The study area corresponds to the hydrological region number 15 – Costa de Jalisco located between 18° 49′ N, 20°30′ N Latitude and 103°50′ W, 105°50′ W Longitude and covers a total area of 12,967 km2. The study area encompasses the pacific coast of the states of Jalisco (from the south of Puerto Vallarta) and Colima (to the south of the Port of Manzanillo) (Figure 1). It has an annual precipitation of 1144 mm with almost 80% of rainfall occurring between June and October, and a mean annual surficial runoff of 3606 hm3 per year (INEGI, 2000). The highly heterogeneous landscape has promoted a high diversity of flora and fauna, especially in the central part, where the Chamela-Cuixmala Biosphere Reserve, a protected area of 131 km2 of well-preserved tropical deciduous forest is located (Cotler and Ortega-Larrocea 2006; Suazo-Ortuño et al. 2008; Avila-Cabadilla et al. 2012). Most of the coast of Jalisco state is represented by tropical dry forest, followed by evergreen forest and an extensive area that includes agricultural fields and pastures (Sánchez-Asofeifa et al. 2009).
The rivers Tomatlán, San Nicolás, Cuixmala, and Purificación, are born in the Sierra de Cacoma and descend almost parallel to the Pacific Ocean. These currents are underdeveloped due to the proximity of the mountains to the coast. The Chacala river (also known as Cihuatlán or Marabasco) serves as the boundary between Jalisco and Colima, and when it flows into the Pacific forms the Barra de Navidad bay (INEGI, 2000).
The landscape changes were analyzed from land use and land cover (LULC) maps derived from the classification of satellite images acquired by the Landsat 5 Thematic Mapper (from 1986), the Landsat 7 Enhanced Thematic Mapper Plus (from 2001) and the Landsat 8 Operational Land Imager (from 2017) and obtained through the USGS Global Visualization Viewer (http://glovis.usgs.gov/). All images were captured during the dry season when the phenological characteristics of the main vegetation types allow us a good discrimination. Three satellite images by date with Path-Row: 29-47, 30-46, 30-47 were used to cover the study area.
The basin limit was downloaded from the Flow Simulator in Drainage Basins (SIATL) scale 1:50,000 from the National Institute of Statistics and Geography (INEGI, http://antares.inegi.org.mx/analisis/red_hidro/siatl/). The study area was isolated from each scene by a mask produced by the rasterization of the basin limits vector. All scenes were individually classified by using K-means unsupervised classification algorithm run in IDRISI TerrSet software (Eastman 2016). The classification scheme involved the following LULC classes: Aquatic surfaces (AS), Evergreen forest (EF), Tropical Dry Forest (TDF), Exposed soil(ES), Crops (CR), Saltmarsh (SM), Mangrove(MN), Littoral (LI) and HS (Human settlements). This later LULC class was hand digitized on-screen over false-color composites of satellite images by date analyzed.
In order to estimate the accuracy of our more recent thematic map, we conducted field verifications during 2016 and 2017 to obtains numbers of verification points and evaluate the LULC map of 2017. Accuracy assessment was accepted until it reached an 80% or higher in overall accuracy (Congalton and Green 2009).
Scenario modelling and validation
LULC scenarios were modeled using Cellular Automata (CA) and Markov chain analysis in IDRISI TerrSet software from Clark Labs. In a first stage, we calculated a transition probability matrix for the period 1986 – 2001 was calculated using Markov chains with the LULC map of 1986 as the first land cover image and the map of 2001 as the later land cover image. The transition areas and the conditional probabilities created in the step previous were used in a Cellular Automata analysis to predict the LULC in 2017. This prediction was used for model validation. A second transition probability matrix was calculated using the LULC maps of 2001 and 2017 and used to predict LULC in 2033. Finally, we calculated a transition matrix for the whole period (1986 – 2017) and we used it to predict the LULC in 2050. These transition probabilities matrices have the information about the probability that each land cover category will change to every other category (Eastman 2016; Wang et al. 2016). During the Markov chain model setting stage, the number of time periods to forward projection was set accordingly the number of time periods between the first and the second image for each period analyzed. In the Cellular Automata/Markov Change prediction setting, the later land cover image used in the Markov Chain analysis was used as the starting point for change simulation, a standard 5*5 contiguity filter was applied meanwhile the number of Cellular Automata iterations it depended on the number of time periods for forward projection specified in the Markov chain analysis (Takada et al. 2010; Wang et al. 2018).
To quantify the predictive power of the model, we compared the result simulation (2017) with the reference map (2017) using Kappa variations (Pontius 2000): Kappa standard (Kstandard), Kappa for no information (Kno), and Kappa for location (Klocation). For all of the Kappa statistics, 0% indicates that the level of agreement is equal to the agreement due to chance and 100% indicates perfect agreement. In comparing the map of reality to the alternative map, Kno indicates the overall agreement. Klocation indicates the extent to which the two maps agree in terms of location of each category (Eastman 2016). We accepted the model to make future projections only if the Kappa values were greater than 80% (Araya and Cabral 2010).