientists and stakeholders who require detailed and high-resolution land cover information [10,13]. This lack of detail is particularly challenging for the accurate identification of cropland, especially given the prevalence of small-scale agriculture in the developing world, which is difficult to detect at coarse resolutions [14]. In Chile, small-scale agriculture is a vital economic sector, and the inability to accurately represent it on global maps can lead to significant misunderstandings of agricultural dynamics. Therefore, there is a critical need to assess the accuracy of land use and land cover (LULC) products and explore options to enhance both the resolution and accuracy of these global land cover datasets.

To address this need, global land cover products have become increasingly available in recent years at a medium spatial resolution, based on images from satellites, such as Landsat, Sentinel 1, and Sentinel 2. Standout examples include the ESA WorldCover 10 m and Dynamic World global land cover maps [10]. Additionally, advancements in large-scale cloud computing and machine learning algorithms in open-source software [15], along with improved access to satellite image collections through platforms, such as Google Earth Engine (GEE) [16], have created new opportunities for utilizing global coverage datasets in agricultural studies [17].

To address the need for high-quality, long-term land cover records, methodologies have been developed to improve the accuracy of labeling in global collections, which may not be optimal in every study area [18]. An effective approach involves using the “filtered” labeling of global collections to generate training samples for classification algorithms [19,20,21]. For example, in 19, the MODIS land cover type MCD12Q1 products were used to extract training data to generate new land cover products [17].

Various researchers have proposed techniques to ensure the extraction of reliable training data from global collections, including the implementation of spatial and spectral filters and the use of only the pixels that have remained consistent over multiple years. By controlling the confidence and reliability of training data, obtaining samples from existing land cover products shows great potential for cropland mapping. The authors of [20] introduced a new method for categorizing a large amount of Landsat data by using high-quality training data from the 500 m MODIS land cover product. Additionally, the authors of [22] successfully created 30 m cropland extent products for Australia and China, with overall accuracies of 97.6% and 94%, using the GEE platform [2]. Furthermore, [23] presents the findings of a cropland extent mapping project for 64 countries across Europe, the Middle East, Russia and Central Asia, achieving an overall accuracy of 93.8% in independent validation samples.

Recent studies have highlighted the importance of utilizing high-quality training samples in LULC classifications. Different researchers have proposed various techniques to guarantee the extraction of reliable training data. For example, the authors of [24] suggested using spatial and spectral filters to remove outliers, while the authors of [20] recommended selecting only pixels from global datasets that have been consistent over a span of three years and refining them using the “metric centroid” method introduced in [25]. These investigations confirmed that by implementing appropriate measures to control the confidence and reliability of training data, acquiring training samples from existing land cover products could greatly benefit the mapping of croplands [26].

In order to meet the need for high-quality data, this study aimed to assess two different methods and three global land use and land cover datasets for mapping the crop cover on a local scale. The objective was to determine the most suitable method and dataset for our specific situation and to provide insights into the applicability of these methods in other regions. Our assessment was based on in situ agricultural surveys (ASs), using the agricultural area as a metric, as it is consistent with the ASs. Additionally, we included a spatial comparison of the crop masks obtained from the various methods and LULC datasets examined.

The first method for calculating crop area involved directly quantifying it from three global LULC databases: MCD12Q1.061 MODIS Land Cover Type Yearly Global 500 m (MODIS) [12], ESA WorldCover 10 m v200 (ESA) [11,27] and Dynamic World V1 (DW) [7]. The second method involved using an alternative approach to reclassify agricultural land cover by utilizing machine learning models. These models incorporated data from Landsat [28] and Sentinel-2 [29,30], along with the three databases previously mentioned, to improve the accuracy of identifying cultivation areas compared to the non-reclassified global maps.

We apply these methods in the Ñuble region of Chile, focusing on the specific year 2021, where both agricultural survey data and satellite-based databases are available.

2. Study Area

The Ñuble region, located in central Chile, is situated between 36° and 38°S and 71° and 73°W. It consists of 21 municipalities, with its capital being the city of Chillán (Figure 1). Agriculture continues to be one of the most vital economic activities in the Ñuble Region, playing a significant role in creating jobs and income for a large portion of this region’s population.

Agriculture also serves as a cornerstone in constructing the identity of its inhabitants [31]. The region grows mainly permanent crops and a rather traditional agricultural landscape, focused on cereals, particularly wheat, as well as other industrial crops, such as sugar beets and rice [32]. This area boasts significant ecological, economic and cultural diversity, with various agroecological units present [31]. According to data from the latest agricultural survey for the reference period 2020–2021, the sown/planted area of the crop category was 154.309 hectares [33].

3. Data and Methods

3.1. Data

In this study, we conducted a comparative analysis between an in situ AS and satellite-based LULC data. The databases utilized for this comparison are detailed as follows.

3.1.1. In Situ Data

We utilized one database from the ASs conducted by the National Institute of Statistics of Chile, known as the Agricultural Census [33]. This survey was carried out in 2020–2021. It provides data on the number of hectares cultivated with various types of crops, categorized by commune, across all the administrative regions of Chile (see

Table 1. Agricultural survey (2020–2021) provided by the Chilean National Statistics Institute [33], including classes and communes in the study area.

Categories		2020–2021
Original classes		Total Cereals Tubers Industrial crops Vegetables fruit trees Vines Flowers Seedbeds Plant nurseries Fodder Improved-grasslands *
Grouped classes		Irrigation, dry, total
Covered area	Ñuble region	Chillán, Bulnes, Chillán Viejo, El Carmen, Pemuco, Pinto, Quillón, San Ignacio, Yungay, Quirihue, Cobquecura, Coelemu. Ninhue, Portezuelo, Ránquil, Treguaco, San Carlos, Coihueco, Ñiquén, San Fabián y San Nicolás.

* The above are divided into three categories: irrigation, dry and without-class.

). Agricultural surveys aim to collect, process and disseminate information about the productive structure of the country’s agricultural and forestry sectors. They are the most important sources of statistical information [33].

The initial step involved harmonizing these two surveys to simplify the data into three primary classes: (i) total irrigated land, (ii) total dryland and (iii) total overall. The total classes encompass the sum of the irrigated and dryland areas, along with those crops not classified under any specific class. These total crop area values were used as a reference to compare the total areas determined by the land cover collections.

3.1.2. Satellite-Based Data

We utilized five global databases: Landsat 8, Sentinel-2, MODIS Land Cover, ESA Land Cover and Dynamic World. For each database, we exclusively extracted pixels classified under the agricultural class. A comparison analysis of these databases, along with the methodology used for the identification of cropland pixels, is presented in

Table 2. Satellite-based databases, with their period, spatial resolution, classes and selected classes that represent agricultural lands.

Database	Available Period	Used Period	Spatial Resol. [m]	Classes or Variables	Selected Classes for Agriculture
ESA WorldCover 10 m v200 [27]	2021		10	Tree cover, shrubland, grassland, cropland, built-up, bare/sparse vegetation, snow and ice, permanent water bodies, herbaceous wetland, mangroves, moss and lichen	Cropland: Land covered with annual crops that are sown/planted and harvestable at least once within the 12 months after the sowing/planting date. The annual cropland produces an herbaceous cover and is sometimes combined with trees or woody vegetation. Note that perennial woody crops are classified as the appropriate tree cover or shrub land cover type. Greenhouses are considered built-up.
Dynamic World V1 [7]	27 June 2015–present (revisit time: 2–5 days depending on latitude)	8/2020–3/2021	10	Water, trees, grass, flooded vegetation, crops, shrub and scrub, built, bare, snow and ice	Crops: Human-planted/plotted cereals, grasses and crops.
MCD12Q1.061 MODIS Land Cover Type Yearly Global 500 m [12,30]	2001–2022 (yearly)	2021	500	Several types of forest: evergreen needleleaf, evergreen broadleaf, deciduous needleleaf, deciduous broadleaf, and mixed, closed shrublands, open shrublands, woody savannas, savannas, grasslands, permanent wetlands, croplands, urban and built-up lands, cropland/natural vegetation mosaics, permanent snow and ice, barren and water bodies.	Cropland: Areas where at least 60% is cultivated cropland.   Cropland/natural vegetation mosaics: Mosaics of small-scale cultivation (40–60%) with natural trees, shrubs or herbaceous vegetation.
Landsat 8 Surface Reflectance	February 2013–present (16-day revisit time)	August 2020 –March 2021	30	NDVI (normalized difference vegetation index) [34], NDWI (normalized difference water index) [35], EVI (enhanced vegetation index) [36] and NBR (normalized burn rate) [37]	-
Sentinel-2 Surface Reflectance	March 2017–present (10-day revisit time)		10

.

Landsat 8

Landsat satellites systematically monitor the Earth’s surface at a resolution of 30 m, capturing multispectral and thermal data every two weeks [28,38]. In this study, six spectral bands were utilized from the surface reflectance collection LANDSAT/LC08/C01/T1_SR available in the GEE platform: blue (B), green (G), red (R), near-infrared (NIR), short-wave infrared (SWIR1) and mid infrared (SWIR2). For each image from Landsat 8, the CFMask algorithm was employed to identify defective pixels, including those affected by clouds, cloud shadows or saturation [39]. This database served not for direct area calculations but to enhance the accuracy of area estimations using the MODIS, ESA and DW databases, according to the procedure in Section 3.2.2.

Sentinel-2

Similarly, the Sentinel-2 satellites [29] acquire multispectral optical data at spatial resolutions ranging from 10 to 20 m, depending on the band. The temporal resolution is 5 days due to the existence of two twin satellites, S2A and S2B. The data were obtained and processed in Google Earth Engine (GEE) from the COPERNICUS/S2_SR collection [30,40] In this study, six bands (B, G, R, NIR, SWIR1 and SWIR2) were used. Quality control bands were employed to exclude pixels affected by clouds [41]. Like the Landsat data, the Sentinel-2 database was not used directly for area calculations but rather to improve the area calculations derived from the MODIS, ESA and DW data.

MODIS

The MODIS land cover product (MCD12Q1.006, available in [30]) offers annual global land cover maps from 2001 to the present at a spatial resolution of 500 m. It utilizes the International Geosphere-Biosphere Program (IGBP)’s classification scheme layers from MCD12Q1 Collection 6 [12]. Each 500 m pixel is categorized into 1 of 17 distinct classes.

ESA World Cover

The ESA WorldCover product provides a global land cover map for the years 2020 and 2021 with a finer spatial resolution of 10 m, created using data from both the Sentinel-1 and Sentinel-2 satellites [11]. This product classifies the Earth’s surface into 11 land cover classes and is freely accessible on the GEE platform [3,30].

Dynamic World

DW presents a near-real-time global land cover dataset with a resolution of 10 m, produced using deep learning techniques and Sentinel-2 imagery with a cloud coverage below 35% [7]. This dataset not only offers class probabilities for nine different land cover classes but also includes near-real-time forest cover probabilities, which were utilized in this research. It is freely available on the GEE platform [30].

3.2. Methods: Agricultural Area Calculation

We employed satellite databases from GEE to calculate the agricultural area for each commune within the Ñuble region, incorporating both the original and processed imagery. The details are described as follows.

We evaluated the MODIS, ESA and DW LULC products, which are all freely available on GEE, to delineate agricultural areas. By selecting data specifically categorized as the cropland class, we conducted a comparative analysis with disaggregated agricultural survey data from 2021 for each commune (see Table 1). Non-agricultural classes were “masked” to isolate the agricultural land, allowing for a precise calculation of the agricultural coverage area in hectares.

Additionally, we developed an alternative approach for reclassifying agricultural land cover using Landsat 8 and Sentinel-2, along with the MODIS, ESA and DW LULC products to enhance the accuracy of these global cover maps in identifying crop areas. The new products are called MODIS v2, ESA v2 and DW v2. A summary of the methodology is presented in a flow diagram in Figure 2.

3.2.1. Reprojection and Filtering

For the DW and ESA datasets, we labeled pixels from Sentinel-2, generating a spatial correspondence between the land covers and the respective bands. In the case of MODIS, the images were resampled to match the spatial resolution of Landsat and transformed to the same map projection (WGS 84 (EPSG:4326)). By default, GEE employs nearest neighbor resampling during reprojection. The land cover class identified in each MODIS pixel was assigned to all 289 corresponding Landsat pixels within that area.

Considering that the timing of crop planting and harvesting varies significantly from one field to another and across larger regions, we used multiple spectral transformations to capture the full spectrum of different crops. Descriptive statistics were calculated from temporal composites of various spectral indices. We computed four key vegetation spectral indices: NDVI (normalized difference vegetation index) [34], NDWI (normalized difference water index) [35], EVI (enhanced vegetation index) [36] and NBR (normalized burn rate) [37]. These indices have been shown to be important for the discrimination of land cover classes [4,23,42], mainly due to their capability to capture the vegetation, urban and water-related characteristics of the landscape [23]. These indices were composited into the 25th, 50th and 75th percentiles for each spectral band using the composite-metric method. Notably, the 25th and 75th percentiles were chosen over the minimum and maximum values to reduce the impact of residual haze, clouds and shadows, which are often artifacts of errors in the CFMask algorithm [26]. The period considered for generating the temporal composition was from August 2006 to March 2007 for the analysis of 2007 and from August 2020 to March 2021 for the analysis of 2021. These periods encompassed the phenological cycle of most crops present in the study region [1].

The overall accuracy of global land cover collections can sometimes be low in certain areas and at specific times [41]. To enhance the accuracy, we implemented debugging steps for pixels designated as training samples (quality filter: edge pixels and yearly consistency, as shown in Figure 2). Initially, we extracted core patches from the labeled crop pixels, these were central pixels surrounded by at least eight pixels of the same type to minimize the misclassification often occurring at the edges [4]. Next, we evaluated the consistency of pixels being classified into the same class during different periods. Only core crop pixels that showed consistent unchanged land cover for two periods prior to the target year of the analysis were chosen for training and validation, as the LULC types of those pixels were likely to be correct [17].

Additionally, the pixel-debugging steps included the following: (1) calculating the spectral centroid for all the pixels initially labeled under a particular class; (2) computing the Euclidean distance (ED) between an individual pixel and its class’s spectral centroid; (3) setting a selection threshold and retaining only those pixels that fall within the threshold distance to the spectral centroid, thereby forming the final reference dataset. A trial-and-error method was employed to find the optimal threshold value for the semi-automatic selection of the training data, as proposed in [43]. We applied threshold ranges based on the interquartile range of the distance to the spectral centroid. It was decided that the optimal threshold value would be defined by the value that conserves the greatest number of homogeneous areas within each class. The table in Appendix A illustrates the Euclidean distance threshold values utilized for selecting the training pixel subset for each class [4].

3.2.2. Training Sample Selection and Classification

In our study, we adopted a semi-supervised approach to classify land cover into 6 categories that we subsequently grouped as crop or non-crop. We consolidated the classes from the global cover collections into six categories (i.e., cropland, grassland, trees, urban, bare and shrub). A step of strata or polygon generation was performed, where each stratum was assigned to one of the six categories. The strata were generated from polygons that delimited areas with homogeneous pixels for each class that remained after the stages of yearly consistency and edge-pixel quality filtering (see Figure 2). Then, we used stratified random sampling (SRS) to gather 1500 training samples for each class [44]. We used SRS because it is a good option for increasing the sample size in classes that occupy a small proportion of the area, thereby reducing the standard errors of the class-specific accuracy estimates for these rare classes and addressing the key objective of estimating the class-specific accuracy [44]. This is particularly important given that some of the six categories defined in the global collections had a small presence in the study area.

For the classification, we employed the random forest (RF) algorithm, which is an ensemble of decision trees trained in parallel [45,46]. This method uses a bootstrap sampling strategy to create a series of decision trees, with each being considered a weak learner, since they utilize only two-thirds of the original training data and a random subset of predictor variables at each node [47]. These trees collectively vote on the classification of unknown samples, effectively creating a robust classifier from multiple weak ones [48].

Research indicates that RF outperforms other classifiers, such as the support vector machines, artificial neural networks and classification and regression trees, especially in scenarios involving high-dimensional data. It offers a high accuracy and is less sensitive to noise and feature selection issues [49,50]. Furthermore, the RF classifier has only two adjustable parameters: the number of selected predictor variables (mtry) and the number of decision trees (ntree). Due to these advantages, the RF classifier is widely used in land cover mapping [20,26,50]. In this study, we set ntree to 100 and mtry to the default value, which was the square root of the total number of input features [4].

Finally, all the land use categories were labeled as either crop or no crop. The areas identified as crop were quantified and compared against those classified as crop from both global cover collections and the AS for the year 2021.

3.2.3. Accuracy Assessment

The accuracy assessment of the crop/no-crop coverage, derived from the improvements made to the global collection data, was conducted using a precision evaluation stage. Several metrics were recommended to assess the classification performance [51]. Central to this evaluation was the overall accuracy (OA), which was used as the metric for model performance in this study and was calculated from the confusion matrix of each algorithm. For this assessment, 3000 validation samples were utilized, sourced from stratified random sampling [44]; 1500 were for the crop, and the others were distributed among the other five classes, grouped subsequently in no crop. These validation samples were independently selected and stratified across each land cover category and were visually confirmed through photointerpretation using Google Earth imagery and Landsat’s RGB composition to ensure an accurate representation of the land cover [42]. Additionally, a careful cross-check was conducted between the training and validation datasets to guarantee that there was no overlap in the time series data being analyzed. Finally, a comparison was conducted by calculating the percentage differences between the crop areas identified by agricultural surveys and the pixels labeled as crop in the LULC, as well as those derived from our training sample generation and classification process.

3.3. Work Strategy

Our work strategy began with calculating the area based on the AS and the data from the MODIS, ESA and DW LULC databases. We then refined the area calculations for the year 2021 (for MODIS, ESA and DW) to facilitate a comparison with the agricultural survey results from 2021 and proposed new approaches, which were named MODIS v2, ESA v2 and DW v2. Subsequently, we determined the error by calculating the percentage difference between the area estimates derived from the satellite databases and those obtained from the AS.

4. Results

In this study we calculated the agricultural or cropland area in two different ways: directly, from original satellite databases (ESA, MODIS and DW) and by processing these databases (ESA v2, MODIS v2 and DW v2). We compared the obtained agricultural areas with those derived from the AS. The results are summarized in Figure 3. The same information can be seen in table format in Appendix B.

The differences among calculated agricultural areas from LULC datasets (ESA, DW and MODIS original and v2) are shown through the error in the decimal numbers regarding the in situ AS. This is shown in

Table 3. Error estimation of agricultural area (km²) retrieval from original and improved (v2) satellite databases (MODIS, ESA and DW) against agricultural surveys. Error in percentages. Negative error means that the database underestimates the cropland area calculation, while positive means the opposite. Year: 2021.

Commune	MODIS	MODIS v2	ESA	ESA v2	DW	DW v2
Chillán	−75	−46	24	−11	67	−32
Bulnes	−85	−53	3	−17	34	−30
Chillán Viejo	−81	−61	68	−8	135	−25
El Carmen	−3	−17	15	3	31	−11
Pemuco	−20	−1	36	24	98	15
Pinto	−45	43	58	143	82	59
Quillón	−100	57	7	286	198	128
San Ignacio	−73	−29	41	25	64	8
Yungay	2	5	30	27	56	19
Quirihue	−92	505	83	726	935	738
Cobquecura	−92	−14	155	455	483	206
Coelemu	−98	201	17	201	239	127
Ninhue	−55	575	236	629	544	439
Portezuelo	−100	176	86	275	427	224
Ránquil	−100	8	−16	145	198	109
Treguaco	−100	312	101	491	614	389
San Carlos	−64	−32	56	16	112	−2
Coihueco	−68	21	22	87	68	7
Ñiquen	49	−26	90	16	142	2
San Fabián	−80	313	190	993	309	303
San Nicolás	−76	−58	57	−23	139	−29

.

From Figure 3 and Table 3, we can see that the coarsest-resolution land cover collection, MODIS, underestimated the extent of cropland (negative error in Table 3). The moderate-resolution collections (ESA and DW) overestimated the extent of cropland in the region. This was expected: as noted in [18], the accuracies of the global collections may not be optimal in some regions, given the specific conditions of each of the regions, in addition to the bias presented by the data used in the training of the global classification models. Against this, the alternative approach of filtering and reclassifying agricultural land, in general, improved the accuracy of quantifying agricultural land. This was observed based on the fact that the agricultural survey quantified the area occupied by agricultural land in the region as 1543 km², a period for which the MODIS collection quantified it at 835.9 km². The improved version of MODIS, that is, MODIS v2, estimated it as 1439 km², improving the precision of the agricultural area calculation by 39%. The DW collection estimated the agricultural land in the region as 2990 km², while DW v2 estimated it as 1728 km², improving the accuracy in calculating the agricultural area by 82%. The ESA and its improved version, v2, showed similar results, overestimating the area of agricultural land present in the region.

For the communes with few areas of agricultural land, such as Quirihue, Quillón, Chillán Viejo, Cobquecura, Coelemu, Ninhue, Portezuelo, Ránquil, Treguaco and San Fabian, the coarse-resolution collection, MODIS, did not label pixels as agricultural land in most cases. On the contrary, MODIS v2 was able to distinguish the presence of cropland in these communes. Global collections with better resolutions (ESA and DW) were able to distinguish the presence of cropland in these communes, demonstrating important qualities that support the use of these types of collections in regions where the presence of agricultural land is limited or where small-scale agriculture is utilized [2].

Considering the communes with extensive cropland greater than 120 km², like Bulnes, El Carmen, San Ignacio, Yungay, San Carlos and Coihueco, MODIS and MODIS v2 underestimated the area occupied by croplands. However, MODIS v2 showed a better accuracy. When comparing these two classifications, the percentage errors went from −83% to −53% in Bulnes, −3% to −17% in El Carmen, −75% to −29% in San Ignacio, 2% to 5% in Yungay, −64% to −32% in San Carlos, −68% to 21% in Coihueco and −68% to 21% in San Ignacio (see Table 3).

By analyzing these same communes, it was found that the better-resolution collections (ESA and DW) and their improved versions (ESA v2 and DW v2) underestimated and overestimated the area calculation in general but with relatively smaller errors than those shown by MODIS (Table 3). For example, in the order of ESA, ESA v2, DW, DW v2, Bulnes showed percentage errors of 3%, −17%, 34% and −30%; El Carmen showed errors of 15%, 3%, 31% and −11%; San Ignacio showed errors of 30%, 27%, 56% and 19%; Yungay showed errors of 30%, 27%, 56% and 19%; San Carlos showed errors of 56%, 16%, 112% and −2%; and Coihueco showed errors of 22%, 87%, 68% and 7%.

Up to this point, the smallest error for the original databases (ESA, MODIS and DW) was made by ESA, while in the processed collections (v2), the best database was DW. This is why in Figure 4, we show the in situ AS, ESA and DW v2 agricultural area maps.

As described in Figure 2, the processing to retrieve v2 for ESA, MODIS and DW involved quality filtering, such as training samples in the random forest (RF) classification algorithms. This reduced the number of pixels that qualified for the training samples. The reduction through the different stages shown in Figure 2 can be observed in Figure 5.

In the application of classification algorithms, an accuracy assessment was performed using confusion matrices and the determination of the overall accuracy (OA). The results of the accuracy assessment showed that the OA of the six-class classification, which was then reclassified into crop and no-crop extent maps, exhibited various levels of OA, as well as the user’s (a measure of commission error) and producer’s (a measure of omission error) accuracies, as provided in

Table 4. Accuracy assessment results for ESA, MODIS and DW classification processes (Tr: trees; Cr: cropland; Ur: urban; Gr: grassland; Sh: shrub; Ba: bare; UA: user accuracy; PA: producer accuracy; OA: overall accuracy).

MODIS		Tr	Cr	Ur	Gr	Sh	Ba	UA
	Tr	238	4	8	7	1	42	0.77
	Cr	21	1156	266	5	0	52	0.92
	Ur	0	53	231	2	0	14	0.40
	Gr	5	10	16	168	45	56	0.62
	Sh	0	0	0	33	265	0	0.84
	Ba	47	36	51	56	4	106	0.39
	PA	0.79	0.77	0.77	0.56	0.89	0.35	OA	0.72
ESA		Tr	Cr	Ur	Gr	Sh	Ba	UA
	Tr	289	1	8	1	0	0	0.95
	Cr	14	1161	268	56	1	0	0.99
	Ur	0	4	251	17	28	0	0.44
	Gr	1	4	21	268	6	0	0.77
	Sh	0	0	19	2	272	7	0.84
	Ba	0	0	0	2	17	279	0.98
	PA	0.97	0.77	0.84	0.89	0.91	0.94	OA	0.85
DW		Tr	Cr	Ur	Gr	Sh	Ba	UA
	Tr	281	2	0	1	15	0	0.88
	Cr	16	1390	36	4	54	0	0.96
	Ur	0	2	291	5	2	0	0.82
	Gr	4	42	19	195	39	0	0.76
	Sh	17	5	8	2	266	0	0.71
	Ba	0	0	2	50	0	244	1.00
	PA	0.94	0.93	0.97	0.65	0.89	0.82	OA	0.89

. The map accuracy varied across satellite collections, with values of 0.72, 0.84 and 0.89 for MODIS, ESA and DW, respectively.

5. Discussion

This study explored the potential of global land cover products, combined with automated methods for collecting training data, to retrieve cropland cover. Understanding and quantifying cropland is essential for the management and conservation of agricultural soils and ecosystem services. Therefore, there is a need for a dataset with good spatio-temporal coverage and a consistent accuracy over time to identify and quantify cropland. This study presents a comprehensive revision of the extent of cropland in the Ñuble region, Chile, derived from various global satellite collections for the target year 2021. Additionally, it utilized the random forest algorithm implemented in Google Earth Engine and employed strategies for the automated generation of training samples from global collections to generate cropland extent maps.

When comparing the results of agricultural surveys with global land cover collections, important differences were evident in the cropland extent and among different data collections. These results were expected, given different factors, such as changes in the environmental conditions. The monitoring of agricultural areas with satellite detection is highly affected by the cloud cover frequency. In [52], the cloud cover frequency over South America was evaluated, and the results showed that Chile is 40% croplands with a high impact on satellite monitoring by cloud cover. Another factor is management practices, which can manifest as spatio-temporal variations in the phenology of the land surface, causing disturbances in the seasonal patterns of the reflectance of the vegetated land surface [53]. This is particularly relevant in our study area, as the Ñuble region is characterized by a significant segment of small and medium-sized farmers [31].

In our study area, there have been similar attempts to calculate the agricultural land cover, such as CONAF land cover monitoring [54]. These maps are obtained from Landsat 8 images using a semi-automated methodology called adapted multi-index, which is based on the combination of different spectral indices (dNBR, dNDVI and CV) that, through integration rules, provide the coverage of land use changes and the directionality of change [54]. Our approach demonstrated better results in the estimation of crop area compared to the results from CONAF (the estimation area for the CONAF database is shown in Appendix C). The discrepancies in the cropland areas between this study and other sources, such as CONAF, can mainly be attributed to differences in the cultivated class definitions, the method of development and the type of data used [23].

Compared to in situ survey methods, the satellite-based method developed in this work, which utilized training samples of existing land cover products, demonstrated many significant advantages, including the possibility of generating sample collections automatically, the production of a large and geographically distributed training dataset and the possibility of generating cropland maps more periodically and at a much lower cost, in agreement with previous research by the authors of [26,55,56]. Therefore, the key innovations made in this study were the production of a high-resolution map with a special focus on the croplands in a large study area with different agro-climatic zones using global LULC products and the successful demonstration of the methodological capability of processing global LULC products using the automated generation of training samples, a machine learning algorithm and the Google Earth Engine platform. We highlight these contributions given the current need to obtain open-source large cropland mapping data [55]. The results obtained indicate that the use of global LULC products must include an evaluation of their suitability with respect to their application in the agricultural sector, considering the spatial resolution needed for the characterization of croplands, the local precision of the collections and biases specific to the crop class. For this reason, we suggest using methodologies to classify cropland cover using spectral information and training labels filtered from global collections.

For future works, we suggest refraining from estimating areas only from the count of pixels obtained directly from global collections but rather to adopt improvements in the filtering of labeled pixels in the global collections and to reclassify the images using classification models fed with spectral variables. This will allow us to quantify and characterize the uncertainty and variability for a given study area.

6. Conclusions

In this study, we carried out a comparison of different global land cover databases, such as MODIS, ESA and Dynamic World, against in situ agricultural surveys. For comparison purposes, we used the area of cropland cover as a parameter. Through the comparison of various resolutions and classification techniques, it has been shown that while lower-resolution collections tend to underestimate cultivated areas, improved methodologies provide much more precise estimates regarding agricultural surveys. This finding underscores the ability of improved approaches to provide a more reliable characterization of the agricultural extent of an area, which is essential for effective resource management and agricultural planning.

This study highlights the importance of using global land cover products combined with automated data-collection techniques to improve the accuracy of cropland identification and quantification. Furthermore, the results illustrate that, despite the limitations of global collections in certain regions, they are useful for describing long-term agricultural trends and may be particularly valuable in regions where high-resolution data are not available. The use of improved collections can be a vital resource for the accurate and effective monitoring of agricultural practices and cropland management. This emphasizes the need to continue improving data-collection and processing techniques to further refine the accuracy and usefulness of global land cover products.

Looking ahead, the adoption of more robust strategies for data filtering and the continuous re-evaluation of classification methodologies are recommended, especially in heterogeneous and small-scale agricultural areas. The integration of emerging technologies and the continuous improvement of global data collections could facilitate the generation of more accurate and periodic agricultural maps at a lower cost. Future studies should focus on validating and calibrating these methodologies in a variety of regional settings to ensure their universal applicability and improve the sustainable management of agricultural resources globally.