Cloud Computing Application for the Analysis of Land Use and Land Cover Changes in Dry Forests of Peru

Dry forests cover 20% of the Earth’s surface, which in turn account for 30% of global productivity [1, 2] They are responsible for capturing atmospheric carbon (CO₂) from biomass and soil, in addition to harboring biodiversity [3, 4], as well as helping to maintain the hydrological cycle and soil conservation [5-7]. They enable climate regulation, conservation of flora and fauna species, and provide raw materials for construction, food and medicines [8, 9]. However, in recent years, forest ecosystems are being impacted by increasing or decreasing temperature and precipitation, changes in land use and forest degradation [10]. These disturbances are often induced by the population settled in this ecosystem, which makes it vulnerable to droughts and fires with a severe biomass reduction process [11].

In the last 60 years (1960-2019), 32% of vegetation cover has been lost due to LULC changes [12], modifying the structure, functionality of forests and loss of biodiversity habitats [9]. It also increases greenhouse gas (GHG) emissions [13]. LULC changes are the main cause of forest fragmentation for the installation of crop plots, pastures and urban growth [14]. It also affects freshwater availability and conservation of natural resources [15]. Loss of soil fertility, water pollution and droughts are related to forest degradation [16]. Therefore, the analysis of forest loss should focus on its socioeconomic uses and landscape dynamics [17, 18].

Analysis of LULC changes has been used as an important tool in the multitemporal analysis of ecosystems, the implementation of policies and strategies for sustainable development [19, 20]. RS allows detecting and spatially analyzing the Spatio-temporal dynamics of LULC using different sensors and techniques [21-24]. Time series of Landsat and Sentinel-2 (S2) images have been used to identify LULC types [24-27]. Similarly, ML models and cloud computing have been applied to analyze LULC changes and map forests accurately and in near real-time [25]. Other studies applied supervised classification by applying RF due to its robustness and overcoming data noise overfitting [28], and it has been widely used in GEE [29-31]. RF has been applied in processing large volumes of data, outperforming other methods in accuracy, such as single-layer neural networks, decision trees and maximum likelihood [29, 32].

In Peru, we find the dry forest ecosystem with biological and cultural richness, scenic beauty and high endemic value [33]. In recent decades, it has experienced biodiversity loss processes due to anthropogenic activities (extractive forestry activities, agriculture and urban expansion) and climatic conditions such as high temperatures, extreme dryness, irregular occurrence of heavy rainfall and the presence of the El Niño-Southern Oscillation (ENSO) [33, 34]. Although previous works mapped the current use in the dry forest, however, the availability of methodologies and cartography is limited for this study area, which could hinder the temporal analysis of this ecosystem [35, 36]. Likewise, there is no information available related to the impacts of LULC changes in natural protected areas. Therefore, in this study we analyze the dynamics of LULC using S2 data and cloud processing throughout the Peruvian dry forest ecosystem. This will provide baseline information on areas with higher dynamics or forest loss that will be potential areas for the development of ecological recovery and restoration projects.

The dry forests of Peru [37], extend along the northern coastal zone, through the departments of La Libertad, Ancash, Lambayeque, Piura, Tumbes and Cajamarca, covering a coastal strip of between 100 to 150 km, with an altitude of up to 1000 m a.s.l. [35]. This forest covers 3.6 Mha, which represents 4.7% of the total forest in Peru [38]. It is characterized by an annual rainfall of 30 to 300 mm between December and March and a mean annual temperature of 23 °C (Figure 1).

The vegetation cover is characterized by being heterogeneous, with trees, shrubs and grasslands that are part of the dry forest [35]. This ecosystem harbors a diversity of forest species with canopy heights of up to 12 meters, which allows the vegetative growth of shrubs and trees [39]. Among the species that inhabit this ecosystem, it is possible to find carob Neltuma pallida, Neltuma limensis, Vachellia macracantha, Vachellia aroma, Colicodendron scabridum, Anonna spp. and Inga spp [40]. In addition, the dry forest of Peru also harbors animal species such as Lycalopex sechurae, Furnarius cinnamomeus, Mazama americana, Iguana delicatissima, Tremarctos ornatus and Penelope albipennis [38]. In the study area, land use is conditioned by anthropogenic activities (agriculture, livestock and urban growth) [24]. While vegetation depends on rainfall during the year [41].

Dry forests are of great economic importance as they provide ecosystem services such as fruits, firewood and fertilizers to the communities settled within the ecosystem. Dry forests are also used for subsistence agriculture and livestock raising, contributing to the food security of these communities. These ecosystems also host important archeological and cultural sites for tourism that help diversify local sources of income and promote the conservation of these natural and cultural environments.

Figure 2 shows process to evaluate the LULC change and its impact on the dry forest of Peru. The construction of time series of S2 images was carried out, then the extraction of clouds and cloud shadows was applied. We then compute the spectral indices and perform the RF classification using training data. Finally, the precision of the generated cartography was calculated.

During the second and third week of June 2022, field trips were conducted to collect data (training and validation) in the dry forest. A GPS navigator (Garmin GPSMAP 64s) and the FocusMap application (https://www.locusmap.app/) were used to georeferenced the LULC classes and generate photographic records [42]. The LULC classes, were represented by a) Open dry forest (ODF), b) Dense dry forest (DDF), c) Bare land (BL), d) Agricultural land (AL), e) Urban area (UA) and f) Water body (WB) (Figure 3). Twenty thousand pixels were extracted from the field-collected data, representing the six randomly grouped LULC classes [24].

S2 L1C (COPERNICUS/S2) images were used due to their spatial (10 meters) and temporal (6 days) resolution. Images of a year considered <30% cloud and no cloud shadow for 2017 and 2021 [43] were selected by using the quality band (QA60). Soil-adjusted Vegetation Index (SAVI), Enhanced Vegetation Index (EVI), Normalized difference Vegetation Index (NDVI), and Normalized Difference Water Index (NDWI) were included to increase the predictor variables for LULC classification. Image processing was performed from the GEE platform [44].

The RF model was used due to its high performance to calculate a set of time series to analyses the time series [28, 45]. The RF has been applied in several studies [28, 46] and has proven to be an excellent classifier in coastal areas [47]. For this purpose, we created multiband image mosaics that included the spectral bands and indices for classification in GEE [48]. The classification results were exported and visually analyzed with high-resolution images in order to improve the classified maps of 2017 and 2021.

The precision was determined based on the confusion (error) matrix technique [49] and 456 validation points that were obtained through the formula established by Cochran [50]. These RS techniques have been widely used [51]. Similarly, we calculated (i) the overall accuracy (OA), (ii) the user’s accuracy (UA), (iii) the producer’s accuracy (PA) and (iv) the Kappa index was used [42, 51, 52]. Additionally, for each class and year 2017 and 2021, the intensity of changes was determined [18]. The loss or gain of each class was determined using cross-tabulation matrices [42, 53]. The annual exchange rate for FAO was calculated using Eq. (1) [54].

This index quantitatively assesses the impact of human actions based on the degree of land use [55]. It is calculated according to the change in LULC compared to the natural state (Eq. (2)) [56]. The higher the degree of land use, the greater the anthropogenic transformation without taking into account the ecological environment [56].

where, $l_a$ is land use degree index; $A_i$ is the rating index of the degree of land use; and $C_i$ is the percentage of the qualified area of the $i$-th land use grade type. In accordance with key studies [56], LULC classes are classified according to Table 1.

The LULC of the dry forest for 2017 and 2021 is shown in Figures 4 and 5. The DDF and ODF covers are the main classes of LULC and representing 39% and 41% of DDF and 22% in ODF in 2017 and 2021, respectively and is distributed mainly in the higher altitude areas. The land area of BL has increased from 29% in 2017 to 31% in 2021 and is distributed in the desert areas of Sechura, Piura and Talara. The proportion of area of the AL class reports a reduction of area from 9% to 6% from 2017 to 2021. The AU class shows an increase, varying from 0.05 to 0.09% in 2017 and 2021, respectively. In turn, the general change of the WB was relatively small and is mainly represented by the surface of rivers located in the study area.

Overall, ODF and BL classes increased significantly, on the other hand, the cultivation area decreased, while the other LULC classes remained unchanged, such as DDF, UA and WB between 2017 and 2021.

The quantitative and spatial changes of the LULC classes were calculated in cross-tabulation matrices, which allowed showing the transformation between the different LULC classes in the dry forest for 2017 and 2021 (Table A1). From a LULC class change intensity perspective, the AU showed an increase in construction area by 2021, which came primarily from BL and ODF land. The BL acreage for 2021 changed to the establishment of new agricultural parcels and new areas of open forest cover. In turn, forest cover (DDF and ODF) showed changes due to the establishment of new agricultural plots, urban areas and soils with little vegetation. In addition, the interaction between both classes DDF and ODF. On the contrary, the area of AL showed a downward trend. Cultivated land changed to BL, ODF and DDF. At a general level between 2017 and 2021, the study area showed changing dynamics. Forest cover, BL and AU gradually increased and AL and WB classes decreased.

The loss of coverage in the dry forest in the evaluation period is mainly concentrated in areas close to urban areas and bodies of water. Between 2017 and 2021, 852.89 km² (2.34%) of forest cover were lost, which changed to crops, urban areas and soils with or without vegetation. Regeneration of forest cover was also reported in approximately 2,273.74 km² (6.24%) (Figure 6a and Table A2). At the level of protected areas (PA), 4,494.81 km² have been conserved so far in the entire dry forest for the regions of Cajamarca, Lambayeque, Tumbes, La Libertad and Piura. However, 110.83 km² (2.47%) lost forest cover (Figure 6b and Table 2).

The effect of human activities on the soil is reflected in the levels of use [57]. The different LULC classes identified in this analysis allowed the calculation of the comprehensive index of the degree of land use for 2017 and 2021. The range in the area varies from 100 to 200 in a similar way in both years, indicating that the areas with the highest land use are concentrated in the north of the study region, with agricultural areas predominating (Figure 7).

In this study, we analyze LULC changes in 2017 and 2021 using cloud computing and RF algorithm. The information generated contributes greatly to generate LULC maps for an important ecosystem in Peru, obtaining OA and Kappa accuracies greater than 89 and 85%, respectively (Table A3 and S3), indicating reasonable and reliable classification results [58]. The dynamics of LULC in the period of analysis reported an increase in the areas of bare land and the dynamics of open and dense dry forest, which could be conditioned by temperature and precipitation [35].

Assessing LULC changes in forest ecosystems is an important tool that helps to multitemporal changes and manage forests with high biodiversity [59]. In the period of analysis, an increase in urban areas was reported, which could be related to urbanization and population growth that demands more and more housing and crop planting [60, 61]. As well as the increase in bare land areas and El Niño phenomenon that impact terrestrial ecosystems and species habitat [62, 63]. This study also reports the reduction of agricultural area, which may be related to the occurrence of ENSO in 2017 that favored agriculture with abundant rainfall and reduced poverty in rural communities by 5% in this ecosystem [64], however, this phenomenon occurs between 3 to 8 years, which conditions agriculture in these areas [39]. The dense dry forest decreased and the open dry forest increased. This reduction may be related to the establishment of new plots for agriculture, urban growth and selective logging [38].

The creation of PA is considered an agent to mitigate deforestation problems and prevent the loss of forest species [65, 66]. At the PA level in the dry forest, it was reported that 92.78% of the conserved area remained unaltered with respect to its natural cover. However, 2.45% of its area lost its forest cover. It is evident that PAs experience a deforestation process both inside and outside their buffer zones [67, 68], with logging being one of the causes of forest loss [69]. Likewise, it has been shown that PA peripheral areas and intangible areas are exposed to deforestation problems [70] due to anthropogenic activities and cattle ranching [67, 71].

The highest degree of intensity is mainly found in the urban and agricultural classes. Cui et al. [58] consider that high values are related to a high degree of anthropogenic impacts, high levels of LULC with flat slopes. On the other hand, the largest area that presented a low degree of land use were the soil classes with or without vegetation, water bodies and forests for being less impacted by human activities and for being on high slopes [55].

The use of the GEE platform offers the ability to process large volumes of data and can be applied for LULC change analysis of large surface areas and crop mapping [25]. A disadvantage when taking images is the problems of cloudiness and the effect of sea swell that can complicate LULC mapping in certain areas, for which atmospheric correction through different algorithms is necessary [72]. In addition, the creation of mosaics with a wider range of dates and complemented with other images such as Landsat, PlanetScope and Sentinel-1A. Special and temporal resolution are other important aspects to be taken into account in LULC studies, the higher the spatial resolution, the greater the LULC detail. While, the temporal resolution will allow obtaining a greater number of images of the same place, which will translate into greater computational power for processing.

The study provides important information as a baseline for monitoring and the formulation of recovery and conservation projects by the competent institutions. The maps present the areas where forest loss was recorded. However, future studies could improve their pressure by using more advanced classification techniques such as image segmentation and deep learning and the use of high-resolution satellite images such as PlanetScope [51]. In addition, future studies need to analyze a larger number of LULC classes and longer time periods to generate multi-temporal historical information [73].

In this research we analysed LULC changes in the dry forest using cloud computing. The LULC maps obtained reported pressures higher than 85%. In the study area, there was a reduction of cultivated land and dense dry forest, while urban areas, dense soils and open dry forest increased significantly. From the perspective of changes in the LULC, it was found that agricultural areas were mainly changed into bare soils and urban areas. On the other hand, natural protected areas showed forest loss, indicating impacts from population and climate change.

The results of this study report the changes of LULC in the dry forest of Peru. This information can be used as a baseline for identifying deforested areas and developing actions for their recovery and conservation by decision makers. In the future, it is important to conduct research that integrates longer time periods and their future prediction using SR techniques, as well as social, economic and environmental aspects to improve the management and conservation of this important dry forest ecosystem.