Remote Sensing Watershed Landscape Dynamics

Authors: Lara Cornejo-Denman, Jose Raul Romo-Leon, Kyle Hartfield, Willem J.D. van Leeuwen, Guillermo E. Ponce-Campos, and Alejandro Castellanos-Villegas

In this study, we developed remote sensing-based land cover classifications and post-classification fragmentation analysis, by using data from Landsat’s moderate resolution sensors Thematic Mapper and Operational Land Imager (TM and OLI) to assess land use changes and the shift in landscape configuration in a riparian corridor of a dynamic watershed in central Sonora during the last 30 years. In addition, we derived a high spatial resolution classification (using PlanetScope-PS2 imagery) to assess the “recent state” of the riparian corridor.

• Land Cover Trends for Moderate and High Spatial Resolution Classifications
• Riparian Vegetation and Agriculture
• Introduced Grassland and Scrub
• Fragmentation and Implications of Habitat Loss
• Multiple Resolution Datasets

Despite the prominent role of LULCC in ecological processes at landscape levels, comprehensive analyses of habitat connectivity and configuration are still sparse for many key environments within large landscapes. Such is the case of riparian ecosystems in arid and semi-arid regions of North America, which are hotspots of biodiversity and ecosystem services. These dynamic areas are crucial for numerous ecological systems and modifications to the areas can create multiple changes on a regional scale. Selecting the Rio Sonora Subwatershed (RSSW) to represent riparian ecosystems in Northwest Mexico, this study aims to quantify changes in land cover and landscape configuration (from 1988 to 2018) of several classes (native vegetation and human land use) associated to the riparian corridor of the Río Sonora. We develop land use–land cover classifications using moderate (Landsat TM and OLI) and high (PlanetScope-PS2) spatial resolution satellite imagery. In addition, for 1988 and 2016, we assess three fragmentation variables: Number of Patches, Mean Patch Area and Class Aggregation Index. Finally, we discuss the advantages and challenges of using different spatial resolution datasets for the assessment of riparian vegetation dynamics.

Supervised classifications of the riparian corridor were derived for 1988 and 2016 using satellite imagery from two different Landsat sensors (TM and OLI). An additional classification for 2018, based on higher spatial resolution imagery from PlanetScope-PS2 sensors, was also conducted to obtain a “recent state” description of LULC for the region. A post-classification change detection analysis was performed using the Landsat TM- and OLI-based classification products. Finally, a fragmentation analysis was executed for 1988 and 2016 to assess the number of patches and landscape connectivity percentages for each land cover class.

For 1988 and 2016, cloud-free Level 1 Precision and Terrain corrected Landsat imagery (L1TP) from Collection 1 were downloaded from the United States Geological Survey (USGS) Earth Explorer website (Table 1). Data from this collection are radiometrically calibrated and orthorectified, ideal products for pixel-level time series analysis. Since individual scenes cover only a portion of the subwatershed, after the classification procedure was completed, the two products were stitched into a larger mosaic to cover the full study area. To increase our classification accuracy, two dates were selected for each year to have information regarding the phenology of vegetation during both the dry and rainy seasons. For 2018, 55 ortho-scenes from the Level 3B Products of Planet Constellation PlanetScope-PS2 (Education and Research Program) were selected ( Table 1). These scenes are orthorectified and radiometrically, geometrically and atmospherically corrected.

A set of variables were selected based on a literature review and derived from the satellite imagery datasets. Table 2 shows the metrics and variables, along with their associated description and the datasets from which they were derived. Ancillary data matching Landsat’s spatial resolution and sub-products were also obtained, including a digital elevation model, slope layer and aspect layer.

Considering riparian vegetation as the focus in this study, several cover types were selected to create a classification scheme. Given the differences in the spatial resolution of the sensors used, two classification schemes were developed, one for each dataset (Landsat TM and OLI and PlanetScope-PS2).

A reference dataset, consisting of an average of 200 points for each land cover class, were acquired, a random sample of 70% was used for training and the remainder (30%) for validation. Land cover training points were selected by: (1) collecting GPS ground data directly from field visits (2017 and 2018), (2) aerial imagery directly obtained in the field (2018) and (3) field data from previous studies in the region. All points were verified and re-located using reference tools such as orthophotos from the Mexican Institute of Statistics and Geography (INEGI, for its acronym in Spanish) and current and historical images from Google Earth.

Three supervised classifications were generated (1988, 2016 and 2018) by applying the classification and regression tree (CART) model. The accuracy of each classification was assessed using the validation data withheld from the reference dataset. A minimum of 75% accuracy was expected in order to validate the classification.

To measure the main conversions between land uses and vegetation types through the studied time period between 1988 and 2016, a pixel-by-pixel change detection analysis was performed through a post-classification change detection approach. For this analysis, we used 1988 and 2016, since both classifications share similar spatial resolution and class schemes. This produced a combined raster and attribute table with the pixel count for every class combination that occurred from one year to another and final cover change for each class was determined by converting the pixel count to hectares (pixel area of 900 m 2  = 0.09 ha).

Land cover transitions for each class from 1988 to 2016, along with total change in hectares, are presented in Table 9. Subsequently, we explain some of the main cover trends for the classes of interest (Agriculture, Riparian Vegetation and Introduced Grassland).

Fragmentationassessment presents the landscape configuration assessment for the LULC maps generated from Landsat data, to address changes between 1988 and 2016. This analysis registers key landscape modifications, represented by an increase in fragmentation for several natural vegetation classes, as well as an increase in connectivity for some anthropogenic-induced classes.

According to our results, riparian vegetation has increased by 40%, although only 9% of this coverage corresponds to obligate riparian species. Scrub area shows a declining trend, with a loss of more than 17,000 ha due to the expansion of mesquite and buffelgrass-dominated areas. The use of moderate resolution Landsat data was essential to register changes in vegetation cover through time, however, higher resolution PlanetScope data were fundamental for the detection of limited aerial extent classes such as obligate riparian vegetation. The unregulated development of anthropogenic activities is suggested to be the main driver of land cover change processes for arid ecosystems in this region. These results highlight the urgent need for alternative management and restoration projects in an area where there is almost a total lack of protection regulations or conservation efforts.