Land cover changes in grassland landscapes: combining enhanced Landsat data composition, LandTrendr, and machine learning classification in google earth engine with MLP-ANN scenario forecasting

Figures & data

Figure 1. Location of Umbria in Italy (red polygon in the inset). The study area is shown in Landsat 8 composite infrared together with the boundaries of the Umbria region and the OCM Landscape layer in the background. White rectangles indicate the locations of the areas chosen to show the classification results in detail.

Figure 2. Methodological workflow of the research.

Table 1. Selected LULC classes, codes, and descriptions.

LULC Class	Code	Description
Woodland	1	Forest vegetation, dominated by different tree species.
Shrubland	2	Shrubs and scrublands (usually dominated by brooms, juniper, blackthorn, and hawthorn).
Grassland	3	Grassland vegetation, predominantly comprised of grasses, forbs, and small shrubs in variable proportion, ranging from dense and continuous to sparse and discontinuous, including both primary and semi-natural secondary grasslands (including habitat types 6210 and 6230*, Annex I, Directive 92/43/EEC).
Cropland	4	Agricultural mosaic mainly composed of annual crops.
Sparse/No Vegetation	5	Areas characterized by very sparse and fragmented vegetation cover, with a remarkable presence of rocky substrate; persistently non-vegetated areas characterized by bare soil; built-up areas.

Table 2. Spectral indices and related formulas.

Index	Formula	Authors
NDVI	$\frac{NIR-RED\hspace{0.17em}}{NIR+RED}$	(Rouse et al. 1973)
EVI	$2.5\times \frac{NIR-RED}{NIR+\left(6\times RED\right)+\left(7.5\times BLUE\right)+1}$	(Huete et al. 2002)
GARVI	$\frac{NIR-\left(GREEN-\left(BLUE-RED\right)\right)}{NIR-\left(GREEN+\left(BLUE-RED\right)\right)}$	(Gitelson, Kaufman, and Merzlyak 1996)
MARI	$\left(\frac{1}{GREEN}-\frac{1}{RED}\right)\times NIR$	(Gitelson, Keydan, and Merzlyak 2006)
BSI	$\frac{\left(SWIR1+RED\right)-\left(NIR+BLUE\right)}{\left(SWIR1+RED\right)+\left(NIR+BLUE\right)}$	(Rikimaru, Roy, and Miyatake 2002)

Table 3. Textural indices calculated using GLCM (gray-level Co-occurrence matrix).

CODE	GEE CODE	Textural index description
T1	ASM	Angular Second Moment: estimates the number of repeated pairs in the image
T2	CONTR	Contrast: estimates the contrast within an image
T3	CORR	Correlation: estimates the correlation between pairs of pixels in the image
T4	VAR	Variance: estimates the extent of dispersion in the distribution of gray-levels in the image
T5	IDM	Inverse Difference Moment: estimates the homogeneity of the image
T6	SAVG	Sum Average: calculates the mean of the gray level sum distribution of the image
T7	SVAR	Sum Variance: estimates the dispersion (regarding the mean) of the gray-level sum distribution of the image
T8	SENT	Sum Entropy: estimates the randomness of the gray-level sum distribution of the image
T9	ENT	Entropy: estimates the randomness of the gray-levels distribution of the image

Table 4. Bands and indices calculated for the composition of the base datacube.

Bands and indices	Number of bands	Codes/names
10^th, 50^th, 90^th percentiles and interquartile range of B1, B2, B3, B4, B5, B7 and NDVI, EVI, GARVI, MARI, BSI	44	Suffixes _p10, _p50, _p90, _iqr
Textural indices	9	T1 to T9
Harmonic components	3	‘phase,’ ‘magn,’ ‘val’
Elevation, slope, aspect	3	‘elevation,’ ‘slope,’ ‘aspect’
EVI slope, local standard deviation, gradient, and direction	4	‘slopevi,’ ‘sdloc,’ ‘grad,’ ‘dir’

Figure 3. Distribution of the importance scores of the variables used in the RF classification for (a) 2000, (b) 2010, and (c) 2020.

Table 5. Accuracy metrics of LULC classifications for each year (UA: User’s Accuracy; PA: Producer’s accuracy).

LC Class	2000		2010		2020
	UA	PA	UA	PA	UA	PA
Woodland	0.92	0.99	0.9	0.99	0.95	0.98
Shrubland	0.69	0.69	0.73	0.58	0.73	0.76
Grassland	0.99	0.71	0.94	0.7	0.95	0.82
Cropland	0.94	0.8	1	0.9	0.92	0.78
Sparse/No Vegetation	0.9	1	0.94	0.88	0.82	0.9
OA	0.90 ± 0.012		0.90 ± 0.012		0.92 ± 0.011

Figure 4. Sankey’s diagram showing the main transition flows (greater than 5 km²) between the five land-cover classes in the study area between 2000–2010, 2010–2020 and 2020–2040.

Figure 5. Land cover maps of 2000, 2010, 2020, and 2040 in two exemplifying areas. The scale and the north sign are indicated in the final image of both areas. The full maps, and a comparison with orthophoto and a global land cover product are provided in the supplementary materials.

Table 6. Area of land cover classes (km²) in three different years (2000, 2010, 2020) and the increase/decrease percentage over 20 years (2000–2020) for the entire study area and the Natura2000 sites. The potential area of land cover classes (km²) and percentage of increase/decrease are also shown for the 2040 forecast.

LC Class	Code	Study Area						Natura2000
		2000	2010	2020	2040	2000–2020%)	2020–2040%)	2000	2010	2020	2040	2000–2020%)	2020–2040%)
Woodland	1	1164	1271	1283	1356	+10.2	+5.7	332	357	359	381	+8.1	+6.1
Shrubland	2	193	137	130	76	−32.6	−41.5	44	34	34	16	−22.7	−52.9
Cropland	3	91	96	72	103	−20.9	+43.1	20	24	19	28	−5	+47.4
Grassland	4	204	165	167	141	−18.1	−15.6	112	103	98	96	−12.5	−2.0
Sparse/No Veg.	5	56	40	57	32	+1.8	−43.9	31	23	30	20	−3.2	−33.3

Rouse, J. W., R. H. Hass, J. A. Schell, and D. W. Deering. 1973. “Monitoring Vegetation Systems in the Great Plains with ERTS.” Third Earth Resources Technology Satellite-1 Symposium NASA 1:309–317. https://doi.org/citeulike-article-id:12009708.

 Google Scholar

Huete, A., K. Didan, T. Miura, E. P. Rodriguez, X. Gao, and L. G. Ferreira. 2002. “Overview of the Radiometric and Biophysical Performance of the MODIS Vegetation Indices.” Remote Sensing of Environment 83 (1–2): 195–213. https://doi.org/10.1016/S0034-4257(02)00096-2.

 Web of Science ®Google Scholar

Gitelson, A. A., Y. J. Kaufman, and M. N. Merzlyak. 1996. “Use of a Green Channel in Remote Sensing of Global Vegetation from EOS-MODIS. Remote Sens.” Remote Sensing of Environment 58 (3): 289–298. https://doi.org/10.1016/S0034-4257(96)00072-7.

 Web of Science ®Google Scholar

Gitelson, A. A., G. P. Keydan, and M. N. Merzlyak. 2006. “Three-Band Model for Noninvasive Estimation of Chlorophyll, Carotenoids, and Anthocyanin Contents in Higher Plant Leaves.” Geophysical Research Letters 33 (11): 33. https://doi.org/10.1029/2006GL026457.

 Web of Science ®Google Scholar

Rikimaru, A., P. S. Roy, and S. Miyatake. 2002. “Tropical Forest Cover Density Mapping.” Tropical Ecology 43:39–47.

 Google Scholar

Data availability statement

All remote sensing data used in the research are openly available in the USGS archives (https://earthexplorer.usgs.gov/) and within Google Earth Engine. The GEE codes developed in this research are available by emailing the authors upon any reasonable request.