A landscape metrics-based sample weighting approach for forecasting land cover change with deep learning models

Figures & data

Figure 1. Study area details showing (a) the Tumbler Ridge District Municipality in British Columbia, Canada, and (b) the 2001 LC of the study area displayed using the Canada Lambert conformal conic projected coordinate system.

Figure 2. A comparison of the landscape composition in years 2001 and 2019 expressed as a percentage of the landscape.

Figure 3. The percentage of annually changed landscape area with respect to each LC class from 2001 to 2019.

Figure 4. Auxiliary spatial variables prior to Min-max normalization and displayed using the NAD 1983 UTM 10N projected coordinate system: (a) elevation, (b) slope, (c, d) distance to population centres, (e) distance to roads, (f) distance to parks and protected areas, (g) distance to agricultural land reserves, (h) distance to lakes, and (i) distance to rivers.

Figure 5. An overview of the hybrid CNN-LSTM model showing the branched structure shared across all model types. Each model branch accepts the spatiotemporal LC data with dimensions $T\times M\times M\times L$ and the explanatory spatial variables with dimensions $M\times M\times A.$

Figure 6. Overview of the LSM-based sample weighting approach implementation.

Table 1. The selected class-level LSMs calculated for the 2001 LC of Tumbler Ridge, British Columbia.

LC Class	Class area (km^2)	IJI (%)	PD (patch/ha)	PLADJ (%)	LSI	CIRCLEMN	CONTIGSD	FRACMN	SHAPESD
Water	6.11	48.23	0.35	59.72	33.18	0.52	0.22	1.05	0.75
Developed	2.02	51.57	0.01	83.89	7.61	0.54	0.29	1.07	0.99
Barren/Sparsely Vegetated	13.06	56.60	0.60	68.98	37.38	0.50	0.18	1.04	0.72
Tree Cover	1373.59	50.13	0.92	97.18	36.09	0.49	0.19	1.04	0.93
Shrublands	16.64	46.22	4.34	29.31	96.15	0.47	0.14	1.03	0.32
Herbaceous	165.93	45.88	3.78	77.64	96.22	0.53	0.22	1.06	0.77

Figure 7. Comparison of relative importance assigned to weight each LC class of the study region.

Table 2. Descriptions of the LSM-based class weighting schemes derived from selected measures of aggregation and shape.

Class-level Metric Type	Class Weighting Scheme Name	Acronym	Class Weight Derivation Strategy (i.e. direct usage or invert metric values?)	Purpose (i.e. which aspects of the configuration or composition determine LC class importance?)
Aggregation	Interspersion and Juxtaposition Index Weight	IJI-W	IJI measures computed for each class are provided as the respective CWs	Increase the importance of LC types that are more interspersed with different LC types
	Inverse Patch Density Weight	PD-IW	Inverting the patch density measures of the LC classes	Increase the importance of classes with less patch density or fewer patches, assuming more aggregated classes will exhibit less frequent change
	Inverse Percentage of Like Adjacencies Weight	PLADJ-IW	Inverting the percentage of like adjacencies measures computed for each class	Increase the weights associated with classes with mostly disaggregated patches; minimize the influence of classes with increasingly homogeneous patches
	Inverse Landscape Shape Index Weight	LSI-IW	Inverting the landscape shape indices of all LC classes	Increase the importance of compact classes, assuming aggregated classes require additional weight allocated to less frequent changes
Shape	Inverse Mean of Circumscribing Circle Weight	CIRCMN-IW	Inverting the measures describing the average circumscribing circle size for patches belonging to each class	Increase the importance of classes in which patches tend to be small; decrease the influence of classes in which patches tend to be large
	Standard Deviation of Contiguity Index Weight	CONTSD-W	The standard deviation of contiguity for each LC class	Increase the importance of classes that have greater diversity of contiguity across all patches belonging to the class
	Mean Fractal Dimension Index Weight	FRACMN-W	The mean fractal dimension index is provided directly for each CW	Increase the importance of classes characterized primarily by irregularly shaped patches; lower the importance of classes in which the average shape across patches is squared
	Standard deviation of Shape Indices Weight	SHAPESD-W	The standard deviation shape index for each LC class is provided directly for each CW	Increase the importance of classes that have greater variation of shape indices, ensuring classes with more variable patch shape complexity are prioritized

Figure 8. Boxplot of FOM measures obtained by comparing the real-world and simulated changes for 2015 across all model types. The three class weighting schemes attributed to the highest average FOM values are shown with the prefix (**).

Figure 9. The FOM, PA, and UA calculated with respect to the cumulative LC changes correctly projected from 2015 to 2019.

Figure 10. Trends of FOM components attributed to each class weighting scheme for the 2019 land cover forecasts. The box plots depict the percent of study area extent identified as (a) correctly forecasted changes, (b) false alarms, (c) missed changes, and (d) changes forecasted as the incorrect class summarized across all model types.

Figure 11. Comparison of cumulative projected and observed changes from 2015 to 2019 for each LC class and the selected class weighting schemes.

Figure 12. Comparison of the average error in allocation distances computed with respect to the real-world 2019 LC for each class and the selected class weighting schemes.

Figure 13. Mapped FOM components of the 2019 LC change forecasts shown for (a) CNN-LSTM_PD-IW and (b) ConvLSTM_PD-IW displayed using the NAD 1983 UTM 10N projected coordinate system.

Data availability statement

The open datasets used this study include the Global Land Cover Mapping and Estimation (GLanCE) annual 30 m (m) Version 1 data product https://lpdaac.usgs.gov/products/glance30v001/ (accessed 3 January 2023), the Terra Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation Model (GDEM) Version 3 (ASTGTM) https://lpdaac.usgs.gov/products/astgtmv003/ (accessed 2 December 2022), Population Centers and Digital Boundary Files of Canadian Provinces and Territories https://www12.statcan.gc.ca/census-recensement/2011/geo/bound-limit/bound-limit-2016-eng.cfm (accessed 2 December 2022), Canadian Census Road Network https://open.canada.ca/data/en/dataset/57d5ffae-3048-4a19-9b4c-eab12f6322c5 (accessed 2 December 2022), British Columbia Railway Network https://catalogue.data.gov.bc.ca/dataset/railway-track-line (accessed 3 December 2022), Lakes https://catalogue.data.gov.bc.ca/dataset/freshwater-atlas-lakes (accessed 3 December 2022), Rivers https://catalogue.data.gov.bc.ca/dataset/freshwater-atlas-rivers (accessed 2 December 2022), Parks and Protected Areas https://catalogue.data.gov.bc.ca/dataset/parks-and-protected-areas-regional-boundaries (accessed 2 December 2022), Agricultural Land Reserve https://catalogue.data.gov.bc.ca/dataset/alc-alr-polygons (accessed 2 December 2022), and Canadian Cartographic Boundary File https://open.canada.ca/data/en/dataset/a883eb14-0c0e-45c4-b8c4-b54c4a819edb (accessed 28 February 2023).