Geospatial-Temporal Dynamics of Land Use in the Juárez Valley: Urbanization and Displacement of Agriculture

Abstract

Urbanization and industrial development in the Juárez Valley, Chihuahua, Mexico, have led to the abandonment and loss of productive agricultural areas. However, the extent and dynamics of this phenomenon are not precisely known due to the lack of updated information. Therefore, it is necessary to geospatially represent these changes over time and predict their probability of persistence into the future to provide decision-making tools for this border region of Mexico. Landsat images were processed, and random forest was applied as a classifier to obtain land uses from 1980 to 2020. The Land Change Modeler options in Terrset™ were executed to generate land use changes, persistence and probabilities. Results showed that urban, built-up areas gained 19,962 ha by 2020 while crops lost 1675 ha. Agricultural permanence has been consolidated over time (persistence until 2020 of 0.83), but evidence suggests that this persistence will decrease in the future due to urbanization (decreasing to 0.59 by 2100). This could jeopardize the availability of primary products and food, lead to land abandonment and exacerbate socio-demographic expansion in this vulnerable region.

1. Introduction

Since the beginning of space exploration, humans have developed a wide range of remote sensing technologies that enable the capture and representation of information from the surface of the Earth. Currently, there is a constellation of platforms or artificial satellites orbiting the planet for various purposes, including SPOT, TERRA, Landsat missions, COPERNICUS, Galileo, MODIS and AQUA, among others [1,2]; these platforms are equipped with sensors that capture information from the Earth and store it in multispectral bands [3], enabling the monitoring of terrestrial space [4]. This facilitates the investigation and updating of historical cartographic documentation of natural and anthropic land cover, as noted by Ancira and Treviño [5], Aquino [6] and Persaud and Milián [7]; furthermore, González et al. [8] and Leija et al. [9] indicated that analyzing spatio-temporal dynamics can help determine the events that influence a change from one type of cover to another.

Esquivel et al. [10], Davila et al. [11] and Qu et al. [12] used GIS and satellite remote sensing to generate cartography of land use and the processes it has undergone over time. Pérez and Romo [13] analyzed the urban planning of Ciudad Juárez in its expansion process towards the southeastern region using GIS. They noted that the municipal government expanded the urban area towards the east, affecting the agricultural lands of the Juárez Valley. This is consistent with Mártinez et al.’s work [14] to determine the spatio-temporal dynamics of land use coverages using Landsat images. The municipal administration attempted to remedy this problem between 1986 and 1989 by limiting growth towards the Juárez Valley, but expansion continued towards the south.

Esquivel et al. [10] emphasized the importance of knowing the historical and current state of land cover and land use of social interest. They pointed out that the lack of comprehensive territorial planning favors urban expansion with deficiencies in infrastructure and services, particularly in remote or peripheral areas of the city that are constantly expanding, displacing agricultural areas. The irregular urban sprawl in the Juárez Valley is a problem identified by Esquivel et al. [10], Davila et al. [11], Pérez and Romo [13] and Mártinez et al. [14]. Agricultural extensions are threatened by encroaching land uses, which have caused cities to exponentially increase their growth rate since the industrial revolution, resulting in economic and territorial losses, abandonment of plots and invasion of agricultural land use over other coverages such as natural grasslands. These changes have created new ecosystemic problems [15].

The general aim of this study was to use GIS and satellite remote sensing (RS) technologies to address the problem of cartographic updates of land cover between 1980 and 2020 and the historical and current analysis of its losses, gains and permanence. Additionally, the study aimed to perform a prospective analysis towards 2100 of the state of urban and agricultural land cover over the Juárez Valley in Chihuahua, Mexico.

2. Data and Methods

2.1. Study Area

The study area (AOI, area of interest) is located in the Juárez Valley, which is situated in the municipality of Juárez (Figure 1) in the northern state of Chihuahua, Mexico. The area covers 962.08 km², with coordinates ranging from −106.623° to −106.180° longitude and 31.784° to 31.228° latitude [16]. The altitude ranges from 1077 to 1817 m above sea level, with slopes ranging from 0° to 70° in small mountain ranges [17]. The Juárez Valley is located on the Juárez aquifer (key no. 0833), with an area of 3386.44 km² [16].

The main land uses and vegetation in the area are urban, agricultural, pasture, and shrub secondary vegetation. There is rainfed and irrigated agriculture [18]. Cervantes and Montano [19], in their book El Valle de Juárez, mentioned that the main crops are alfalfa, sudan grass, pistachio, corn, wheat, cotton and sorghum, with regional environmental characteristics of average temperatures of 18.2 °C, a minimum of 7.5 °C and a maximum of 28.4 °C and an average annual precipitation of 265.3 mm and a population of 1,391,180 inhabitants as of 2015.

The processes used in this study (Figure 2) utilized satellite images hosted in the Google Earth Engine™ (GEE) data catalog provided in cooperation with the United States Geological Survey (USGS). This cloud-based petabyte geomatics platform allows for the processing, analysis and visualization of imagery from various satellite platforms such as Landsat’s historical legacy platform of ETM+, TM and OLI/TIRS sensors [20].

For the 1980s and 1990s, images were obtained for the Landsat 5 TM Collection Tier 1 mission calibrated for TOA (top of atmosphere) (LANDSAT/LT05/C02/T1_TOA courtesy of the U.S. Geological Survey) [21]. For the 2000s and 2010s, Landsat 7 Collection 2 Tier 1 mission imagery calibrated to TOA (top of atmosphere) (LANDSAT/LE07/C02/T1_TOA courtesy of the U.S. Geological Survey) was used [21]. The technical characteristics of these two platforms include a temporal resolution of 16 days, a multispectral band of 30 m, far infrared and thermal distances of 60 m and a Landsat 7 panchromatic band of 15 m, and their spectral ranges are from 0.45 to 2.35 μm [22]. For the 2020s, images from the Landsat 8 OLI/TIRS Collection 2 Tier 1 mission calibrated for TOA (top of atmosphere) (LANDSAT/LC08/C02/T1_TOA courtesy of the U.S. Geological Survey) were used [21]. The technical characteristics of this platform show a temporal resolution of 16 days operating under the same acquisition program as Landsat 5; the spatial resolution of the multispectral bands is 30 m, the panchromatic band of Landsat 8 is 15 m and its spectral ranges are from 0.443 to 1.375 nanometers [22,23].

Algorithms were developed on the GEE platform to analyze the multitemporal images. Date, boundary and cloud percentage filters were applied, and, subsequently, multispectral bands with spectral and spatial resolution were selected for the development of supervised and unsupervised classifications. Since the scenes available in the platform already have an atmospheric correction, this process was not applied. Date filters covering months with a lower percentage of cloud (<10%) were used next, and the boundaries (AOI) were loaded to the data warehouse and served as a mask file for extraction. Based on the filtered scenes, statistical reducers where the output is calculated per pixel were applied so that each pixel of the output was composed of the mean value of all images in the collection at that location (Figure 3) [20,21,22,23,24].

To obtain a multitemporal image, mosaics were generated. Images that did not comply with an acceptable percentage of cloudiness or coverage of the AOI of the studied regions were excluded, and cloud masking was applied using the quality band [25] to generate better-quality images. For platforms with a panchromatic band, the HSV pansharpening technique [26] was applied to increase the spatial resolution of the multispectral bands from 30 to 15 m.

For the Landsat 7 platform, the ETM+ sensor has an error from 2003 which causes banding or parallel lines and the absence of some information due to radiometric errors. To correct this drawback, interpolation and clustering techniques were applied following what was suggested by the USGS and EROS in 2004 [27].

A classification based only on spectral variables can show confusion among objects with similar spectral responses. Spectral indices [27,28] of vegetation, soil, water or even indices based on RGB (red, green and blue) [29] were used to discriminate pixels that could conflict in their spectral response. Among the indices used (

Table 1. List of indexes calculated for multispectral and RGB images. Source: own elaboration based on USGS (2022) [27], Santos et al. (2022) [28], Roque (2021) [29] and the index database documented by IDB project (2011–2023) available at: https://www.indexdatabase.de/db/ia.php (accessed on 20 January 2023).

Index	Meaning
NDVI	Normalized Difference Vegetation Index
GNDVI	Green Normalized Difference Vegetation Index
EVI	Enhanced Vegetation Index
AVI	Advanced Vegetation Index
SAVI	Soil Adjusted Vegetation Index
NDMI	Normalized Difference Moisture Index
MSI	Moisture Stress Index
GCI	Green Coverage Index
BSI	Bare Soil Index
NDWI	Normalized Difference Water Index
NDSI	Normalized Difference Snow Index
VARI	Visible Atmospherically Resistant Index
NGRDI	Normalized Difference Green/Red Index
GLI	Green Leal Index
VDVI	Difference vegetation index

) to discriminate the pixels that could conflict in their spectral response, NDVI, GNDVI, EVI, AVI, SAVI, NDMI, MSI, GCI, BSI, NDWI, NDSI, NDSI and VARI were calculated by map algebra using the corresponding bands. The wavelength covered by each band per satellite platform varies; for example, it is not the same wavelength for Landsat 5 band 1 (blue) as for Landsat 8 band 1 (coastal aerosol). For the RGB indices, NGRDI, GLI and VDVI were calculated according to what Poley and McDermin mentioned [30].

Physiographic variables were generated to support the classifier and indicate to the pixel its topographic or geographic position in the terrain [31]. This information helps the classifier to recategorize a pixel in relation to its neighboring pixels by the value of its topographic position if it has the same spectral response and value in an index. A DEM (digital elevation model) obtained from the NASA-USGS (USGS/SRTMGL1_003 courtesy of the U.S. Geological Survey) was used for this purpose. This is a SRTM (shuttle radar topography mission) [32] product with a spatial resolution of 30 m. The slopes and slope orientation variable were generated in degrees [33].

As the first training and calibration variables, spectral separability was performed at the pixel level from an unmixing with the end members of each class (

Table 2. Structure of a matrix for OBIA and spectral separability. Source: own elaboration.

Array of Values
Object	Variable 1	Variable 2	Variable n
Id 1	$a_{i1j1}$	$a_{i1j2}$	$a_{i1jn}$
Id 2	$a_{i2j1}$	$a_{i2j2}$	$a_{i2jn}$
Id n	$a_{i3j1}$	$a_{i3j2}$	$a_{i3jn}$

). A transposed pseudo-inverse matrix was used to calculate the composite image for each class with the same number of bands. This technique is used in OBIA (object-based image analysis) [34,35].

To complement the OBIA processes, a segmentation of the physical and spectral characteristics representative of the objects that may contain the pixels was performed [36]. The composite image was segmented into polygons of objects corresponding to their spectral and index values, shape and texture. Neighborhood groupings were integrated in this process, which are based on spatial contiguity and correspondence. This analysis determines if the behavior of a pixel is similar to its neighbor in its vertical, horizontal and vertex limits [37,38]. The integral efficiency analysis allows for accurate discrimination for the identification of objects and generation of training polygons based on statistics and geospatial behavior.

To train the classification algorithm, training points were generated by photointerpretation and the values corresponding to the spectral signatures that best represented each class or object. The identified classes were Crops, Urban, Sands and Livestock. The machine learning prediction algorithm of random forest [39,40] available in GEE was used. It was executed with a number of decision trees [39] and the training points generated for each year and the variables or bands integrated in a single image.

The validation process consisted of comparing the results of the predictions or classifications against a set of independent data and calculating the kappa index that measures the level of agreement. This index allows us to know the degree to which the observers have certainty in their measurements or predictions [41]. The results of the kappa index (Equation (1)) are obtained by subtracting a proposed theoretical contribution given to randomness (Equation (3)) from reliability. This yields a result that is less than the overall reliability (Equation (2)). For categorical mapping, a significant percentage of the random map should be correct for user agreement [42].

$$Kappa=\frac{P_o-P_c}{1-P_c}$$

(1)

$$Po=\sum _{k=1}^n{P}_{kk}$$

(2)

$$Pc=\sum _{k=1}^n{P}_k+P+{}_k$$

(3)

where Po = proportion of the area correctly classified (overall reliability), and Pc = reliability resulting at random. Source: own elaboration based on Mas et al. [42].

Table 3. Basic structure to generate a confusion matrix for the calculation of the Kappa index. Source: own elaboration based on Mas et al. [42].

Georeferenced Sites (GCP)
	Category	Class 1	Class 2	Class n	Total
Map (Classification)	Class 1	$P_{i1j1}$	$P_{i1j2}$	$P_{i1jn}$	$P_{i1}+$
	Class 2	$P_{i2j1}$	$P_{i2j2}$	$P_{i2jn}$	$P_{i2}+$
	Class n	$P_{inj1}$	$P_{inj2}$	$P_{injn}$	$P_{in}+$
	Total	${P+}_{j1}$	${P+}_{j2}$	${P+}_{jn}$

shows the basic structure of a confusion matrix of size n for the comparison of the georeferenced and map sites.

The results obtained for each period of each year were processed using a desktop GIS. The final layers were generated, and areas in hectares were calculated while isolating the land covers of interest. Formats for the classifications were prepared for their subsequent analysis in image processing software in the Land Use Change Modeler.

The LCM (Land Change Modeler) module was used to obtain the trends of land use change coverages, which allow for the mapping of transitions, permanences, losses and gains. The probability of change was calculated through Markov chains [43,44,45] for the years 2025, 2030, 2050 and 2100 using 1980 as the base year and 2020 as the final and starting consecutive years to be predicted.

To generate the models, the input parameters were the results of each year with a normalized legend by ID and category. The first LCM result was calculated by overlapping the pixels of each category of the initial year against that of the end year. The units shown were indicated in pixel units, hectares or others.

From these options, cartographic results of losses, gains, persistence and transitions were constructed for each analyzed period and the historical period [46,47]. Finally, with the information stored in each of the models, tables of Markov chains or probability of future change were obtained [48,49].

2.2. Statistical Analysis

The GEE platform was used to call collections by their identifier corresponding to the Landsat 5, 7 and 8 platforms and filtered by dates and area of interest. For 1980, it was a collection from 22 August 1982 to 1 May 1985, depending on image availability and cloud percentage (<10%). The multispectral bands to be used were selected, and this process was repeated for the other platforms, modifying the date filter. For 1990, it was from 1 January to 31 December 1990. For 2000, 2010 and 2020, a mosaic was generated, and a statistical reducer was applied to the median.

The quality of the filtered images was improved by optimizing the value of the pixels through a masking operation, selecting the quality statistics band and a bit-by-bit operation [50]. The HSV technique of pansharpening, used by Poveda et al. [26], was applied to the platforms that had a panchromatic band of 15 m spatial resolution. The necessary bands to construct the spectral indices were selected, as indicated by Santos et al. [28] and Roque [29]. The equations of the University of Bonn portal were accessed and solved by means of an expression in map algebra. The necessary bands to construct the spectral indices were selected [31], and slope variables in degrees and slope orientation [33] were constructed using the SRTM DEM.

For the OBIA spectral separability analysis, the spectral bands of the image, indexes and physiographic variables were integrated into a single, multiband image according to Ela and Claire [34] and Ramandhan et al. [35]. The SNIC (simple non-iterative clustering) algorithm was used for segmentation, and a reducer was applied to the clustering band [36,37,38].

In the final stage of the random forest (RF) classifier, the arguments from Phan et al. [39] and Amini et al. [40] were incorporated. An image composed of the bands obtained by unmixing, segmentation, physiographic and multispectral bands, their respective labels of each band with 500 decision trees [39] and the series of training points generated from a base sampling taken from Mas et al. [42] were used. This is a hybrid method combining systematic and clustering methods (MH). Ground truth points were obtained through photointerpretation and expert knowledge, representing each category of land use and vegetation analyzed.

The land use classifications for each year were validated using the kappa index and INEGI cartography scale 1:250,000, as indicated by Mas et al. [42]. Values were extracted from all series that were constant in their categories between 1980 and 2020. The sampling used the MH method, generating a 2000 × 2000 m grid with points created on each line at a distance of 15 m (133 points), to which points by clusters were added [42]. These were created within each polygon of the grid, where the minimum distance between them or the area of influence was 500 m so that each category or class within the quadrant had the same probability of being evaluated.

To run the LCM, classifications were spatially adjusted by resampling, ensuring that the spatial configuration was homogeneous in pixel size and number of columns and rows These were converted to ASCCI format for export to Terrset™ software and added to the soil change modeling processes in the LCM module, as mentioned by Anand and Ainam [43], Hamad et al. [44] and Sundara et al. [45]. Models were run for each decade between 1980 and 2020, and predictions of changes for the dates of 2025, 2030, 2050 and 2100 were obtained, feeding the processes of gains, losses and transitions [46,47] to obtain the net contributions through Markov chains [48,49].

3. Results

Figure 4 displays the land use and vegetation coverages represented before being generalized for the years 1980 and 2020. Urban land use is represented in black, built to the northwest of the AOI, and crops are shown in light green on the border strip of the Rio Bravo.

The coverages were generalized into four classes: crops (crop, which integrates agricultural crops and walnut), built urban, sands (desert), and livestock (forest—forest, pasture land—pasture, scrub—scrub, association PlSc—association of pasture and scrub and water body—water bodies of small watering places).

Table 4. Hectares of generalized classes for the years 1980 and 2020. Source: own elaboration based on the results obtained from Figure 4.

Class/Year	ha per Classes
	1980	2020
Crop	6737.42	5027.06
Built urban	1617.70	26,485.40
Sands	54.59	2189.58
Livestock	87,796.50	67,308.20

shows the calculated areas in hectares for each of these classes for the years 1980 and 2020.

Table 5. Changes and persistence between classes for the period from 1980 to 2020 (C = changes, P = persistence). Source: own elaboration based on the results of the generalization of coverages to classes.

Changes and Persistence in ha between 1980 to 2020
Class	ID Class 1980	Hectares in 1980	ID Class 2020	Hectares of Change to 2020	Percentage of Change from 1980 to 2020	C/P
Crop	1	6561.49	1	5018.06	76.48	P
	1	6561.49	2	646.84	9.86	C
	1	6561.49	3	6.56	0.10	C
	1	6561.49	4	887.89	13.53	C
Built urban	2	2373.66	1	124.53	5.25	C
	2	2373.66	2	1841.50	77.58	P
	2	2373.66	3	8.34	0.35	C
	2	2373.66	4	398.27	16.78	C
Sands	3	NA	1	NA	NA	C
	3	54.00	2	0.01	0.01	C
	3	54.00	3	7.27	13.46	P
	3	54.00	4	46.73	86.53	C
Livestock	4	87,217.80	1	931.48	1.07	C
	4	87,217.80	2	3804.28	4.36	C
	4	87,217.80	3	853.34	0.98	C
	4	87,217.80	4	81,626.21	93.59	P

displays the changes and persistence in hectares for the historical period from 1980 to 2020.

Figure 5 shows the classifications for the years 1990, 2000 and 2010 before they were grouped or generalized by classes.

Table 6. Area in hectares of generalized classes for the years 1980 and 2020. Source: own elaboration.

Class/year	ha per Classes
	1990	2000	2010
Crop	6561.49	6075.92	3400.46
Built urban	2373.66	6294.20	22,752.61
Sands	54.00	875.56	3778.39
Livestock	87,217.80	82,960.53	66,275.40

displays the areas in hectares for each of the classes for the years 1980 and 2020 based on the information in Figure 4.

Table 7. Cross-tabulation for the years 1990, 2000 and 2010. Source: own elaboration based on the results of the generalization of coverage to classes.

Changes and Persistence in ha between 1990 to 2000 and to 2010
Class	ID Class 1990	Hectares in 1990	ID Class 2000	Hectares of Change to 2000	Percentage of Change from 1990 to 2000	C/P	ID Class 2000	Hectares in 2000	ID Class 2010	Hectares of Change to 2010	Percentage of Change from 1980 to 2010	C/P
Crop	1	6561.49	1	5018.06	76.48	P	1	6294.20	1	146.62	2.33	P
	1	6561.49	2	646.84	9.86	C	1	6294.20	2	5621.12	89.31	C
	1	6561.49	3	6.56	0.10	C	1	6294.20	3	8.37	0.13	C
	1	6561.49	4	887.89	13.53	C	1	6294.20	4	516.36	8.20	C
Built urban	2	2373.66	1	124.53	5.25	C	2	6075.92	1	2801.92	46.12	C
	2	2373.66	2	1841.50	77.58	P	2	6075.92	2	1861.07	30.63	P
	2	2373.66	3	8.34	0.35	C	2	6075.92	3	14.46	0.24	C
	2	2373.66	4	398.27	16.78	C	2	6075.92	4	1394.79	22.96	C
Sands	3	NA	1	NA	NA	C	3	875.56	1	4.35	0.50	C
	3	54.00	2	0.01	0.01	C	3	875.56	2	105.02	11.99	C
	3	54.00	3	7.27	13.46	P	3	875.56	3	511.52	58.42	P
	3	54.00	4	46.73	86.53	C	3	875.56	4	254.55	29.07	C
Livestock	4	87,217.80	1	931.48	1.07	C	4	82,960.53	1	444.60	0.54	C
	4	87,217.80	2	3804.28	4.36	C	4	82,960.53	2	15,162.12	18.28	C
	4	87,217.80	3	853.34	0.98	C	4	82,960.53	3	3243.59	3.91	C
	4	87,217.80	4	81,626.21	93.59	P	4	82,960.53	4	64,103.62	77.27	P

displays the cross-tabulation for these years, showing the area calculated in hectares for the change and persistence and the percentage of each of the categories in these years.

Figure 6 displays the years 1980, 1990, 2000, 2010 and 2020 with their four classes, where the green shades identify the cultivated areas, the black shades identify the built-up urban areas, the yellow shades identify the sandy areas and the brown shades identify the livestock areas.

Figure 7 shows the multitemporal geospatial dynamics in hectares of change for the classes. It shows that the urban class experienced high growth demand over the other classes.

Table 8. Matrix with class values for points extracted from INEGI’s land use and vegetation series (GCP) and classifications with generalized classes for the years 1980, 1990, 2000, 2010 and 2020.

GCP Persistent in the Series	Class	Classification 1980					Class	Classification 1980
		Crop	Built Urban	Sands	Livestock	Total		Crop	Built Urban	Sands	Livestock	Total
	Crop	3635	103	0	198	3936	Crop	0.0521	0.0015	0.0000	0.0028	0.0564
	Built urban	2	3247	19	400	3668	Built urban	0.0000	0.0465	0.0003	0.0057	0.0525
	Sands	0	0	1	24	25	Sands	0.0000	0.0000	0.0000	0.0003	0.0004
	Livestock	1627	2963	1	62,924	67,515	Livestock	0.0233	0.0424	0.0000	0.9014	0.9672
	Total	5264	6313	21	63,546	69,807	Total	0.0754	0.0904	0.0003	0.9103	1.0000
GCP INEGI Serie I	Class	Classification 1990					Class	Classification 1990
		Crop	Built Urban	Sands	Livestock	Total		Crop	Built Urban	Sands	Livestock	Total
	Crop	4426	697	0	2444	7567	Crop	0.0649	0.0102	0.0000	0.0358	0.1110
	Built urban	2	813	0	3197	4012	Built urban	0.0000	0.0119	0.0000	0.0469	0.0588
	Sands	0	0	25	1147	1172	Sands	0.0000	0.0000	0.0004	0.0168	0.0172
	Livestock	308	409	25	62,934	63,676	Livestock	0.0045	0.0060	0.0004	0.9228	0.9337
	Total	4736	1919	50	69,722	68,198	Total	0.0694	0.0281	0.0007	1.0223	1.0000
GCP INEGI Serie II	Class	Classification 2000					Class	Classification 2000
		Crop	Built Urban	Sands	Livestock	Total		Crop	Built Urban	Sands	Livestock	Total
	Crop	3970	822	11	956	5759	Crop	0.0673	0.0139	0.0002	0.0162	0.0976
	Built urban	407	4057	37	9493	13,994	Built urban	0.0069	0.0688	0.0006	0.1609	0.2372
	Sands	12	20	336	4319	4687	Sands	0.0002	0.0003	0.0057	0.0732	0.0795
	Livestock	332	156	244	50,626	51,358	Livestock	0.0056	0.0026	0.0041	0.8582	0.8706
	Total	4721	5055	628	65,394	58,989	Total	0.0800	0.0857	0.0106	1.1086	1.0000
GCP INEGI Serie IV	Class	Classification 2010					Class	Classification 2010
		Crop	Built Urban	Sands	Livestock	Total		Crop	Built Urban	Sands	Livestock	Total
	Crop	1857	359	17	298	2531	Crop	0.0288	0.0056	0.0003	0.0046	0.0393
	Built urban	580	16,633	810	5623	23,646	Built urban	0.0090	0.2584	0.0126	0.0874	0.3673
	Sands	16	43	624	1069	1752	Sands	0.0002	0.0007	0.0097	0.0166	0.0272
	Livestock	162	1219	1279	45,258	47,918	Livestock	0.0025	0.0189	0.0199	0.7031	0.7444
	Total	2615	18,254	2730	52,248	64,372	Total	0.0406	0.2836	0.0424	0.8117	1.0000
GCP INEGI Serie VI	Class	Classification 2020					Class	Classification 2020
		Crop	Built Urban	Sands	Livestock	Total		Crop	Built Urban	Sands	Livestock	Total
	Crop	2115	44	2	315	2476	Crop	0.0324	0.0007	0.0000	0.0048	0.0379
	Built urban	1157	16,929	480	6450	25,016	Built urban	0.0177	0.2594	0.0074	0.0988	0.3833
	Sands	0	8	298	376	682	Sands	0.0000	0.0001	0.0046	0.0058	0.0104
	Livestock	317	541	756	45,923	47,537	Livestock	0.0049	0.0083	0.0116	0.7036	0.7284
	Total	3589	17,522	1536	53,064	65,265	Total	0.0550	0.2685	0.0235	0.8131	1.0000

displays the values of the validation matrix and kappa index for each validation year. For overall reliability, the lowest value was 0.72 for the year 2000, and the highest value was 0.92 for 1980. For random reliability, the lowest value was 0.70 for 2020 and the highest 0.99 for 2000.

The values for kappa ranged from 0.44 for the year 2000 to 0.71 for 2020 (

Table 9. Values of the global and random reliability indicators and kappa index for each validation year. Source: own elaboration based on sampling and information from INEGI’s land use and vegetation series.

Indicators	Year
	1980	1990	2000	2010	2020
Overall reliability	0.92	0.88	0.72	0.82	0.84
Random reliability	0.89	0.96	0.99	0.71	0.70
Í. Kappa	0.69	0.54	0.44	0.69	0.71

).

From the validated mapping of the classifications generated, grouping land cover or land use and vegetation into classes, it was possible to obtain the gains, losses and net contributions of the classes for the years 1980, 1990, 2000, 2010 and 2020.

Table 10. Net losses, gains and contributions for each class from 1980 to 2020. Source: own elaboration.

Period 1980–2020
Class	Loss (ha)	Gain (ha)	Loss–Gain (ha)
Crop	−3225	1550	1675
Built urban	−204	20,166	−19,962
Sands	−51	2222	−2171
Livestock	−22,040	1582	20,458

Note: negative values in the third column indicate a gain. Source: own elaboration based on LCM processes.

and

Table 11. Contributions in hectares between net profits and losses for each class in the period 1980–2020.

Period 1980–2020
Class	Crop	Built Urban	Sands	Livestock
	Net Contributions (ha)
Crop	0	1733	7	−65
Built urban	−1733	0	−23	−18,206
Sands	−7	23	0	−2186
Livestock	65	18,206	2186	0
Net loss	1740	0	23	20,457
Net gain	65	19,962	2193	0

Note: Negative values indicate a negative contribution to the class analyzed. Source: own elaboration based on LCM processes.

shows the results in hectares for the historical period from 1980 to 2020.

For the contributions in net hectares of gains or losses between classes for the period analyzed, a cartographic representation is shown for the net changes between classes (Figure 8, Figure 9 and Figure 10) in Table 6 and Table 7 for the period from 1980 to 2020.

Figure 9 shows the class persistencies for the historical period from 1980 to 2020, and Figure 10 identifies the agricultural or crop gains, losses and persistencies for the same period.

Table 12. Net losses, gains and contributions for each class between 1980 and 1990, 1990 and 2000, 2000 and 2010 and 2010 and 2020 (negative values in the third column indicate a gain): Source: prepared by the authors based on LCM processes.

Period 1980–1990				Period 1990–2000
Class	Loss (ha)	Gain (ha)	Loss–Gain (ha)	Class	Loss (ha)	Gain (ha)	Loss–Gain (ha)
Crop	−1426	1258	168	Crop	−1555	1080	475
Built urban	−726	1495	−769	Built urban	−558	4459	−3901
Sands	−52	50	2	Sands	−47	901	−854
Livestock	−2539	1939	600	Livestock	−5644	1363	4281
Period 2000–2010				Period 2010–2020
Class	Loss (ha)	Gain (ha)	Loss–Gain (ha)	Class	Loss (ha)	Gain (ha)	Loss–Gain (ha)
Crop	−3287	610	2677	Crop	−508	2149	−1641
Built urban	−703	17,093	−16,390	Built urban	−4306	3212	1094
Sands	−377	3293	−2916	Sands	−2305	707	1598
Livestock	−18,847	2218	16,629	Livestock	−4128	5179	−1051

shows the net losses, gains and contributions in hectares.

Table 13. Losses, gains and net changes for each class between the 1980s and 2020s.

Period 1980–1990					Period 1990–2000
Class	Net Contributions (Ha) Per Class				Class	Net Contributions (ha) per Class
	Crop	Built Urban	Sands	Livestock		Crop	Built Urban	Sands	Livestock
Crop	0	69	0	99	Crop	0	522	7	−55
Built urban	−69	0	−8	−692	Built urban	−522	0	9	−3388
Sands	0	8	0	−7	Sands	−7	−9	0	−838
Livestock	−99	692	7	0	Livestock	55	3388	838	0
Net loss	168	0	8	699	Net loss	529	9	0	4281
Net gain	0	769	7	99	Net gain	55	3910	854	0
Period 2000–2010					Period 2010–2020
Class	Net contributions (ha) per Class				Class	Net contributions (ha) per Class
	Crop	Built urban	Sands	Livestock		Crop	Built urban	Sands	Livestock
Crop	0	1709	10	957	Crop	0	−1250	−8	−383
Built urban	−1709	0	−100	−14,581	Built urban	1250	0	−405	249
Sands	−10	100	0	−3005	Sands	8	405	0	1185
Livestock	−957	14,581	3005	0	Livestock	383	−249	−1185	0
Net loss	2676	0	−100	17,586	Net loss	0	1499	1598	383
Net gain	0	16390	3015	957	Net gain	1641	405	0	1434

Note: negative values in the last column indicate a gain. Source: prepared by the authors based on the results of the LCM module of phase 3 and the classes of the generalized classifications.

shows the contributions in hectares between net losses and gains for each class between 1980 and 1990, 1990 and 2000, 2000 and 2010 and 2010 and 2020.

Finally,

Table 14. Markov chain matrix showing the probabilities of persistence (green) and changes (red) between classes based on the behavior of the historical period from 1980 to 2020 and from there to 2025, 2030, 2050 and 2100. Source: own elaboration with the results of the LCM analysis.

Probability of Change to the Years Calculated from 2020
LCM for Markov Chains from 1980 to 2020 into the future	Probability of Change from 2020 to 2025					Probability of Change from 2020 to 2050
	Clase	Crop	Urban	Sands	Livestock	Clase	Crop	Urban	Sands	Livestock
	Crop	0.83	0.08	0.00	0.09	Crop	0.60	0.22	0.00	0.18
	Urban	0.02	0.97	0.00	0.02	Urban	0.03	0.90	0.00	0.07
	Sands	0.01	0.36	0.23	0.40	Sands	0.02	0.44	0.13	0.40
	Livestock	0.00	0.04	0.03	0.93	Livestock	0.01	0.16	0.03	0.80
	Probability of Change from 2020 to 2030					Probability of Change from 2020 to 2100
	Clase	Crop	Urban	Sands	Livestock	Clase	Crop	Urban	Sands	Livestock
	Crop	0.78	0.11	0.00	0.11	Crop	0.28	0.42	0.01	0.21
	Urban	0.02	0.95	0.00	0.03	Urban	0.05	0.80	0.05	0.15
	Sands	0.01	0.37	0.21	0.40	Sands	0.04	0.55	0.02	0.39
	Livestock	0.00	0.07	0.03	0.90	Livestock	0.03	0.36	0.02	0.59

displays the prediction results of the Markov chains, which are change probability matrices generated from the historical period analyzed with the start year of 1980 and 2020 as the end year, and this result was taken as a starting point for the predictions for each of the years 2025, 2030, 2050 and 2100.

4. Discussion

In this study, geospatial techniques and tools in geographic information systems (GIS) and remote sensing (RS) were utilized to obtain historical land use information from the Juárez Valley. The results demonstrate that these technological trends allow for the obtaining of quality and certainty of results, which is consistent with previous studies by Su et al. [51] and Aguilar et al. [52] and Osuna et al. [53] and Figueredo and Rámon [54].

The increase in the consolidated or built urban area can be clearly observed in Figure 3 (historical period 1980 to 2020) and the intervals by decade (Figure 4 and Figure 5), which coincides with the analysis of Esquivel et al. [10], who analyzed this growth in terms of the impact on land use cover and its effects on the urban aquifer. This study determined that the areas most affected by urban sprawl growth were those corresponding to livestock land use (grassland–scrubland associations) that were found on the periphery of the urban sprawl for each decade, which is consistent with Fuentes [55] and the results of Graph 1 and Figure 4.

It was also detected that the agricultural and crop areas that developed towards the eastern region in the initial decade from 1980 to 2020 were consolidated as built urban areas, which was confirmed by Quijada and Chávez [56], indicating that this phenomenon has been registered since 1940, expanding the urban over flat areas, such as the agricultural areas located to the west of the city and which is reiterated in Pérez and Romo [13], an effect that is noted in the results of Figure 3, Figure 4 and Figure 5.

For the year 1980, between 6561.49 and 6737.42 ha of crop area were counted, while, for the year 2020, 5018.06 and 5027.06 ha were detected, with a loss of 1710.36 ha. This last result coincides with the problems mentioned by Martínez et al. [14] and was located and delimited by the municipal administration in 1986–1989. The previous result coincides with what was obtained and is framed cartographically in Figure 6 with the legend in red tones, where the change of crops towards the built urban in the historical period is shown. The net loss of crops is identified in Table 9, showing that it was a value of 1675 ha, while the net gain over all urban land uses was 19,962 ha between 1980 and 2020, where crops had a net contribution to the urban built-up area of 1773 ha in the historical period; this type of dynamic was also mentioned by Martínez et al. [14] in a region close to the Juárez Valley, that is, over the Rio Conchos basin.

Based on Table 13, the results of Figure 7 and Figure 8 (losses, gains and persistence and their respective tables) are supported, where the agricultural land use indicated 0.83 persistence towards 2025, with probabilities less than 0.01 of change to other classes. By 2030, these increased to 0.11 for the urban and livestock classes, while its permanence decreased to 0.78. This decrease was reflected in the year 2050, where its persistence resided at 0.60, and, for the changes, it indicated a probability of 0.22 for the urban class and 0.18 for livestock. For 2100, these continued to increase, from agricultural to urban with a change of 0.42 and 0.21 from agricultural to livestock, with the persistence of agricultural having a value of 0.28. These results coincide with those mentioned by Ramos et al. [57], who, although they did not work on the same area, obtained similar results of probability of land use change to urban by 2030.

The results of this study coincide with the hypothesis of the work of Pérez and Romo [13], where they indicated the relevance of urban growth in the borders since the 1960s, adding infrastructure planning problems, specifically in the southeastern area of Ciudad Juárez, where it is also indicated that this expansion/urbanization and effect are applied to agricultural or crop land uses.

To explain the permanence and consolidation of agricultural and cultivation plots in the study area, reference can be made to the findings of Ascencio et al. [58], which indicated that some relevant factors for this phenomenon may be the means of production, the characteristics of these and the valuation and development of the land, as well as the integration of productive systems. Therefore, this type of study opens a panorama to continue analyzing the study area and identify the processes that allow permanence to be offered to this type of productive system, which is of great importance for the population.

5. Conclusions

The objectives of this study were successfully achieved by employing geographic information systems and the use of satellite products, which proved to be a powerful support tool for monitoring land use changes in border regions such as the Juárez Valley, where the socio-economic phenomena associated with urbanization in recent years are significant and unfortunately have captured the world’s attention. The first objective was to generate the historical cartography—updated by decade (1980–2020)—of the main land uses and coverages for the region of the Juárez Valley in Chihuahua, Mexico, which was accomplished. From this, the second objective was fulfilled by calculating the areas of gains, losses and transitions and their respective mapping. It was found that the greatest growth occurred in the decade from 1990 to 2000 and in 2010, evidencing continuous urban growth since 1980 in areas of livestock use (associations of pasture–scrubland) to the south and west in agricultural regions.

The study showed that agricultural land use was consolidated, with plots that were not initially productive being grouped within the same zone and cultivated. Additionally, the border between agriculture and livestock increased towards the south, with the latter being adapted as a region of agricultural production. The third objective relating to the probabilities of persistence of the agricultural productive region in the future was fulfilled and indicated that, in the years analyzed, this will decrease. Urbanization is the main land use that can influence this type of loss, which represents a possible productive decrease, thus affecting the availability of primary products and food and income of the producers and can lead to land abandonment and problems for socio-demographic expansion.

Given what was generated and obtained in this study, it is advisable to continue analyzing the dynamics of land use changes, characterizing the activities that are being developed in the region and encouraging the collection of field information for better classification and validation of the information analyzed. Therefore, greater financing of this type of study is demanded by institutions involved in the planning and managing of territorial development so that they can benefit from more and better tools for decision making and solving problems associated with the phenomenon of urbanization, loss of productive activities and the degradation of natural resources in these border areas of developing countries such as Mexico.