Mapping Vegetation Species Succession in a Mountainous Grassland ecosystem using Landsat and Sentinel-2 data

Vegetation species succession and composition are significant factors determining the rate of ecosystem biodiversity recovery after being disturbed and subsequently vital for sustainable and effective natural resource management and biodiversity. The succession and composition of grasslands ecosystems worldwide have significantly been affected by the accelerated changes in the environment due to natural and anthropogenic activities. Therefore, understanding spatial data on the succession of grassland vegetation species and communities through mapping and monitoring is essential to gain knowledge on the ecosystem and other ecosystem services. This study used a random forest machine learning classifier on the Google Earth Engine platform to classify grass vegetation species with Landsat 7 ETM+ and ASTER multispectral imager (MI) data resampled with the current Sentinel-2 MSI data to map and estimate the changes in vegetation species succession. The results indicate that ASTER IM has the least accuracy of 72%, Landsat 7 ETM+ 84%, and Sentinel-2 had the highest of 87%. The result also show that other species had replaced four dominant grass species totaling an area of about 49 km2 throughout the study.


38
Vegetation succession has many economically significant attributes, including high overall 39 biomass and productivity, a wider variety of species, and minimal nutrients or energy from the 40 ecosystem (1). Nevertheless, an ecosystem with naturally occurring succession stages will be more 41 resilient to natural and anthropogenic disturbances, suppose these disturbances increase in severity, 42 frequency, and magnitude because of human activities and weather conditions. In that case, the 43 pressure on plant communities increases, thereby causing an accelerated succession and creating 44 a new vegetation community and allowing the succession of non-native species (2). 45 Luken (1) describes vegetation succession as a change in vegetation composition over 500 years 46 without been disturbed to achieve a stable species composition called a climax. Climate change 47 and other disturbances within a short period may result in fluctuations of species composition, 48 promote non-native species and delay the natural vegetational succession from reaching its climax. 49 The succession of non-native species could impact biodiversity and the natural ecosystem. Non-50 native invasive species quickly inhabit disturbed spaces and delay native species from achieving 51 seral or climax states. In some cases, the succession is entirely taken over and held for an extended 52 period at an intermediate state, affecting biodiversity (3-5). Invasive vegetation species threaten 53 native vegetation species and water resources because they grow faster, consume more water, and 54 spread more than the native species (6, 7). The encroachment of these vegetation species tends to 55 alter the balance of ecosystems, thereby accelerating succession. Vegetation species succession 56 and composition are significant factors determining the rate of ecosystem biodiversity recovery 57 after being disturbed (8) and subsequently vital for sustainable and effective natural resource 58 management and biodiversity. For example, in South Africa, changes in vegetation succession 59 resulting from disturbances have led to significant losses in biodiversity (9).

60
Worldwide, the succession and composition of grasslands ecosystems have been significantly 61 affected by the accelerated changes in the environment due to natural and anthropogenic activities 62 (10, 11). It has resulted in shortages in grasslands taxonomy and efficient functioning of ecosystem 63 services (12). Grasslands' changing diversity and composition impact ecosystem services like 64 precipitation and temperature controls, freshwater supply, erosion control, and soil formation (13-65 15). They can likely result in biodiversity loss (16). About one-third of South African land surface 66 is covered by the grassland biome (17). It has just less than 3% located in protected areas, and 40-67 60% have been altered with little chance of been salvaged and returned. It makes the grassland one 68 of the most vulnerable biomes in South Africa (18). Therefore, understanding spatial data on the 69 succession of grassland vegetation species and communities through mapping and monitoring is 70 essential to gain knowledge on the ecosystem and other ecosystem services (9,19,20).

71
Remote sensing provides an efficient approach for mapping grassland vegetation species by 72 reducing rigorous fieldwork necessitated by standard mapping methods. It does this effectively by 73 offering a wide range of recent data on vegetation species distribution from hyperspectral and 74 multispectral imagery (21,22). Extensive studies have been undertaken in monitoring spatio-75 temporal changes in vegetation species composition and diversity using remote sensing data (23-76 26). However, these studies focus briefly on a short period, usually between one to five years, 77 because, before now, only high-resolution hyperspectral images could give accurate vegetation 78 species discrimination at individual levels (9,19,(27)(28)(29)(30). Nevertheless, recent studies have shown 79 that free low-resolution satellite images like Sentinel-2 MSI and Landsat 8 OLI can be used to 80 accurately map and monitor grassland vegetation species (31-33). Vegetation species succession 81 and diversity monitoring can be done over an extended period using these low-resolution imageries 82 in combination with machine learning. Therefore, this study used Landsat 7 ETM+ and ASTER 83 MI data fused with the current Sentinel-2 MSI data to map and estimate the changes in vegetation

107
The study used satellite images from different sensors to cover the period of study. The sensors previous study (31) that used deep learning and machine learning models to discriminating grass 122 species at the individual level. Their study recommended Sentinel-2 MSI bands 6, 7 (red edge), 123 bands 8 and 8A, band 11, and band 12 to produce optimum classification accuracy. Therefore, the 124 spectral resolution of these bands was used to match and select the bands in the Landsat ETM+ 125 and ASTER MI as presented in table 1. The spectral library was used to generate 100 random 126 sample locations for training and cross-validation. The locations were randomly split into a 127 training set (70%) to train the classifiers (31, 35) and a test set (30%) for testing purposes (31, 36).

128
The GEE code editor random forest machine learning classifier with ten trees was used to process   Table 2 shows the accuracy of each sensor. ASTER IM has the least accuracy of 72%, Landsat 7 147 +ETM 84%, and Sentinel-2 had the highest of 87%. The ASTER IM had the lowest level of 148 accuracy, possibly because (40) stated that each scene does not have all 14 bands. Therefore, some 149 scenes may have fewer bands than others. Hence, only bands 1 to 3 were available for that period.

150
However, these bands' spectral range can be compared to bands 1-5 of Landsat +ETM and OLI.

151
Another possible reason is that the three bands available didn't adequately separate the grass 152 species from each other, as shown in the spectral reflectance curve in Figure 2. Nevertheless, if 153 all the bands were available, ASTER IM should discriminate the grass species effectively to attain 154 a higher accuracy using machine learning classifiers. The accuracy of the ASTER image agrees 155 with a study done by (41). Their study used ASTER NDVI and EVI to discriminate rice and citrus 156 fields with 75% and 65% accuracy, respectively. They also used Landsat 5 TM NDVI and EVI,    Figure 2 shows the spectral reflectance of the twelve grass species extracted from all the sensors.

168
In the Landsat 7 ETM+, the species were discriminated in wavelengths of 0.52 -0.77 µm and 1.55 169 -2.08 µm, representing bands at the start of the wavelength for the panchromatic near-infrared, 170 shortwave infrared 1, and shortwave infrared 2. The ASTER MI has its best spectral separation 171 wavelength of 0.780-0.860µm (VNIR near-infrared, nadir pointing band). At the same time, the 172 Sentinel-2 separated it best in the bands 6, 7 (red edge), bands 8 and 8A, band 11, and band 12 as 173 recommended by the study by (31, 44-46), hence the difference in classification accuracy.     P. australis is a decreaser species quickly affected by overgrazing and has a slow recovery rate 287 after a disturbance (48). It is a tall grass found across South Africa, especially around river beds and wet environments, and is not at risk of extinction. (47,49). Nevertheless, figure 11 shows that 289 it is being replaced mainly by T. triandra (2.55 km 2 ), E. plane Nees (2.3 km 2 ), and M. junceus (1.4 290 km 2 ). It is also gaining back by replacing S. centrifugus and E. curvula. It is found across the study 291 area around the river channels and is replaced mainly in the southern and northern parts of the 292 study area (Figure 12). replacing others or has been replaced. Some of these species are used for human activities like 305 thatching and medicinal purposes, and some are palatable for grazing. Climate change and fires 306 are common factors that can also affect these successions (50, 51). The study area is a region 307 constantly affected by wildfires, and the fire severity and magnitude have been mapped by (31).

308
Their research showed some parts of the study area had constantly been burnt with high fire 309 severity over 20 years. These parts of the study area may be experiencing changes in species 310 composition, leaving only the fire-tolerant species or invasive species like S. plumosum, which is