Dynamics of Irrigated Land Expansion in the Ouémé River Basin Using Field and Remote Sensing Data in the Google Earth Engine

Abstract

The availability of reliable and quantified information on the spatiotemporal distribution of irrigated land at the river basin scale is an essential step towards sustainable management of water resources. This research aims to assess the spatiotemporal extent of irrigated land in the Ouémé River basin using Landsat multi-temporal images and ground truth data. A methodology was built around the use of supervised classification and the application of an algorithm based on the logical expression and thresholding of a combination of surface temperature (Ts) and normalized difference vegetation index (NDVI). The findings of the supervised classification showed that agricultural areas were 16,003 km², 19,732 km², and 22,850 km² for the years 2014, 2018, and 2022, respectively. The irrigated land areas were 755 km², 1143 km², and 1883 km² for the same years, respectively. A significant increase in irrigated areas was recorded throughout the study period. The overall accuracy values of 79%, 82%, and 83% obtained during validation of the irrigated land maps indicate a good performance of the algorithm. The results suggest a promising application of the algorithm to obtain up-to-date information on the distribution of irrigated land in several regions of Africa.

1. Introduction

Agriculture is the backbone of the Beninese economy, contributing 20.87% to the gross domestic product (GDP) [1]. However, it is highly dependent on rainfall, which makes it extremely vulnerable to the negative effects of climate change, threatening food security and people’s livelihoods [2]. Significant spatial and temporal variations in rainfall, prolonged dry spells, extreme heat, and late rains have made agricultural production very uncertain, leading to decreases in crop yields [3,4]. In this situation, the provision of additional water through irrigation becomes necessary for boosting agricultural production [5]. According to McAllister et al. [6], irrigation is a means of providing crops with the quantity of water they need to grow. This practice improves agricultural production while maintaining or increasing yields in the face of variable climatic conditions [7].

Benin has significant potential in terms of water and land resources, which are essential for the development of irrigation [3,8]. These resources have been divided into four major hydrographic entities, including the Ouémé River basin (ORB) [5]. The ORB is the agricultural breadbasket of Benin, with an estimated potential arable land of 2.6 million hectares. The basin’s surface and groundwater resources have been estimated at around 5.4 billion m³/year and 755 million m³/year of water, respectively [9]. Although such potential exists, irrigation activities are not widespread in this area [10]. To make better use of the available potential, the Master Plan for Water Development and Management (MPWDM) for the basin proposes the installation of hydraulic infrastructures (multifunctional and hydro-agricultural dams) on the Ouémé transect to ensure the irrigation and agricultural development of over 132,300 hectares of land [9,11]. Considering the future implementation of hydraulic infrastructures for irrigation in the basin, it is essential to have explicit knowledge of the evolution and current extent of irrigated land for better planning of water resource use [12,13].

According to the literature, existing studies on irrigation in Benin have focused on agricultural production [14,15,16] and the typology of irrigation systems [3]. To date, no study has been carried out on the mapping of irrigated croplands. Moreover, estimates of the geographical extent of irrigated croplands remain incomplete and are limited to the statistics of the Food and Agriculture Organization of the United Nations (FAO) available in the AQUASTAT database, as well as to agricultural data collected by government agencies and non-governmental organizations [17]. Unfortunately, these statistics mostly do not reflect the real conditions of irrigation evolution, as they may be influenced by political decisions or also collected from conventional methods that require a lot of financial means and time [18]. In this context, a comprehensive mapping study of the current extent of irrigated croplands in Benin, and more specifically in the Ouémé basin, is needed to understand the possibilities for future expansion.

Notable advances in remote sensing and geographic information systems (GIS) have enabled the scientific community to implement various cost-effective methods for mapping irrigated areas in different parts of the world [13,19,20,21]. The works carried out by Ozdogan et al. [22] and Massari et al. [23] offer a synthesis of the different approaches developed in the literature to assess the spatiotemporal dynamics of irrigated croplands. For example, Magidi et al. [20] mapped irrigated areas in the Mpumalanga region of South Africa using high-spatial resolution satellite images (Sentinel-2 and Landsat 8). In their study, the authors employed a supervised classification procedure based on the random forest algorithm available on the Google Earth Engine platform, and thresholding techniques were applied to the normalized difference vegetation index (NDVI). Similarly, Chandrasekharan et al. [24] applied the same approach to multi-temporal Landsat images to map irrigated and rainfed areas in Ethiopia during 2015–2016. In addition to NDVI time series, other indicators such as rainfall, NDMI, and slope were used in their study. Traoré et al. [25] and Traoré et al. [26] attempted to apply other classification techniques (i.e., the maximum likelihood classifier, MLC; and the support vector machine classifier, SMV) to multi-temporal Landsat images to assess changes over time in the area of irrigated lands in Burkina Faso. The first study was carried out in the Kou watershed, while the second was carried out in the Mogtedo region. In addition, studies conducted by Wu and De Pauw [27], Pervez et al. [28], McAllister et al. [6], Ghebreamlak et al. [7], Kant and Mishra [29], and Abuzar et al. [30] have developed a simple and modifiable algorithm to extract and evaluate changes over time in irrigated cropland from satellite images from different sources (Landsat, ASTER, MODIS, SPOT). The algorithm developed consists of a logical expression based on a combination of thresholds for surface temperature (Ts) and vegetation indices (NDVI). Although the application of this algorithm revealed very promising results, the authors noted that land occupations such as forests and scrubland may present similar vegetation and surface temperature (Ts) indices as irrigated areas during the irrigation season, which may affect the classification results. Therefore, they proposed an intermediate differentiation and thresholding technique based on seasonal vegetation changes to separate agricultural from non-agricultural areas before applying the algorithm. As the implementation of this method has yielded satisfactory results in parts of Asia [27,28,29], Australia [6,30], Morocco [31], and Sudan [7], the authors highlighted the need to extend this approach to other environments, including the regions of West Africa, where numerous efforts have been made in recent decades to develop irrigation. For this reason, the algorithm based on the combination of NDVI and Ts thresholds was applied to agricultural zones in this study to assess the expansion dynamics of irrigated land in the Ouémé River basin. To delineate the agricultural zones in this study, the supervised classification method based on the RF algorithm available in the GEE platform was employed. The results of this study would be essential for implementing irrigated land expansion scenarios in the Ouémé River basin.

2. Materials and Methods

2.1. Study Area

The research was conducted in the Ouémé River basin in West Africa, an area known for its significant agro-hydrological potential. This region is drained by the Ouémé River, which spans 510 km in length. At the Bonou outlet, the basin encompasses an area of 49,256 km², situated between latitudes 10°09′33″ N and 6°20′14″ N and longitudes 1°30’E and 2°30′ E (Figure 1) [32]. The topography of the basin is generally flat, with an average slope ranging from 0.9 m/km to over 2% [11]. The study area exhibits three distinct rainfall patterns [33]: (i) a unimodal regime in the north, characterized by a rainy season from May to October and a dry season, typical of the southern Sudanese climate; (ii) a bimodal regime in the south, featuring two wet seasons, a longer one from March to July and a shorter one from September to mid-November, as well as a prolonged dry season from November to March, representative of the Sudano–Guinean (sub-equatorial) climate; and (iii) a transitional regime in the central region, resembling the climate observed in Benin, marked by a rainy season spanning from March to October, occasionally with a brief dry spell in July and August. The average annual precipitation across the basin ranged from 900 to 1300 mm during the period ranging from 1951 to 2015 [33]. Over the period ranging from 1991 to 2021, the mean annual temperature fluctuated between 26 and 30 °C (Figure 1). The predominant land cover in the study area is savannah, representing 64% and 60% of the catchment area in 2000 and 2013, respectively. However, this vegetation has significantly declined in favor of agricultural land, which increased by 8.1% during the period of 2000–2013 [32]. Small-scale rainfed agriculture is prevalent in the study area [11]. During the dry season, which generally extends from November to March, water availability for crops becomes a significant challenge [8]. Consequently, agricultural activities are primarily observed on irrigated areas during this period [10]. Three categories of irrigation have been identified in the basin area: (1) urban and peri-urban irrigation based on vegetable production (onions, tomatoes, chillies, okra, carrots, etc.) near major cities, (2) small-scale irrigation in inland valleys (lowlands and floodplains), and (3) large-scale irrigation systems for rice, sugarcane, and palm oil production [3,8].

2.2. Methods

A two-step methodology inspired by the work of Wu and De Pauw [27], Kant and Mishra [29], and Chandrasekharan et al. [24] was implemented in this study to assess the temporal evolution of irrigated areas in the Ouémé River basin. It firstly consists of the application of the supervised classification procedure based on the random forest algorithm to Landsat images in the GEE to differentiate agricultural land from other land uses, in particular evergreen vegetation. Next, the existing relationship between soil moisture, surface temperature, and vegetation indices (Ts/NDVI) illustrated by a logical operation and thresholding of a combination of NDVI and Ts as detected by the remote sensing technique was exploited to differentiate irrigated from rainfed areas in the delimited agricultural areas. Figure 2 illustrates the methodology adopted in this study.

2.2.1. Satellite Image Collection and Pre-Processing

Landsat image time series (L7, L8, L9) for the years 2014, 2018, and 2022, provided by the United States Geological Survey (USGS), were consulted and processed on the GEE platform to achieve the various specific objectives of this study. Landsat imagery was used because it provides a long time series coverage of satellite images across the study area. Furthermore, it offers the highest spatial resolution of the thermal infrared (TIR) band, essential for determining surface temperature (Ts) [34]. To mitigate the adverse effects of cloud cover on image quality within the study area, images were acquired during the dry season (November to March). In addition, a series of pre-processing procedures were carried out to improve image clarity [20]. Given the challenge of procuring single images encompassing the entire study area for each designated year, image composites were generated. These composites were derived by filtering image scenes along the study area boundaries, retaining only scenes having less than 20% cloud cover, and then computing the median of all pixels satisfying this specified filter.

Table 1. Characteristics of the Landsat images used in this study.

Main Uses	Satellite Images	Used Bands	Wavelength (um)	GEE Dataset IDs	Spatial Resolution (m)
Land use and land cover (LULC) classification	Landsat 7	Blue: B1 Green: B2 Red: B3 NIR: B4 SWIR1: B5 SWIR2: B7	0.45–0.52 0.52–0.60 0.63–0.69 0.77–0.90 1.55–1.75 2.09–2.35	C01/T1_SR	30
	Landsat 8	Blue: B2 Green: B3 Red: B4 NIR: B5 SWIR1: B5 SWIR2: B7	0.45–0.51 0.53–0.59 0.64–0.67 0.85–0.88 1.57–1.65 2.11–2.29
	Landsat 9	Blue: B2 Green: B3 Red: B4 NIR: B5 SWIR1: B5 SWIR2: B7	0.45–0.51 0.53–0.59 0.64–0.67 0.85–0.88 1.57–1.65 2.11–2.29
NDVI extraction	Landsat 7	Red: B3 NIR: B4	0.63–0.69 0.77–0.90	C01/T1_SR	30
	Landsat 8	Red: B4 NIR: B5	0.64–0.67 0.85–0.88
	Landsat 9	Red: B4 NIR: B5	0.64–0.67 0.85–0.88
Ts retrieving	Landsat 7	TIR: B6	10.4–12.5	C01/T1_TOA	60
	Landsat 8	TIR: B10	10.6–11.19		100

provides an overview of the different characteristics of the utilized images.

2.2.2. Agricultural Area Mapping

Following the example of Magidi et al. [20] and Chandrasekharan et al. [24], the classification procedure based on the RF algorithm available in GEE was applied to Landsat time series to extract agricultural land in the study area. Among the different classification algorithms documented in the literature, the RF machine learning classifier stands out as one of the most widely employed methods [35,36]. The RF algorithm functions based on the generation of multiple decision trees. Given the proven effectiveness of employing 500 trees, this number was chosen to build the RF classifier for this study [21,35]. While the primary objective was to generate agricultural land maps for the designated years, additional land use categories such as water body, vegetation, and built-up area/bare land were incorporated into the classification process. This step aimed to prevent the misclassification of certain non-agricultural pixels as agricultural classes, given their spectral reflectance similarities to agricultural land [24]. To this end, sample points for training the RF algorithm and validating the final land cover maps were collected on the basis of land cover detected by the visual interpretation and knowledge of the characteristics of the study area. In addition, reference land cover maps (ESA land cover, Dynamics World Cover…) archived in the GEE were also used. The training and validation samples for the target year comprised 1615, 1660, and 1646 land use points, respectively (Figure 3a). As employed in the existing literature, the 70/30 ratio was applied to the sampling points to train the classifier and validate the results [35,36].

Table 2. Description of the land cover categories used.

N°	Land Cover	Description
1	Water	Water bodies, e.g., temporary and permanent watercourses, wetlands, reservoirs, etc.
2	Agricultural area	Land used for seasonal crops, fields, and fallow land under palm trees, areas developed and equipped for irrigation, etc.
3	Vegetation	Shrubs, trees, grasslands, natural forests, plantations, savannahs, degraded forests, etc.
4	Built-up and bare land	Dwellings, industrial zones and road infrastructures, hills, bare ground, etc.

provides an overview of the various land use categories employed in the classification.

2.2.3. Irrigated Area Mapping

The maps of irrigated areas were produced by analyzing seasonal changes in vegetation and soil moisture status in agricultural areas during the dry season, applying an algorithm based on the logical operation and thresholding of a combination of surface temperature (Ts) and vegetation index (NDVI). This algorithm assumes that in agricultural areas during the dry season (generally the irrigation period), due to high soil moisture and the presence of crops, irrigated areas have a lower Ts surface temperature than rainfed areas and also a higher NDVI than rainfed areas [27,28,29]. The development and application of the algorithm to extract irrigated areas can be summarized as follows:

-

Conversion of spectral and thermal bands of Landsat images for the determination of NDVI and surface temperature (Ts);

NDVI is a greenness scale successfully used to map irrigated areas at several scales, often with original spectral bands and other spectral indices. In general, NDVI values range from −1 to +1. Equation (1) [20] was used to determine NDVI values.

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{e}\mathrm{d}}}$$

(1)

The NIR and red bands derived from the Landsat images are shown in Table 1.

The surface temperature (Ts) was retrieved from multi-temporal Landsat satellite sensors that provide high-spatial resolution observation in the thermal infrared range [37]. This parameter was calculated using the following Formula (2):

$$\mathrm{T}\mathrm{s}={\displaystyle \frac{\mathrm{T}\mathrm{b}}{\left(1+\frac{\mathsf{\lambda }\mathrm{T}\mathrm{b}}{\mathsf{\rho }}\right)\times \mathrm{l}\mathrm{n}\mathsf{\epsilon }}}$$

(2)

where Ts is the temperature of the earth’s surface in Kelvin (K), Tb is the brightness temperature in Kelvin (K), λ is the wavelength of the emitted radiance (Table 1), ε is the emissivity of the surface, and ρ is the coefficient, 0.01438 Mk.

From Equations (1) and (2), the NDVI and Ts images of the irrigation period (November to March) for each target year (Figure 4) were obtained.

-

Determination of NDVI and Ts thresholds from seasonal NDVI and Ts profiles derived from the overlay of NDVI and Ts images and irrigated land samples;

NDVI and Ts images obtained for the irrigation period (November to March) were overlaid onto irrigated cropland samples. For each study year, 150, 148, and 158 sampling points were collected, respectively, from reference irrigated sites through the basin (Figure 1) and also from high-resolution images on the Google Earth Pro platform (Figure 3b). The NDVI and Ts values for the monthly NDVI and Ts images were extracted, and then the seasonal averages of NDVI and Ts for the sampling points were deduced. These new NDVI and Ts values for the irrigated areas samples were visualized and fitted using the Tukey outlier detection method proposed by Crawley [38]. This method was applied to the seasonal averages of NDVI and Ts for the target years in order to identify and remove values above Q3 + 1.5 × IQR and below Q1 − 1.5 × IQR. Note that Q1, Q3, and IQR represent the lower quartile, upper quartile, and inter-quartile ranges, respectively. Figure 5 shows the NDVI and Ts variation of irrigated and rainfed areas during the irrigation period. The description of Figure 5a–c confirms the assumption that, in the agricultural area during the dry season, irrigated croplands have high NDVI and low Ts values, unlike rainfed croplands. The analysis of these figures enabled us to determine the NDVI and Ts threshold values for extracting irrigated land for each target year. According to Figure 5, NDVI thresholds of 0.40–0.60, 0.35–0.55, and 0.35–0.55 and Ts thresholds of 302.65–303.15 °K, 302.85–303.25 °K, and 302.50–302.80 °K for the years 2014, 2018, and 2022 were set to identify irrigated land in the study area.

-

Combining the thresholds through a logic equation;

For each year, the irrigated areas were extracted using logical Expressions (3)–(5) developed from NDVI and Ts thresholds.

$${\mathrm{I}\mathrm{r}\mathrm{r}\mathrm{i}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{s}}_{2014}:0.40 \le \mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I} \le 0.60 \mathrm{A}\mathrm{N}\mathrm{D} 302.65°\mathrm{K} \le \mathrm{T}\mathrm{s} \le 303.15°\mathrm{K}$$

(3)

$${\mathrm{I}\mathrm{r}\mathrm{r}\mathrm{i}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{s}}_{2018}:0.35 \le \mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I} \le 0.55 \mathrm{A}\mathrm{N}\mathrm{D} 302.85°\mathrm{K} \le \mathrm{T}\mathrm{s} \le 303.25°\mathrm{K}$$

(4)

$${\mathrm{I}\mathrm{r}\mathrm{r}\mathrm{i}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{s}}_{2022}:0.35\le \mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I} \le 0.55 \mathrm{A}\mathrm{N}\mathrm{D} 302.50°\mathrm{K}\le \mathrm{T}\mathrm{s} \le 302.80°\mathrm{K}$$

(5)

• ▪ Applying the algorithm to NDVI and Ts images of agricultural areas to differentiate irrigated from rainfed land

Logical Expressions (3)–(5) were applied to NDVI and Ts images of agricultural areas to differentiate between irrigated and rain-fed areas. This algorithm was implemented in the Raster Calculator tool of ArcGIS 10.8.1 mapping software.

2.2.4. The Accuracy Assessment

Evaluating the precision of the generated maps is crucial for ascertaining the credibility of the findings. Based on the classification results, accuracy indicators such as the overall accuracy (OA) and the Kappa index (K) [39] of the land cover maps were directly calculated in the GEE using Formulas (6) and (7). These two indicators present the level of consistency between the classification results and the actual land cover of the study area based on the validation points.

$$\mathrm{O}\mathrm{A}={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{q}}{\mathrm{t}}_{ii}}{\mathrm{N}}} \ast 100$$

(6)

where q is the number of classes (04), t_ii is the number of pixels in class i correctly classified in class i, and N is the total number of prediction pixels. OA indicates the proportion of all validation samples correctly mapped during a classification. This metric is typically represented as a percentage (%) and can range from 0% to 100%. While overall accuracy (OA) is straightforward to compute, it alone may not be adequate for validating classification outcomes. To enhance accuracy, the kappa index was computed in this investigation.

$$\mathrm{K}={\displaystyle \frac{{\mathrm{p}}_0-{\mathrm{p}}_{\mathrm{e}}}{1-{\mathrm{p}}_{\mathrm{e}}}}$$

(7)

where ${\mathrm{p}}_0$ is the proportion of validation points correctly classified and ${\mathrm{p}}_{\mathrm{e}}$ is the expected proportion of validation points correctly classified by chance. The Kappa coefficient utilized in this study ranges from 0 to 1, with a value nearing 1 indicating a near-perfect classification. Following the accuracy evaluation, non-farmland land uses were excluded from the final maps to generate the farmland map for each specified year.

As for the validation of the generated irrigated land maps, the overall accuracy (IA), and error of omission (EO) [40] of the algorithm, results were determined through Equations (8) and (9).

$$\mathrm{I}\mathrm{A}={\displaystyle \frac{\mathrm{G}\mathrm{P}\mathrm{C}\mathrm{I}}{\mathrm{G}\mathrm{P}\mathrm{I}}}\times 100$$

(8)

$$\mathrm{E}\mathrm{O}={\displaystyle \frac{\mathrm{G}\mathrm{P}\mathrm{N}}{\mathrm{G}\mathrm{P}\mathrm{I}}}\times 100$$

(9)

where GPCI is the ground-truthed irrigated points classified as irrigated; GPI is the ground-truthed points of irrigated areas; and GPN is the ground-truthed irrigated points falling on non-irrigated class. IA and EO reflect a good prediction of the algorithm used, with values ranging from 0% to 100%.

2.2.5. Evaluation of Changes in Agricultural and Irrigated Areas from 2014 to 2022

To assess the dynamics of irrigated land expansion during the target years, the average annual expansion rate (TA) was calculated using the Bernier’s [41] formula represented by Equation (10).

$$\mathrm{T}\mathrm{A}={\displaystyle \frac{\mathrm{l}\mathrm{n}\mathrm{S}2-\mathrm{l}\mathrm{n}\mathrm{S}1}{\left(\mathrm{t}2-\mathrm{t}1\right)\times \mathrm{l}\mathrm{n}\mathrm{e}}}\times 100$$

(10)

where S1 and S2 are the area of irrigated land at date t1 and t2; ln is the natural logarithm and e is the base of the natural logarithm with e = 2.71828.

3. Results and Discussion

3.1. Distribution of Agricultural Areas for the Periods 2014, 2018, and 2022

Supervised classification of Landsat images using the random forest algorithm provided information on the distribution of agricultural areas in the Ouémé River basin over the study period. Figure 6 shows morphological changes in the distribution of agricultural areas over the years 2014, 2018, and 2022. The analysis of these maps illustrates a very dynamic agricultural practice in the study area, with a high concentration of agricultural land in all regions of the basin. Such observations confirm that the river basin is a zone of agricultural activities [11].

Table 3. Statistics of the agricultural areas for the years 2014, 2018, and 2022.

Year	Agricultural Land		Precision
	Area (km2)	Area (%)	OA (%)	Kappa
2014	16,002.8	32.29	92.77	0.90
2018	19,732.4	39.81	93.85	0.91
2022	22,850	46.1	94.37	0.92

shows the values for the extent of agricultural areas and the precision indicators calculated. The description of the table shows that in 2014, agricultural land covered an area of 16,002.8 km² which represents 32.29% of the total surface area of the basin. These areas have increased to 19,732.4 km² (39.81%) in 2018 and to 22,850 km² (46.1%) in 2022. In terms of spatiotemporal dynamics, these results show that agricultural areas increased significantly over the period of 2014–2022. Similar observations were illustrated by the work of Lawin et al. [32] and Bodjrènou et al. [42]. Using the land use maps produced by the project “Dynamics of Land Use in West Africa”, Lawin et al. [32] found that agricultural land occupied 7.1, 23.1, and 31.4% of the watershed in 1975, 2000, and 2013, respectively. The same results, i.e., an increase in agricultural areas, ranging from 7.50% in 1975 to 22.87% and 31.79% in 2000 and 2013, were obtained by Bodjrènou et al. [42] when they assessed the current and future dynamics of land use units in the Ouémé River basin. The authors emphasized that this progression of agricultural areas has taken place at the expense of vegetation formations such as forests and savannahs. Our results also concur with those found by Annan et al. [43], who reported that agricultural areas increased in size from 1986 to 2023 in the Ouémé River basin. As described in Section 2.2.4, an assessment of the accuracy of the agricultural area maps (Figure 6) was carried out by comparing these maps with randomly generated ground truth points in the study area. The comparison was facilitated by determining the overall accuracy (OA) and Kappa index (Table 3). From Table 3, overall accuracy values of 92.77%, 93.85%, and 94.37% are obtained for the years 2014, 2018, and 2022, respectively, with corresponding Kappa coefficients of 0.9, 0.91, and 0.92. Compared with the threshold value of 85%, these accuracy values (OA and K) obtained in this study are high and reflect a good cartographic accuracy of our results [20,44]. These high accuracy values mirror the results of the study conducted by Annan et al. [43], who reported overall accuracy values above 90%. According to the authors, such values reflect agreement between land use maps and validation samples collected at reference sites. Although the calculated indicator values are higher than those obtained by Magidi et al. [20] and Pareeth et al. [44] in their respective studies, the accuracy of our final maps can be refined using the post-classification improvement methods proposed by Paredes-Gómez et al. [45] and Nhamo et al. [46].

3.2. Distribution of Irrigated Land in 2014, 2018, and 2022

This part focuses on the evolution of irrigated areas in the study area.

Table 4. Characteristics of irrigated croplands in 2014, 2018, and 2022.

Year	Irrigated Croplands		Precision
	Area (km2)	(% of Agricultural Area)	IA (%)	EO (%)
2014	754.94	4.72	79.21	20.79
2018	1136.27	5.76	81.72	18.28
2022	1882.64	8.24	83.15	16.85

shows the total area of irrigated land for the three years, as well as accuracy indicators. According to the table’s analysis, the total area of irrigated land in the Ouémé River basin is estimated at 754.94 km², 1143.23 km², and 1882.64 km² for the years 2014, 2018, and 2022, respectively. These areas represent 4.72%, 5.76%, and 8.24% of the total agricultural areas for the corresponding years. These findings confirm the predominance of rainfed agriculture in the Ouémé River basin [2,11]. A similar trend was reported by the results of the study conducted by Chandrasekharan et al. [24] to assess the extent of irrigated land in Ethiopia. The authors showed that irrigated areas accounted for only 6% of the total cultivated land area during 2015–2016. Magidi et al. [20] noted an opposite trend in the Mpumalanga province. Their study revealed that irrigated areas accounted for 65.88% and 75.01% of cultivated land in 2019 and 2020, respectively, indicating high water use for agricultural production in this region. These statistics also indicate that, although irrigated areas are relatively small compared with rainfed agricultural areas, significant increases have been noted in the study area throughout the study period. However, as shown in Figure 5, this increase was not seen in all regions of the basin (Figure 7). For example, the northern parts experienced a regression in irrigated land areas during the period 2014–2022 in contrast to the central and southern regions, where irrigation activities were intensified with the expansion of irrigated land areas. This situation can be explained by the new orientations that have intervened in Benin’s agricultural sector since 2016 with the creation of agricultural development poles (PDA) [47]. Through these PDA, each part of the basin has been identified for the promotion of a given agricultural commodity chain. The southern and middle part, covering the country’s agricultural development poles 5 and 6, is one of the areas identified for the promotion of sectors including rice growing. As rice is an irrigated crop at the center of the basin’s hydro-agricultural development policies [9], its production has required the multiplication of structuring hydro-agricultural investments, leading to the extension and concentration of irrigated croplands observed in the southern and central parts of the basin in 2018 and 2022. It is also known that the southern and central parts of the basin are home to some of the country’s benchmark irrigation sites, namely the Koussin-Lélé, Kaffa, and Bamè perimeters and the Complexe Sucrière du Bénin (SUCOBE) [3,8]. The assessment of the accuracy of the final irrigated land maps (Figure 8) indicated overall accuracy (IA) values of 79.21%, 81.72%, 83.15%, respectively, for the years 2014, 2018, and 2022, with corresponding errors of omission (EO) values of 20.79%, 18.28%, and 16.85% (Table 4). The overall accuracy (IA) values are lower than 99% and 98%, which were the accuracy values found by Wu and De Pauw [27] and Kant and Mishra [29] in their respective studies. However, in an area as complex as the Ouémé River basin, the overall accuracies of 79.21%, 81.72%, and 83.15% obtained in this study seem entirely relevant for the analysis of irrigated land expansion dynamics. In addition, these values are higher than those obtained by Lamhamedi et al. [31] and Ghebreamlak et al. [7] when they applied the algorithm to identify irrigated cropland in the region. The low relative precision of their results may be due to the defined NDVI and Ts thresholds. For this reason, Pervez et al. [28] reported that the success of this method largely depends on the thresholding process used.

3.3. Changes in the Area of Agricultural Land in the ORB from 2014 to 2022

Table 5. Change in agricultural and irrigated land areas between 2014 and 2022.

Periods	2014–2018		2018–2022
Features	Area (km2)	AE (%)	Area (km2)	AE (%)
Agricultural land	3729.6	5.24	3117.6	3.67
Irrigated land	381.33	10.22	746.37	12.62

shows the evolution of agricultural land in the Ouémé River basin over the periods of 2014–2018 and 2018–2022. The analysis of this table revealed that the area of agricultural land increased by 3729.61 km² between 2014 and 2018 and by 3117.6 km² between 2018 and 2022. The annual expansion rate (AE) of agricultural land during these two study periods was estimated at 5.24% and 3.67%, respectively. Similar trends were observed through the results of previous studies carried out on land occupation in the Ouémé River basin [32,42,43]. Indeed, each of these studies highlighted a strong expansion of agricultural areas, i.e., a strong transformation of agricultural land when moving from one period to another. In addition, Annan et al. [43] noted that this increase in agricultural land was due to strong anthropogenic pressure on protected areas, savannahs, and forests. Alongside agricultural land, irrigated land also increased between the 2014–2018 and 2018–2022 periods (Table 5). The respective area expansions for each period were quantified at 381.33 km² and 746.37 km², resulting in annual expansion rates (AEs) of 10.22% and 12.62%, respectively. These values show a strong transformation of irrigated areas that can be attributed to the new strategic orientations operated in the irrigation sector in Benin, namely the implementation of the National Irrigation Development Program (PND Irr) and the establishment of the Territorial Agencies for Agricultural Development (ATDA) in 2018 for the valorization of hydro-agricultural potential. Our results concur with those found by Traoré et al. [26], Pervez et al. [28], Alexandridis et al. [12], and Zhang et al. [48] in their respective studies. For example, in the Mogtedo region of Burkina Faso, the study by Traoré et al. [26] revealed a 54% increase in irrigated land area between 1987 and 2015. Similarly, the study by Zhang et al. [48] highlighted a substantial 24.8% increase in the area of irrigated cropland in China during the period of 2000–2019.

3.4. Factors Contributing to the Expansion of Irrigated Areas

The expansion of agricultural areas in general and of irrigated areas in particular may be due to several factors that are both environmental and anthropogenic. Annan et al. [43] reported that population growth, access to water sources, and soil fertility are factors behind the expansion of agricultural areas in the Ouémé River basin. Indeed, land close to water sources is used more for agricultural production in the dry season. Climatic factors such as rainfall can also affect the expansion of agricultural areas in the Ouémé River basin. The results obtained in Section 3.2 show that the majority of agricultural land is exploited during the rainy season. Delayed rainfall and pockets of drought adversely affect the seasonal dynamics of agricultural areas, especially rainfed areas. Furthermore, in the southern part of the basin, Osseni et al. [49] found that soil type and the proximity of the road network affected the vegetation cover of the area while at the same time leading to an expansion of irrigated areas. According to the authors, the development of road infrastructure, especially secondary roads, will facilitate access to fertile land, the preferred area for agricultural activities. The same factors were observed by Magidi et al. [20] when they asserted that factors such as slope, hydrographic network, and distance from inhabitants affect the distribution of cultivated land in the Mpumalanga region of South Africa. All of these findings confirm the results of the study by Jellason et al. [50]. According to their study, population growth and government policies are behind the expansion of agricultural land. The authors add that other factors, notably the development of agricultural mechanization, level of land fertility, promotion of agricultural sectors such as cash crops, and climate change, can either positively or negatively influence the expansion of agricultural areas in sub-Saharan Africa. Specifically, Zhang et al. [48] found that population growth, extension of agricultural land, and investment in water infrastructure have contributed to the expansion of irrigated land in China. The same findings were obtained in the Ouémé River basin. According to data from the General Census of Population and Housing (RGPH) conducted in 2002 and 2013, the population of the Ouémé River basin surged by 51.3%, equivalent to 986,785 inhabitants [51]. Extrapolating this trend to the period from 2014 to 2022 yields a projected population of 3,021,878 inhabitants in 2014 and 4,083,895 inhabitants in 2022, representing a 35.14% increase. This population growth drives increased demand for staple foods like rice and market garden produce, which are primarily cultivated on irrigated lands [3,8]. Additionally, agricultural land has expanded significantly between the periods of 2014–2018 and 2018–2022. For instance, our results indicate a 23.31% increase in the area in the basin during 2014–2018 and a further 15.80% increase during 2018–2022. This expansion in agricultural land has implications for the expansion of irrigated areas in the basin. To quantitatively demonstrate the impact of increased agricultural land on the expansion of irrigation, an analysis of the evolving trends in agricultural land acreage and irrigation from 2014 to 2022 is presented in Figure 8. The figure illustrates an increase in irrigated land areas, reflecting the ongoing development of agricultural land. Other contributing factors to the expansion of irrigated land in the Ouémé River basin include the development of small-scale private irrigation schemes, the establishment of producer cooperatives for managing large irrigation sites, and new agricultural land development initiatives outlined in the MPWDM and government action programs (GAP) [5,9,11].

3.5. Advantages and Limitations of the Methodology Employed

Given the difficulty of identifying pixels of irrigated areas, many studies have shown that the use of high-spatial resolution satellite data such as Landsat images is desirable for mapping and quantifying irrigated cropland [6,20]. In this study, Landsat image composites were employed to determine Ts and NDVI images as it was difficult to obtain a single image of the entire study area. This difficulty guided the choice of the years 2014, 2018, and 2022, as they are years for which the available imagery made it possible to obtain NDVI and Ts images for the irrigation period with less cloud cover. This study leveraged the capabilities of the GEE platform, chosen for its capacity to efficiently handle and process substantial datasets within compressed time frames [20]. In this research, the RF algorithm was used to separate agricultural land from other land uses. The results indicated very good cartographic accuracy and confirmed the studies by Traoré et al. [26] and Kombate et al. [35] on the performance of the RF algorithm. In addition, it integrates easily with big data platforms such as the GEE for processing large geospatial datasets in a short time [36]. Despite the multitude of existing high-resolution Landsat images in the GEE, some regions of our study area do not have quality images to conduct this research. Given this situation, pixels from several images with large temporal differences in acquisition dates were used to form the composites.

Implementing the algorithm utilizing NDVI and Ts thresholds for delineating irrigated regions yielded superior outcomes across the various years examined in this study. Previous research has indicated that this methodology is particularly effective in areas where irrigation schedules and crop calendars are clearly defined [6,29]. Although the addition of surface temperature (Ts) has been considered promising for extracting irrigated land, some studies have relied solely on the NDVI threshold [19,20]. The selection of NDVI and Ts threshold values in this study was based on sampling points collected on irrigated sites. This empirical method has been used in numerous studies [27,28] with good findings. Although NDVI and Ts thresholds were effective in differentiating irrigated areas, it is important to acknowledge that this method may introduce certain biases. According to Ghebreamlak et al. [7], the quality of the samples used to define the NDVI and Ts threshold values affects the results of the algorithm. For this reason, it is preferable to focus on collecting samples of irrigated land when defining thresholds. In addition, the absence of national statistics on the evolution of irrigated land areas limited the validation of the final maps obtained. The same observation was made by Pervez et al. [28] and Kant and Mishra [29] in their respective studies.

4. Conclusions

This study devised an algorithm combining thresholds of vegetation index (NDVI) and surface temperature (Ts) to delineate the spatiotemporal distribution of irrigated cropland in the Ouémé River basin. Prior to implementing the algorithm, a preliminary land use classification was conducted using the random forest (RF) classifier to outline agricultural lands. The classification outcomes indicated agricultural areas of 16,002.8 km², 19,732.4 km², and 22,850 km² for the years 2014, 2018, and 2022, respectively, reflecting a gain of 3729.61 km² between 2014 and 2018 and 3117.6 km² between 2018 and 2022. The Kappa values obtained, ranging from 0.9 to 0.92, demonstrated good accuracy for the respective maps. Moreover, the application of the algorithm based on a combination of thresholds for surface temperature (Ts) and vegetation indices (NDVI) exhibited superior cartographic performance, yielding accuracy values of 79.21%, 81.75%, and 83.15% for the years 2014, 2018, and 2022, respectively. The algorithm results revealed irrigated areas of 754.94 km², 1143.23 km², and 1882.64 km² for the years 2014, 2018, and 2022, respectively, with an increase of 381.33 km² and 746.37 km² during the periods of 2014–2018 and 2018–2022. The average annual expansion rate of irrigated areas for each period was estimated at 10.22% and 12.62%, respectively. This expansion was observed across various regions of the basin. Although the area of irrigated cropland remains relatively small compared with agricultural land in general, the increase in the basin’s population and land development facilitated by the proliferation of hydraulic infrastructures have contributed to an increase in irrigated areas. This spatial and temporal information on irrigated croplands would offer valuable insights into the expansion dynamics of irrigated areas in the Ouémé River basin, aiding water managers in guiding hydro-agricultural investments for the sustainable and efficient utilization of water resources in the agricultural sector.