A Survey of Changes in Grasslands within the Tonle Sap Lake Landscape from 2004 to 2023

Abstract

The Tonle Sap Lake (TSL) landscape is a region of vast natural resources and biological diversity in the heart of Southeast Asia. In addition to serving as the foundation for a highly productive fisheries system, this landscape is home to numerous globally threatened species. Despite decades of recognition by several government and international agencies and the fact that nine protected areas have been established within this region, natural land cover such as grasslands have experienced considerable decline since the turn of the century. This project used local expert knowledge to train and validate a random forest supervised classification of Landsat satellite imagery using Google Earth Engine. The time series of thematic maps were then used to quantify the conversion of grasslands to croplands between 2004 and 2023. The classification encompassed a 10 km buffer surrounding the landscape, an area of nearly 3 million hectares. The average overall accuracy for these thematic maps was 82.5% (78.5–87.9%), with grasslands averaging 76.1% user’s accuracy. The change detection indicated that over 207,281 ha of grasslands were lost over this period (>59.5% of the 2004 area), with approx. 89.3% of this loss being attributed to cropland expansion. The results of this project will inform conservation efforts focused on local-scale planning and the management of commercial agriculture.

1. Introduction

The Tonle Sap Lake (TSL) floodplains and surrounding landscape are a region globally recognized for biodiversity and natural productivity [1,2,3,4]. Millions of people throughout the Lower Mekong River Basin rely on TSL’s fisheries, water resources, and natural vegetation for their livelihoods [5,6]. Recognizing the importance of this landscape, UNESCO created the Tonle Sap Biosphere Reserve (TSBR) in 1997, which was further established by governmental royal decree in 2001 [3,5,7]. The TSBR is a major breeding area for at least a dozen globally threatened bird species [3,6,7]. These include the Lesser Adjutant (Leptoptilos javanicus), Painted Stork (Mycteria leucocephala), Asian Openbill (Anastomus oscitans), Black-headed Ibis (Threskiornis melanocephalus), Oriental Darter (Anhinga melanogaster), Sarus Crane (Grus antigone), and Bengal Florican (Houbaropsis bengalensis). These bird species, along with several environmental functions of this flood pulse ecosystem, are dependent on the mosaic of grasslands and other vegetation found throughout the floodplains. Despite the efforts of the Cambodian Ministry of Environment (MoE) and agencies such as the Wildlife Conservation Society (WCS) and BirdLife International, this landscape and its grasslands are facing ongoing degradation due to commercial agricultural development [5,7,8,9,10]. Agricultural development threatens countless species, with the true extent of these impacts being uncertain due to limited ecosystem assessments.

Grasslands throughout the TSL landscape provide a variety of ecosystem services. Among these services are carbon storage and sequestration [11,12,13], water nutrient regulation, livestock food provisioning [14], and the provision of breeding grounds and habitat for wildlife [2,3,9,13]. Long-term conservation efforts for these grasslands across the protected landscapes are aimed at the unique and globally threatened community of bird species found there. Historically, information on these bird populations has been gathered through assessments, made from field surveys, such as those documented by Seng et al. [15], van Zalinge [16], and Packman [6]. These surveys, however, are resource-intensive and lack the ability to provide comprehensive information on the landscape on a time scale relevant to management action. Jointly, monitoring the expansion of croplands, such as rice, into these land cover could lead to vast increases in knowledge on regional environmental status, food security, and local socioeconomics [17,18]. The integration of remote sensing tools and geographic information systems (GIS) could lessen the resource requirements for conducting such surveys, while at the same time providing accurate and up-to-date information on land cover dynamics.

The use of remote sensing for earth observation (EO) has expanded in recent decades to become an established and increasingly impactful method for obtaining information on land cover and land use (LCLU) [6,19,20,21]. Land cover acts as a key component of environmental monitoring and modelling for many projects [19,22,23]. Time series data, from programs such as Landsat, support landscape level analyses of land cover which would be cost-prohibitive if conducted in the field (i.e., in situ). A widely used application of land cover products derived from remotely sensed imagery is the detection of LCLU change [19,24,25]. Change detections provide crucial, site-specific information on the spatial distribution and abundance of land cover classes over time [21,26,27]. Such information can then be used to track the impacts of human development (such as cropland expansion), social pressures, or policy decisions by a wide range of stakeholders [28]. While these methods are increasingly available, they have not been applied ubiquitously [1]. Grasslands have historically been underrepresented in remote sensing analyses [6,12]. Few studies published over the last 20 years for this region specifically have reflected on changes in grassland abundance or distribution, which is a vital component for accurate and timely conservation action [9,24]. Sourn et al. [29] demonstrated a large change from forests to croplands in the nearby Battambang province between 1998 and 2018. The authors noted that drivers of this deforestation were likely policy, legal frameworks, and socio-economic pressures [29]. Mahood et al. [13] detailed that rice cultivation was likely the cause of substantial grassland and shrubland loss between 1993 and 2018 within the TSL and surrounding floodplains. Other studies of the early 2000s also portrayed large amounts of cropland expansion at the cost of natural vegetation cover throughout this region [5,10,30]. Conservation organizations working in the TSL landscape are facing limited resources and knowledge on which to base their decisions due to the age of their data on remanent grassland patches [2,4]. Measuring the current area and distribution of natural land cover types would aid the management of this ecosystem [2]. Our study aims to inform local conservation efforts, as well as provide an analytical framework for similar investigations of natural vegetation loss in neighbouring regions. This research focuses on the full TSL landscape, including a 10km buffer from the WCS-defined landscape (‘stronghold’) boundary. Within this area are nine protected areas, which face ongoing influence from a wide range of stakeholders [1,5,31,32]. This research also specifically addresses the loss of dry grasslands, which are a critical component of the local ecology. For these reasons, our objectives are to quantify losses in grasslands within the TSL landscape since the early 2000s. More specifically, this research aims to undertake the following:

• Quantify trends in land cover change leading to losses in grasslands throughout the Tonle Sap Lake (TSL) landscape.

  a.

  Evaluate these changes in the context of dry grasslands, a vital wildlife habitat within this region.

This analysis will help conservation efforts to specifically address losses at the landscape and key protected area scales.

2. Materials and Methods

2.1. Study Area

The TSL and its surrounding floodplains are a key component of the Lower Mekong River Basin [1,13]. The core area is defined by the TSBR, which serves as a research site designated to protect biodiversity and limit habitat disturbance [7,13]. The area of open water within this flood pulse system has been measured to expand as much as five times between the dry and wet seasons [3]. Intermixed with flooded forests, spreading outwards from the open waters are a mosaic of grassland patches and scrub/shrublands [3,4,7], leading into croplands. To ensure that the entire landscape was considered in this analysis, a 10 km buffer was applied to the TSBR area (Figure 1). This 2.99-million-hectare (ha) area, in black, represents the primary extent of this analysis.

Nine protected areas are located throughout the TSBR (Figure 1). These protected areas make up approximately 170,056 ha within the study region (~5% of the total area) [33]. The protected areas range in size from 288 ha (Phnom Krang Dey Meas) to over 38,000 ha (Bakan). These areas have each been established for the conservation of critically endangered species, such as the Bengal Florican, the Sarus Crane, and the vulnerable Manchurian Reed Warbler (Acrocephalus tangorum) [9,34,35]. The updated boundary files for these protected landscapes were downloaded from Protected Planet [33].

2.2. Satellite Imagery Selection and Pre-Processing

A combination of Landsat 5 Thematic Mapper (TM) and Landsat 8 Operational Land Imager (OLI) Level 2 (surface reflectance corrected) imagery were used to generate landcover maps of this region throughout the study period [36,37]. Landsat images are widely established for land cover mapping and time series analyses, even for grassland areas [11,14,19,28]. Google Earth Engine (GEE) (Mountain View, CA, USA) was used to compile Landsat imagery, conduct the pre-processing, and perform the classification. This cloud-based GIS greatly reduced the burden of computational complexity for this analysis [18,28,38]. Images of the dry seasons from 2004, 2008, 2013, 2018, and 2023 were queried. Dry season images were defined as those taken between November of the previous year and early April of the defined year. For example, the 2004 dry season would be defined as 1 November 2003, through to 1 April 2004. This period was selectively sampled due to the consistent cloud and shadow coverage throughout the rest of the year [29]. The only other filter applied to the image collections was to restrict the collections to the Tonle Sap 10 km buffer area. A more consistent 5-year time step was desired for this analysis; however, imagery from the 2003 dry season were of poor quality with considerable cloud coverage.

For each of the five years mapped in this study, the image collections were comprised of roughly 30–100 individual Landsat images (i.e., scenes). More specifically, the image collections were refined to 104 (2004, Landsat 5), 65 (2008, Landsat 8), 54 (2014, Landsat 8), 43 (2018, Landsat 8), and 28 (2023, Landsat 8) images based on date and boundary criteria [36,37]. The specific Path/Row information for each image collection can be viewed in Appendix A. The pre-processing of the surface reflectance products consisted of two functions. First, both Landsat 5 and Landsat 8 images were scaled based on the default scaling factors on their respective GEE download pages [36,37]. Second, a pixel-based composite was made from each collection of images. These composites are a common technique for creating a singular image (mosaic) of a study region with reduced noise, from which image analysis and classification can be performed [19,39,40,41,42]. The specific technique for creating these composites was based on the median pixel algorithm. This algorithm returns the median value for each overlapping pixel for each band across an image stack, and has been found to generate a more radiometrically consistent result [43,44,45]. More information on this process can be found in the article by Francini et al. [43] and the Google Earth Engine Developer pages [46].

2.3. Reference Data

Three sources of reference data were used to support the training and validation of the 2004, 2008, 2014, 2018, and 2023 thematic maps. These sources of reference data were integrated because they most closely matched the land cover dynamics of the studied years, were based on reliable methods, and ensured that the earliest maps (2004 and 2008) were comparable to historic surveys of grasslands within this geography. First, historic habitat data were used to guide the 2004 and 2008 classifications. This habitat data consisted of 1250 grassland samples, 1240 shrubland/forest samples, and 780 cropland samples generated through the interpretation of high-resolution imagery (1:25,000), which was collected in 2005 and 2006 [47,48]. These data were also used to establish estimates of grassland habitat from earlier surveys [6]. Second, the WCS Cambodia GIS team provided maps of the landscape digitized from high-resolution 2021 and 2022 imagery. These maps were used to generate samples from homogeneous areas of shrublands (n = 1200), forests (n = 700), grasslands (n = 1200), and rice croplands (n = 1000). These samples were used to guide the classification of the 2014, 2018, and 2023 imagery. Both sets of reference points were manually reviewed by both the GIS technician and landscape experts. Lastly, the WCS Cambodia team independently conducted an analysis of changes in grasslands within the Bakan and Ang Trapeng Thmor (ATT) protected areas. From this analysis, 75 reference samples per class were integrated into the classification of the larger landscape. Samples for the class ‘water’ were generated through the interpretation of both the high-resolution basemap imagery and the respective Landsat composite.

The previously established sources of reference data were used by a trained image interpreter as a guide for manually entering reference samples using the GEE geometry tools. For each of the five classifications, 900 to 1300 reference samples were entered (

Table 1. Total reference data sample sets used to train (i.e., classify) and validate each of the five land cover maps.

Reference Data (Total: Training and Validation)
	2004	2008	2014	2018	2023
Grassland	335	225	225	205	285
Forest/Shrub	375	285	355	375	365
Cropland	255	225	315	295	285
Water	225	215	235	315	345
Village/Road	Cambodia Roads Mask
Total	1190	950	1130	1190	1280

). These samples were geographically distributed with a minimum of 100 samples per class per map, and were based on a higher level of detail than the Landsat imagery provided [27,49].

2.4. Land Cover Classes

Land cover classes within the TSL landscape were distinguished based on definitions adapted from the European Space Agency (ESA) WorldCover project classification scheme [50]. The classes defined here included grassland, forest/shrub, cropland, water, and village/road. The full definition for each class is provided below in

Table 2. Land cover class definitions. Adapted from European Space Agency (ESA) WorldCover project classification scheme [50].

Land Cover Class	Definition
Grassland	Any area dominated by >10% herbaceous plants (i.e., those with a persistent stem but lacking woody/firm structure). These may include grasslands, prairies, pastures, or savannahs. Woody plants, such as trees and shrubs, may be present with coverages < 10%. Abandoned croplands are also included in the class if herbaceous coverage is >10%.
Forest/Shrub	Any area dominated by trees or shrubs with a combined coverage >10%. Other land cover classes such as grasslands, croplands, or water may be present beneath the tree canopy. Areas planted for commercial agriculture (including rubber plantations) are not included in this class. This class does include seasonally flooded areas.
Cropland	Any area covered by planted/sowed crops. These croplands may consist of herbaceous or woody crops, including paddy rice, rubber, cashew, cassava, or a mixture of other crops. Croplands may be irrigated or rainfed within this region.
Water	Any area dominated by open water during the majority of the dry and wet seasons. These may include lakes, reservoirs, or rivers.
Village/Road	Any area covered by buildings, roads, or other artificial (i.e., built-up) structures.

. Additional attention was given when defining the characteristics of the grassland, cropland, and village/road classes due to their influence on the analysis. The definition for grassland, which is also consistent with Sourn et al. [29], is marked by the absence (minimal coverage) of trees or shrubs [12]. Croplands are understood to be one of the most abundant land cover types throughout Cambodia [31]. While many crop types could exist within these areas, previous studies have reported that paddy rice makes up more than 70% of the croplands within the TSL floodplains [13,31]. Lastly, the village/road class was important to distinguish due to its considerable distribution throughout the study region. Scattered buildings and unpaved roads, however, were difficult to distinguish from barren areas or dry-season croplands due to the limited spatial resolution found with the Landsat imagery [51]. For this reason, we used a Cambodia road layer (updated in 2013) as a mask to serve these areas, rather than including this class within the supervised classification [52]. The road layer lines were buffered 10 m to cover the average road diameter. The road layer mask was integrated after the GEE classification.

To further define dry from wet grassland communities across the floodplains, the European Union Joint Research Centre (JRC) Global Surface Water layer (v4) was used [53]. Specifically, the (monthly) occurrence frequency of surface water presence estimated from 1984 to 2021 was used to define a threshold between these vegetation communities. An occurrence value greater than or equal to 20% (≥20%) defined wetland areas (Figure 2). Grassland areas outside of this polygon were defined as dry grasslands.

2.5. Land Cover Classification and Change Analysis

The image mosaics for the full study region, Bakan, and ATT were independently classified in GEE using a random forest supervised classification [14,54]. These classifications resulted in three separate land cover maps for 2004, 2008, 2014, 2018, and 2023. The random forest algorithm, like other machine learning methods, is documented to handle complex classes and a high dimensionality of input data better than conventional classifiers [19,55]. The random forest algorithm is also commonly used for paddy rice mapping which is known to be in high abundance throughout this region [17]. In addition to the original Landsat bands, seven spectral indices and Shuttle Radar Topography Mission (SRTM) slope data were used in the classification [56]. The goal of this combination of data was to maximize the spectral uniqueness of each land cover class [20,21,38]. The spectral indices included the Normalized Difference Vegetation Index (NDVI) [12,21,57], the Soil Adjusted Vegetation Index (SAVI) [14,58], the Enhanced Vegetation Index (EVI) [38,57], the Greenness Index (GI), the Moisture Stress Index (MSI) [59], the Normalized Difference Water Index (NDWI) [59], and Normalized Burn Ratio (NBR) [60,61]. The SRTM elevation layer was used to improve the differentiation of flat croplands and upland forests [56,62].

Several hyperparameters of the random forest classification algorithm were tuned during the initial testing of the results. These hyperparameters included the number of decision trees, the proportion of training and validation data, and the total selection of input features. The number of decision trees was kept at 500, which is documented to be a basic minimum for increasing the consistency of results [54]. The reference data were split, with a random 55% of the samples per class being used for training and 45% being used for validation (i.e., testing). These proportions ensured that a minimum valid sample size for each class was available for independently conducting the accuracy assessments [49]. Lastly, the selection of input features was determined through a relative feature importance calculation [38,63]: the feature importance test. The results of the relative feature importance test, in combination with the impacts on overall accuracy when removing specific features, were used to achieve the best classification performance [38,64,65,66].

Following the land cover classifications in GEE, each map was exported using Google Drive and opened using ArcGIS Pro (v3.1, Redlands, CA, USA). Minor editing of misclassified areas was performed on each map by local experts to improve the quality of the results and the subsequent change-detection analysis [27]. The maps were converted to shapefiles to perform these edits. These edits included first running the eliminate tool on each map. The eliminate tool was used to dissolve any isolated singular-pixel-sized polygon into the neighbouring polygon of the largest size. For example, isolated polygons less than 0.02 ha in size classified as water but located in the middle of a large cropland area were dissolved into the surrounding croplands. Running the eliminate tool reduced the noise commonly found when using pixel-based classifications [67,68]. Secondly, while reviewing the land cover maps, minor manual revisions were made when areas were recognized as being misclassified and the correct class could be confirmed by a combination of interpretating the corresponding image mosaic and the available reference data.

The final land cover maps for each of the three sites were then merged and used to conduct a post-classification change detection. Post classification change detections have historically been among the most applied change-detection methods, providing a straightforward approach to changes over time [24,26,57]. The ArcGIS Pro categorical change-detection tool (‘compute change detection’) was used so that the spatial distribution and specific types of categorical change could be evaluated [21,24]. We specifically selected changes from grasslands to croplands for each of the four-time steps: (1) 2004 to 2008, (2) 2008 to 2014, (3) 2014 to 2018, and (4) 2018 to 2023. Changes from forest/shrub to cropland were independently assessed to compare our model’s results with those of previous studies [5,13].

2.6. Accuracy Assessment

The accuracy of each land cover map was independently assessed in GEE using a thematic map accuracy assessment error matrix [49,69]. In total, 15 error matrices were produced, which provided quantifications of the overall accuracy (OA), producer’s accuracy (PA), and user’s accuracy (UA) for each map. The final map depicting changes from grasslands to croplands over time (2004 to 2023) was assessed using a change-detection error matrix calculated from an ArcGIS Pro accuracy assessment [24,26,49,70,71]. For this assessment, there was one ‘change’ class (‘grasslands to croplands’) and two ‘no change’ classes (‘grasslands’ and ‘croplands’). One hundred samples were randomly distributed within each stratum (i.e., class) to perform this assessment. The validation of these samples was performed using the 2004 Landsat image mosaic, 2023 Landsat image mosaic, and ArcGIS Pro high-resolution basemap imagery for image interpretation.

The map and accuracy assessment produced from the change-detection analysis were also subsequently evaluated based on the recommendations of Olofsson et al. [27,72]. Using the sample-based accuracy assessment performed in ArcGIS Pro, the area-weighted accuracy and uncertainty were used to generate an area-based error matrix and corresponding confidence intervals for an adjusted estimate of the amount of changed and unchanged areas. Confidence intervals (95%) were calculated for these adjusted areas [27,72]. A graphical representation of the full methodology can be found below in Figure 3.

3. Results

3.1. Land Cover Classification and Change-Detection Analysis

The five thematic maps produced for the full study region achieved an average overall accuracy of 83.5%. The classified maps for the 2004 and 2023 time period are displayed in Figure 4 and Figure 5. The average class-specific user’s accuracies were 88.5% for forests, 76.1% for grasslands, 82.5% for croplands, and 90.8% for water. The overall accuracy, user’s accuracy (UA), and producer’s accuracy (PA) for each class are reported in

Table 3. Accuracy for mapping each of the land cover classes throughout the TSL landscape for each of the five thematic maps. Reported are the overall accuracy of the maps, the user’s accuracy (UA), and the producer’s accuracy (PA) for the respective land cover classes.

	2004	2008	2014	2018	2023
Overall Accuracy	88.5%	86.7%	87.9%	78.5%	79.8%
Grassland UA	84.6%	88.2%	78%	84.6%	90%
Grassland PA	83.9%	68.2%	72.2%	59.5%	78.3%
Cropland UA	92.2%	71.9%	93.3%	84.6%	31.7%
Cropland PA	87.7%	86.8%	87.4%	84.6%	48.1%
Forest/Shrub UA	86.4%	96.5%	89.4%	82.6%	81.3%
Forest/Shrub PA	91.2%	94.3%	91.0%	76.9%	75.7%
Water UA	97.9%	87.3%	85.2%	91.4%	98.6%
Water PA	93.9%	96.5%	98.1%	85.5%	98.6%

. Confusion between the forest and cropland classes outside of the floodplain decreased the accuracy for both classes for the 2023 classification.

The change-detection analysis reported that throughout the 2004 to 2023 study period, a total of 207,281.4 ha of grassland area (both wet and dry types) was lost. This loss reflects a 59.5% decrease in grassland area since 2004 (approximately a 3.13% annual decline). This rate, however, was not consistent. Between 2014 and 2018, the rate of decline was 6.7%, while between 2008 and 2014, the rate of decline was 22.0%. The average 4-to-5-year loss was 14.9%.

In 2004, dry grasslands represented 297,004.7 ha out of the 348,300 ha of total grassland present within the study area (85.3% of the total area). The estimated loss of dry grassland over the study period was 174,400.2 ha (Figure 6). This is an estimated decline in area of 58.7% since 2004, a rate of 3.09% annually (

Table 4. Annual grassland cover estimates for the study area based on the classified land cover maps.

Dry Grassland Loss
Year	Hectares	Percent Decrease
2004	297,004.71	Baseline
2008	253,596.98	15%
2014	192,452.50	20.59%
2018	171,072.58	7.20%
2023	122,604.53	16.32%
Total Decrease (2004–2023)	174,400.18	58.7%

).

Within the nine protected areas, the cropland area increased by 37.1% over the study period (approx. 36,400 ha to 49,900 ha). Dry grasslands within these same areas experienced a decline in area of 45.4% between 2004 and 2023. Within the largest protected area, Bakan, the estimated loss of dry grassland was 26.6% (approx. 2806.1 ha of the 10,531.7 ha of dry grassland present in 2004). This same protected area has experienced an increase in cropland land cover of 137.3% since 2004.

The total increase in cropland land cover within the study region was 347,059 ha, a 27.3% increase in area since 2004. The increase in cropland area accounts for 55.7% of the losses in natural vegetation (grasslands, forests, and shrubs) from 2004 to 2023. For dry grasslands specifically, 89.3% of the decline in land area can be accounted for by expansions in croplands (155,725.8 ha of the total 174,400.18 loss).

3.2. Accuracy Assessment

An area-weighted and error-adjusted assessment of the change-detection analysis is reported in

Table 5. Area-weighted accuracy assessment and error-adjusted land cover area estimates for the change-detection analysis based on Olofsson et al. [27,72].

Class	Map Area (Hectares)	Adjusted Area (Hectares)	95% CI (Hectares)	User’s Accuracy	Producer’s Accuracy	Overall Accuracy
Change	174,400.18	216,965.70	20,102.80	69.0%	55.5%	83.89%
Dry Grasslands	122,604.53	248,575.82	18,989.90	64.0%	31.6%
Croplands	1,618,810.00	1,450,273.18	18,308.45	87.0%	97.1%

. This analysis of error and uncertainty for the change (dry grasslands converted to croplands) and no change (unchanged dry grasslands and unchanged croplands) classes shows that the total amount of grassland lost between 2004 and 2023, as reported by the maps, may be underrepresented. Instead of 174,400 ha of loss, the error-adjusted area depicts a total loss of 216,965.7 ha (a 24.4% higher estimate). The 95% confidence interval for this estimated loss is 196,900 ha to 237,100 ha (216,965.7 ha ± 20,103 ha). This area-based accuracy assessment achieved an overall accuracy of 82.89%.

4. Discussion

The results of this analysis, like earlier studies, demonstrate a drastic decline in natural land cover throughout the TSL landscape due to cropland expansion. Kummu et al. [2] made a call for improving knowledge of this ecosystem through surveys of natural land cover types and resources nearly 20 years ago; yet, such knowledge is still critically limited. Reliable information specifically on grassland ecosystems, such as their distribution and quality, is important for successful conservation [6,14,38]. To address the study objectives and support the management of the critically endangered wildlife habitat within the TSL landscape, losses of both grassland and dry grassland were evaluated between 2004 and 2023. The post-classification change-detection analysis reported declines in grassland and dry grassland of 207,281 ha (59.5%) and 174,400.2 ha (58.7%), respectively. The five thematic maps used for this analysis, which achieved an average overall accuracy of 83.5%, depicted a 3.09% annual decline in dry grassland area. The area-weighted and error-adjusted change-detection analysis reported that the actual decline in dry grassland area is closer to 216,965.7 ha. The 95% confidence interval on this calculation showed a decline of 66% to 80% between 2004 and 2023 [27,72]. Within the nine protected areas, 45.4% of dry grasslands were lost between 2004 and 2023. Based on these estimates, the 5% of land currently designated as protected within this region showed a 13.3% lower decline in dry grasslands than the study area average. This 13.3% lower rate of decline is notable, and demonstrates that further conservation action is needed within the TSL landscape, even within protected areas.

The loss of grasslands, especially tropical grasslands, is substantial across Southeast Asia, and represents the loss of a highly valuable ecosystem [6]. Land cover change has a long-recognized influence on ecological systems [27,73,74]. Packman et al. [10] predicted in 2014 that if rates of grassland decline (habitat loss) continued for the following decade (2012–2022), then the critically endangered Bengal Florican bustard would become extinct. Our estimates show that 122,604 ha (41.3%) of the 2004 grassland area still remains within this landscape. This estimate, however, does not take into account the number of patches meeting the minimum habitat requirement for this species. The studies by Niu et al. [31] and Senevirathne et al. [30] marked historic trends in the expansion of agriculture. In agreement with these findings, this study estimates an increase in cropland area of 27.3% (347,059 ha) within the TSL landscape. Chen et al. [5] found a substantial decline in forest cover in their study of the TSL landscape between 1992 and 2019 (2.3% annual decline). Our analysis of the 2004 to 2023 forest cover trends found identical rates of decline for this land cover class (2.3% annually). Regarding the influence of croplands on the loss of grasslands, Packman et al. [6,10,75] quantified that 95% of the grassland losses in the south-eastern region of the Tonle Sap floodplain were attributable to rice cultivation. Our analysis across the entire landscape suggests that 89.3% of the loss of dry grasslands can be accounted for by increases in cropland area. The difference in these estimates could be attributed to either differences in study area or differences in the definition of grasslands. The 2011 study defined grasslands based on soil types, while this study used the frequency of water occurrence [6].

The use of machine learning classification and regression methods for grassland monitoring is increasing, yet is still underrepresented in current remote sensing [12,14]. Like other studies investigating complex land cover and land use change dynamics over large areas, this project faced several challenges. First, change-detection accuracy assessments are particularly difficult due to the availability and reliability of historic reference data, especially for large areas [19,24,57]. This study relied on reference data generated from the interpretation of historic and more recent high-resolution imagery, as well as land cover maps generated from local landscape experts. While a large number of reference data samples for each land cover class were generated, they were not specific to each of the five years classified during this study. Compounding the challenge of adapting the available reference data to this study was the diverse mixture of vegetation found within the floodplains. Image analysis of areas with diverse vegetation mixtures can be difficult with moderate-resolution imagery such as Landsat [20]. The flooded forests, shrublands, and grasslands in many areas are a patchwork mosaic of emergent vegetation, especially when relying on the 30 m Landsat spatial resolution [3]. Due to this challenge, reference samples with limited certainty when compared to the Landsat image mosaics were removed from the analysis. A second potential source of uncertainty in this analysis was the use of a post-classification change detection. No single change-detection method can be optimal for all cases [19,24,57]. This study used the given approach due to its ease of implementation and established reliability [24,26,57]. Future studies should leverage the potential of more advanced land cover trend analysis methods, such as Continuous Change Detection and Classification (CCDC) or LandTrendr, to generate more precise estimates of land cover change [76,77,78,79,80]. Lastly, the classifications in this study relied on a combination of optical imagery and elevation data. While Landsat imagery provided a stable data source, optical imagery collections for each mapping period were limited in this region to only the dry seasons [14,18,28]. Imagery queried from May to October of each year contained considerable amounts of clouds, shadows, and noise, degrading the quality of the median pixel mosaics. The obstruction of clouds throughout the wet season also meant that seasonal image composites, stacks containing bands from two or more seasons within a year, were unfeasible [21,39,81,82]. A combination of optical and Synthetic Aperture Radar (SAR) remotely sensed data has shown promise in paddy rice mapping due the changes in this crop’s growth stages across seasons, but this was not approached in this study [17,18]. The integration of SAR remotely sensed data or increased spatial resolution would potentially also improve the ability to map individual crop types. Having such data would aid the use of the resulting products for regional conservation agencies. A supplemental goal of this analysis was to provide a methodology which could be easily trained and adapted to neighbouring regions, which was accomplished by limiting the technical complexity to only optical imagery.

The results of this study are not unique to the TSL landscape and should be used to inform broader conservation policies and monitoring efforts. Grasslands are an ecosystem fundamental to global sustainability goals [83]. These systems experienced a decline of over 39% during the first half of our study period [84]. In one example, Baeza et al. [85] report a loss of 2.4 million ha (9% of the 2001 area) in South American grasslands from 2001 to 2018. Agricultural expansion (a primary form of human disturbance) into areas of natural vegetation is an issue [29].

Rice cultivation, in this region specifically, has existed for over 1000 years [9]. It has only been for the last 20 to 30 years that industrial-scale rice production has been a major threat to grassland habitats [10]. The onset and escalation of this disturbance was caused by a mixture of social, economic, and political challenges [13]. Many of the villages within the TSL landscape are reliant on agriculture [86,87]. Studies have shown that simply increasing the land in production for each farmer does not cause a net increase in their income [87]. The rate of population growth, coupled with the high population density in the TSL landscape, will further burden these efforts in the coming decades [4,5]. The results of this conflict between social and ecological needs will result in the further decline, and potential extinction, of grassland-dependent species such as the Bengal florican [9,13]. To avoid this potential outcome, locally integrated conservation organizations must continue to increase their engagement with local farmers and commercial stakeholders [6,13,86]. Management decisions must also be supported and enriched by the most capable cloud computing and EO methods [38,88].

5. Conclusions

The Tonle Sap Lake (TSL) landscape supports a rich, diverse habitat, ecosystem services utilized by over a million people in the Lower Mekong River Basin, and numerous critically endangered species. To support local biodiversity, nine protected areas totalling more than 170,000 ha have been established throughout this region. Despite ongoing conservation efforts, historic studies and local experts have reported drastic declines in grasslands over the last two decades. This analysis surveyed land cover changes from 2004 to 2023 within the TSL landscape and a surrounding 10 km buffer area to generate current and reliable estimates of the decline in grasslands and dry grasslands. Five thematic maps, achieving an average overall accuracy of 83.5%, were used to perform a change-detection analysis. From 2004 to 2023, the estimated decline in dry grasslands was 174,400.2 ha (58.7% of the area in 2004). Of the areas no longer containing dry grasslands, 89.3% were classified as croplands in 2023. An analysis of the protected landscapes estimated that 45.4% of the dry grasslands present in 2004 were lost by 2023. Future conservation efforts should build on the methods and results used here as a pathway to local-scale management and outreach. Without an increase in current, reliable, and comprehensive data on the trends and distribution of land cover such as dry grasslands within this region, conservation efforts aimed at protecting the numerous critically endangered and threatened species in this region would be restricted in their potential.