Using multispectral imagery to monitor urban food forests in a changing climatic environment: a case from Zimbabwe

1 Introduction

The City of Harare Report (Working Paper No. 19) provides a synopsis of the challenges that the city is experiencing, including rapid urbanization and the prevailing political environment. However, while the report acknowledges the significant role that Zimbabwe’s informal economy plays in enhancing poverty reduction, job creation, income generation, livelihoods and food security enhancement, urban farming is only reported as a threat to the environment, while its benefits were not mentioned. In a country where 37% of the urban population lives below the total consumption poverty line of US$70 per person per month (Zimbabwe National Statistics Agency, 2017; Masimba and Walnycki, 2024), and where urbanization and suburbanization are increasing rapidly, poverty and food insecurity are likely to worsen (Kamusoko and Chikati, 2017). A specific research field of urban food systems that recently emerged, whose scope is gaining traction among academics is Urban Food Forestry (UFF), whereby perennial woody food-producing species or food trees are cultivated (Clark and Nicholas, 2013). Clark and Nicholas (2013) argue that UFF could address the challenges of urban food insecurity arising from urbanization and climate change.

However, though UFF could contribute to urban food security, it is not immune to climate change. Therefore, there is a need to investigate the impact of climate change on UFF. It is against this background that the current study is situated. Whereas there is a growing body of scientific literature on the potential of UFF in diversifying ecosystem services (ES) and biodiversity, this literature is mainly from Europe and North America and not from the Global South (Eiden, 2021). Therefore, the research question addressed in this study is: “Can urban forests be used as a tool to enhance adaptation and resilience against climate change?” Recent literature on climate change adaptation and resilience highlights the role of Urban Ecological Infrastructure (UEI) in urban ecosystems management. A critical component of UEI that must be managed is Green UEI, referring to the vegetation that occurs in urban environments. This study focuses primarily on the management of one aspect of Green UEI, namely UFF and its value to sustainable cities. A subsidiary question that is addressed in this study relates to how the impact of climate change on own consumption UFF production systems can be monitored using recently emerged smart technologies in order to enhance the sustainability of urban food systems. The study is prompted by the gaps that exist in urban policy regarding the impact of climate change on UFF food systems in Zimbabwe. Therefore, the objective of this research is to use a case study to investigate how plant health in an urban food forest, where fruit is produced for own household consumption is being impacted by climate change. Hitherto, this research area has received little scholarly attention. The remaining part of this paper is organized as follows. The next section describes the study area. This is followed by the literature review, results and discussion sections, and finally the conclusion section of the paper.

2 Literature review

2.1 Urban food security and climate change

Insights drawn from literature on urban food systems reveal that one of the major threats to urban food security in Africa is rapid urbanization on the continent (Crush and Battersby, 2016). Many of the rural migrants fail to secure employment, hence they are unable to afford food from formal markets. In order to survive they open small fields in their residential areas, where they practice urban farming (Mukwada, 2026). The draconian top-down planning policies that were inherited from colonial administrators in Zimbabwe, where some legislation was closely modeled on British planning law (Marumahoko and Nhede, 2025) prohibited urban agriculture altogether. Historically, some of the most stringent colonial planning laws included the Municipal Act of 1930, Town Planning Act of 1933, Urban Councils Act of 1973, Regional, Town and Country Planning Act of 1976, all of which did not support urban agriculture, but “pathologize” informal activities. However, recent literature has shown that climate change is worsening a problem that is rooted in Zimbabwe’s colonial past.

Though a considerable amount of research has been done elsewhere on the impacts of climate change on urban food security, in the Global South such research has yet to gain inroads. To many poor urban households in the Global South, urban farming has become a climate change adaptation strategy (Dubbeling and De Zeeuw, 2011). Despite this dearth of information, Nkrumah (2018) maintains that Africa’s urban population has a role to play in finding appropriate solutions that can sustain food security in urban environments. While many cities in developed countries already have robust plans for sustainable urban food systems, cities in developing countries are still grappling with their colonial legacies. In industrialized countries, where municipal governments are recasting themselves as food system innovators, responses to urban food insecurity are increasingly emerging at the local level (Sonnino, 2016). For instance, Sonnino (2016) documented 15 food strategies that are already being applied in the USA, UK and Canada.

Unlike in the Global North, urban planners in the Global South have neglected to plan for food systems because it is still “perceived as a rural issue with no connection to the built environment,” yet food insecurity is worse in the latter than in the former (Steenkamp et al., 2021). Effective and viable urban food system planning requires the collection and management of geospatial information. As explained below, this is where remote sensing and Geographic Information Systems (GIS) play a crucial role.

2.2 Importance of multispectral image databases in the study of changing urban environments and food systems

Multispectral medium spatial resolution datasets, such as Landsat images are useful in monitoring complex heterogeneous urban landscapes (Mushore et al., 2017; Mushore et al., 2019). A considerable amount of research has been undertaken in Harare on both climate change and land use and land cover changes (LULC; for instance, work done by Mupedziswa, 2011; Mushore et al., 2022; Kamusoko and Chikati, 2017; Munanga and Mafuku, 2021, among others). Mupedziswa (2011) has concluded that the intensification of climate change over the years has resulted in increased and more intense episodes of pollution, flooding, deforestation, drought and shortage of water, as well as impacted urban farming and worsening of poverty. These challenges can be investigated efficiently through use of remote sensing and GIS.

The ambit of the research done so far leads to the conclusion that climate change is detrimental to the urban environment due to its impacts LULC, including urban agriculture, which is the mainstay of survival for many urban dwellers in the developing world. It can therefore be further concluded that multispectral medium spatial resolution databases could play a valuable role in the investigation of the impact of climate change on evolving urban environments.

One of the most recent developments in remote sensing is the emergence of tools that lessen the burden of analyzing complex urban landscapes on a large scale. One such tool is Google Earth Engine (GEE), which” has become a pivotal tool for large-scale geospatial analysis, with growing applications in urban environmental monitoring” (Stamou and Stylianidis, 2025). The monitoring of urban environments is a critical element of attaining sustainable urban development, as enshrined in the UN Agenda 2030 and Sustainable Development Goals (SDGs). For instance, SDG 2 calls for the total elimination of hunger by 2030. However, the dynamics characterizing rapid urban development are associated with many challenges, including loss of agricultural land and the worsening of urban food insecurity. These dynamics can be analyzed more efficiently by using the GEE platform. For instance, GEE has been applied in environmental monitoring in China to assess and evaluate the liveability and ecological quality of the urban environment (Yang and Su, 2023; Li et al., 2024). For example, using GEE and multi-source data on ecological environmental quality, urban heat island effect and socio-economic benefits, Li et al. (2024) developed an Urban Environment Liveability Index for Kunming City in China. In the Kunming study, areas of remote sensing application were quite varied, ranging from forest and vegetation studies to medical fields such as malaria. The Landsat image collection was the most widely used dataset, while the Landsat image collection is the biggest component of the GEE data portal.

3 The study area

The study area is an own fruit production orchard, an UFF where fruits are largely produced for household subsistence purposes. Its perimeter and area were 70.9 meters and 223 square meters, respectively. The orchard is located in Mt. Pleasant Heights, formerly known as the Bannockburn Farm, which is located along the northern fringes of Harare (Figure 1). Its altitude is approximately 1,490 meters above sea level. The area is characterized by clay loamy soils. Mt. Pleasant Heights is one of Harare’s newer suburbs. It was established in the late 1990s as an extension of the original Mt. Pleasant residential suburb. It comprises two phases, referred to as Phase 1 and Phase 2, with the latter being geographically located along the northern flanks of the suburb. Residential stands in the area are generally large, ranging between 1,300 and 2,300 m². Many properties within the suburb are characterized by vacant spaces used for expansive loans and orchards. The study location is an orchard that is situated on one such stand. A variety of fruit trees were planted in the orchard, including Avocado (Persea americana), Mango (Mangifera indica), Mutohwe (Azanza garckeana), Peach (Prunus persica), Pawpaw (Carica papaya), Pomegranate (Punica granatum), Lemon (Citrus × limon), Egyptian pear, Orange (Citrus × sinensis), Guava (Psidium guajava), True Loquat (Eryobotrya japonica), Apple (Malus domestica), Zumbani (Lippia javanica). Non-wood species that were planted in the orchard include Grapes (Vitis), Banana (Musa paradisiaca) and granadilla (Passiflora edulis). Perennial woody food-producing species or food trees are an important source of food and nutrition to the household that owns it. The fruits are used for subsistence purposes, even though surplus fruit is shared with relatives and friends. However, there is potential for commercialization of produce in the future.

Mean Maximum and mean Minimum temperatures for Harare hover around 29.9 °C and 10. 6 °C, respectively (Magadza and Tendaupenyu, 2025). However, temperatures may exceed 30 °C and fall to levels that are close to zero, in October and June, which constitute the hottest and coldest months of the year, respectively. In Harare mean annual rainfall ranges from 750 mm to 900 mm, while the rain season extends from November to March (Magadza and Tendaupenyu, 2025).

4 Methodology

The first step involved the classification and delineation of the UFF into its constituent components. Figure 2 illustrates the two constituent aerial components that resulted from the unsupervised classification of the UFF, namely Cluster “0” and Cluster “1.” Sentinel 2 data was used in the classification. The clustering was intended to determine the spectral signatures of the dominant plant species in the UFF. These species include Persea americana, Mangifera indica, Prunus persica, Punica granatum, Citrus × limon, Psidium guajava and Eryobotrya japonica. “Zero” and “one” denote the default cluster values appropriated by GEE during the clustering process. The rationale for the clustering was to provide a detailed analysis of the main parts of the UFF that exhibited unique characteristics.

NDVI and NDWI values were then calculated for each cluster, while SPEI values for the clusters were extracted from the SPEI database noted in Table 1. Both NDVI and NDWI are widely considered as suitable for environmental analyses. They were preferred compared to other indices due to their simplicity. NDWI was chosen over alternative indices such as the Normalized Difference Moisture Index, Modified Normalized Difference Water Index, Automated Water Extraction Index and other indices because of the ease with which it can be calculated. Another advantage that it offers is that it offers good reflectance in the Green Band and a strong absorption in Near-Infrared (NIR) Band, which maximizes contrast between water, vegetation and soil. This makes it ideal for assessing UFFs. In addition, NDWI is sensitive to changes in liquid water content of vegetation canopies (Gao, 1996). Similarly, SPEI was chosen as the preferred drought assessment index, compared to the Standard Precipitation Index (SPI), because it is sensitive to both changes in precipitation and temperature.

Table 1. Data types and attributes.

Time-series analyses were performed on the three indices, as well as on Precipitation, Maximum temperature and Minimum temperature data that were derived from Climate Explorer. The analyses covered the period between 2015 and 2023. Unavailability of data imposed restrictions to the length of the period of study. Sentinel 2 data are not available for the pre-2015 period, while SPEI_12 data for the post 2023 data were not available from the Global SPEI database (SPEIbase). Initially the intention was to analyze the impact of climate change on plant health according to species. The time-series analyses for environmental indices were based on data derived from a collection of Sentinel 2 images from the ee. ImageCollection(“COPERNICUS/S2_SR_HARMONIZED”).

NDVI and NDWI calculations and the extraction of SPEI_12 data were performed using a JavaScript. The next step involved exporting the output data for NDVI, NDWI (for both Cluster 0 and Cluster 1) and SPEI_12 to Google Drive in CSV format. The CSV data were then transformed to text data and used for further analysis in Excel environment. The entire data acquisition, processing and exportation process is illustrated in Figure 3.

Monthly averages for NDVI and NDWI values were calculated for each month based on the days when cloud cover was less than 10%. This process of aggregation eliminated the effects of temporal autocorrelation and bias from the analyses. The analyses culminated in two multiple regression models, catering for the two clusters mentioned above. The regression was undertaken in an Excel environment. The general form of the models was.

Y ̂ = β 0 + β 1 · NDWI + β 2 · Temp _ min + β 3 · Temp _ max + β 4 · Precipitation + β 5 · SPEI _ 12.

Where.

Y^ = NDVI, a measure of greenness or vegetation health, β 0 = the intercept, β 1 = NDWI , ameasure water content in vegetation or detect water bodies. β 2 = Tem p min, Minimum Temperature, β 3 = Tem p max , Maximum Temperature, β 4 = Precipitation and β 5 = SPE I 12 = the Standardized Precipitation-Evapotranspiration Index (SPEI) over 12 months, a measure of wetness or drought severity.

In order to improve the predictive value of the models, all months that did not have a complete data set for variables included in the analyses, namely NDVI, NDWI, Maximum Temp, Minimum Temp, Precipitation and SPEI_12, were dropped from the analysis. Thus, the study period was confined to the years between 2015 and 2023 because of the restrictions imposed by the unavailability of SPEI_12 data for the post 2023 period in the CSIC/SPEI/2_10″ Global SPEI database (SPEIbase) and unavailability of Sentinel 2 data in the pre-June 2015 period. Due to these restrictions, each of the variables consisted of only 71 monthly data values, spanning a period of 8 years.

GEE has become a tool of great interest to the academic and research world because it is a cloud-based platform that facilitates geoprocessing of datasets (Velastegui-Montoya et al., 2023). GEE provides powerful capabilities for remote sensing images due to the numerous advantages it offers, including capabilities that enable users to manage vast amounts of satellite images and perform complex analyses efficiently (Xu et al., 2025). Hence, it has been used in several fields, such as ecological and environmental monitoring (Kotikot et al., 2024), climate change studies (Yadav et al., 2024), natural resource management (Vizzari et al., 2024), hence the significant contribution it has made to global environmental protection and sustainable development research (Ke et al., 2024). One of the biggest advantages offered by GEE is its capacity to process and store vast amounts of data. GEE has a vast data repository comprising a collection of over 40 years of satellite imagery from different parts of the world (Kumar and Mutanga, 2018). This makes time-series and trend analyses possible. In addition, GEE is an open-source cloud-based platform that makes it easy and convenient to carry out geospatial analyses on a large scale (Ashok et al., 2021).

4.1 Normalized difference vegetation index

The indices were calculated using the following formulae:

NDVI = ( NIR – Red ) / ( NIR + Red )

NIR = Near − Infrared Band ( B8 )

Red = Band 4 ( B4 )

The NDVI is used to assess vegetation health. For instance, Mukwada and Manatsa (2018) used NDVI to assess vegetation health in the Drakensberg Mountains of South Africa. It is an index that reflects greenness. Its values range between −1 and +1. The higher the NDVI values the healthier the vegetation they correspond to. Values that are close to +1 denote dense, lushy, and healthy vegetation, especially forests or thriving crops. Such conditions are typical where NDVI values exceed 0.7. NDVI values that range between 0.2 and 0.6 depict moderate crop growth or productivity. Such values often reflect grasslands, sparse vegetation cover and moderate crop cover. When NDVI values are close to 0, especially when they range between 0 and 0.1, there is presence of minimal plant growth, which is normally in cases where surfaces are bare, especially in urban environments, where vegetation is sparse.

4.2 Normalized difference water index

NDWI was calculated as follows.

NDWI = ( Green – NIR ) / ( Green + NIR )

Where:

Green = Band 3 (B3).

NIR = Near-Infrared Band (B8).

Just like the NDVI, the NDWI values range from −1 to +1. It is used to assess the presence of water in the environment; hence it is referred to as the index of wetness. Its values are positive in locations characterized by water bodies like rivers, lakes or water that has accumulated on vegetation or where wetlands exist. It is an index that reflects wetness. Higher values signify the abundance of water in the environment. NDWI values that are equal to or greater than 0.5 depict high water content in the environment. Excessive amounts of water are detrimental to plant health, especially in terrestrial ecosystems and farmlands, unless the plants are hygrophytic in nature. NDWI values range between 0.3 and 0.5 under moderate wet conditions. NDWI values ranging from 0 to 0.3 are denote low water content in the environment. They signify water stress or bare soil and surfaces with minimal vegetation cover. Values that are less than 0 are typical of environments characterized by bare soil, built-up areas, or other surfaces moisture retention is poor. This is normally the case in urban environments.

4.3 Standardized evapotranspiration index

SPEI was calculated using the following formula

SPEI = ( P − PET ) − μ σ

Where:

• P − PETis the water balance for the period,
• μis the long-term mean of P − PET,
• σis the standard deviation of P − PET.

The SPEI data used were calculated using the Penman-Monteith formula, where Evapotransipiration was calculated as follows:

λET = Δ ( R n − G ) + ρ a C p ( e s − e a ) r a Δ + γ ( 1 + r s r a )

where R_n is the net radiation, G is the soil heat flux, (e_s - e_a) represents the vapor pressure deficit of the air, r _a is the mean air density at constant pressure, c_p is the specific heat of the air, D represents the slope of the saturation vapor pressure temperature relationship, g is the psychrometric constant, and r_s and r_a are the (bulk) surface and aerodynamic resistances. The interpretation of SPEI results is summarized in Table 2.

Table 2. Interpretation of SPEI results.

In general, positive SPEI values denote wet conditions. The higher the SPEI value the wetter the environment. The converse is true for negative values. The lower the SPEI value is the greater the aridity.

Before multiple regression analysis was undertaken the values for NDVI, NDWI and SPEI, as well as the values for precipitation, Maximum temperature and Minimum temperature, were further analyzed to determine their suitability for building the regression model that can be used to predict the state of plant health in the orchard. The first step involved the selection of suitable variables for the model by correlating them with NDVI, which denotes plant health, the dependent variable in the model. The Pearson Product Moment correlation coefficient was calculated to determine how NDVI correlated with potential predictors. Variables that exhibited a strong relationship with NDVI or those that had an indirect effect on NDVI, for instance NDWI, were considered as potential candidates for inclusion in the model. The results of the correlation analysis are summarized in Table 3.

Table 3. Correlation matrix used for variable selection and VIF calculation.

Bold values depict the correlation values of variables which exhibit multicollinearity.

4.4 Test for normality, multicollinearity and homoscedasticity

The inclusion of variables in the models depended on the level of normality, multicollinearity and homoscedasticity they exhibited.

4.5 Test for normality

The Jarque-Bera (JB) test was used to check for normality in the data for the dependent variable, namely NDVI. The formula used in the JB test is as follows.

JB = n 6 ( S 2 + C 2 4 )

Where JB = Jarque-Bera test statistic, S = Skewness, C = Kurtosis.

The JB test statistic indicates normality of data when its p-value is greater than 0.05. Initially, when the raw data for NDVI Cluster 0 was subjected to the JB test the results showed that NDVI values were not normally distributed. However, after a log transformation of the NDVI Cluster 0 data the Jarque–Bera (JB) statistic revealed normal distribution (JB = 2.28, p=0.32). The process was repeated for NDVI Cluster 1. Nevertheless, when the JB test was performed on NDVI Cluster 1 after a log transformation the results still showed that the transformed data was not normally distributed (JB = 8,796,647, p= 0.012). Hence, robust Standard Error (SE) had to be computed and used in the interpretation of the model for Cluster 1.

4.6 Test for homoscedasticity

The Breusch–Pagan (JP) test was performed to check for homoscedasticity in the model. The following formula was used:

JP = n . R 2

Where JP = Breusch-Pagan test statistic, n = number of observations, and R² = coefficient of the residuals from auxiliary regression.

Homoscedasticity occurs when the p-value for the Breusch–Pagan statistic is greater than 0.05, and the opposite its true for heteroscedasticity. The raw NDVI data for both Cluster 0 and Cluster 1 models exhibited heteroscedasticity.

Thus, while none of the variables were disqualified on account of multicollinearity the dependent variable fell short on normality and homoscedasticity. Hence, a Log transformation was executed before regression was performed on the dependent variable (NDVI Cluster 0, and NDVI Cluster 1). The results are reported in the next section of the paper.

5 Results and discussion

Figure 4 shows the variability of precipitation in the UFF. As shown in Figure 4, between 2015 and 2023 precipitation received in the study area exhibited high seasonality, but with a statistically insignificant trend. As illustrated in Figure 5, NDVI values recorded in the UFF fluctuated, but indicate a statistically significant increase between 2015 and 2023. This applies for both Cluster 0 and Cluster 1 parts of the UFF. This increase explains the vigorous the plant growth that took place during after the 2014–2026 drought. NDVI values were positive throughout the study period, though they tended to be lower in summer, yet that is the time when plants are expected to grow with vigor. Summer is the period when rains are mostly received in Zimbabwe.

In Zimbabwe, where most agricultural practices are rain-fed, farmers assume that there is no need to water their orchards during summer. Hence, the drop in NDVI values in April 2016, March 2018 and December 2019 is purely the result of lack of water supplementation when plants are exposed to moisture stress due to exposure to high temperatures during this season. In Cluster 0, the highest NDVI values was recorded in June 2022, when NDVI reached 0.642, while the lowest was recorded in November 2019 when NDVI dropped to 0.1325, yielding a range of 0,5,095. Cluster 1 recorded its highest NDVI value in 2020, when it reached 0,473 and its lowest NDVI value of 0,1,635 in November 2019, giving a range of 0,3,095.

The significant rise in NDVI values after October 2018 may be a culmination of wet conditions that prevailed during the preceding period. An unprecedented total of 1,048 mm was received between October 2016 and February 2017, making it the wettest period of the time-series. This significant increase is a sign of recovery that took place after the 2014–2016 drought.

The period between 2018 and 2020 was characterized by uncertainty due to the Covid 19 pandemic. The management of orchards slackened due to poor ergonomics. Since the orchard partly depends on supplementary water from irrigation during the dry season and dry spells, borehole pump breakdowns were left unattended for lengthy periods due to the Covid 19 pandemic. The pandemic imposed mobility restrictions, thus limiting access to technical services needed for pump repairs. This led to deterioration of plant health in the orchard, despite the relatively high NDWI values that were recorded in this month compared to other months in the time-series (Figure 5).

Average NDVI values recorded for Cluster 0 are higher than those for Cluster 1. While it was not possible to determine the spectral signatures for specific plant species in the orchard during unsupervised classification, it can be concluded from this analysis that the dominant species in Cluster 0, namely Eryobotrya japonica, exhibits higher levels of NDVI compared to Persea americana, Mangifera indica, Azanza garckeana and Prunus persica, which dominate the remaining part of the orchard, i.e., Cluster 1. The difference between NDVI values for the two clusters increased between 2015 and 2023, even though the values for both clusters increased with time. For Cluster 0, mean NDVI values ranged between 0.1635 in November 2019 to its maximum of 0.642 in June 2022. On the other hand, mean NDVI values for Cluster 1 ranged from a minimum of 0.1325 in November 2019 to a maximum of 0.473 in January 2020. However, the patterns of variability for the time-series of both clusters are similar, indicating that the two clusters are under similar external influences, because of their geographic proximity.

The generally low NDVI values are characteristic of sparse vegetation in the orchard. In this case, sparsity is the result of planned plant spacing, which leaves gaps between trees. These gaps are characterized by bare patches of soil from which unwanted plants are weeded, hence the relatively low NDVI values.

NDWI values were negative throughout the study period, ranging from −0.215 in November 2019 to −0.6865 in June 2022 for Cluster 0, and from −0.1955 in November 2019 to −0.533 in April 2020 for Cluster 1 (Figure 5). Just like NDVI, variability patterns for NDWI for the two clusters were similar. Trends for both mean NDVI and NDWI were statistically significant (p = 0.0001). The highest NDWI value was recorded in April 2016. The high correlation between NDWI and NDVI shown in Table 3 (p < 0.5) strongly suggests that the presence of water in the environment plays a crucial role in plant health in the orchard, regardless of the low NDWI values recorded during the period of study. The inverse relationship between NDVI and NDWI is typical in an UFF which depends on irrigation, especially during the dry season or dry spells. When plant growth is vigorous NDVI values increase while NDWI values decrease due to diminution of soil water content (Figure 6). SPEI_12 was poorly correlated with NDVI values, hence vegetation health. However, the decreasing trend of NDWI while NDVI is increasing could also imply deteriorating environmental conditions that might be accompanying plant growth in the UFF. As plants grow, their rooting systems spread and become more extensive and enable them to tap water from the phreatic zone, thus depleting from the zone and cause the water table to recede (Table 4). Combined with losses resulting from evapotranspiration and increasing reliance on groundwater in a changing climatic environment, the use of water by growing plants in the UFF leads to loss of soil moisture and subsequently to a decrease in NDWI. Such a situation might become worse in areas where the density of boreholes is high or areas where the intensity of water abstraction is high.

Table 4. VIF values derived from the correlation matrix.

Szabó et al. (2022) have reported several places where a significant decline in groundwater levels has already been observed, indicating an imbalance between the natural groundwater recharge and groundwater abstraction.

However, as illustrated in Figure 7, a close inspection of the variability of SPEI values indicates that the greater part of the time-series was dominated by near-normal and normal rainfall, with values ranging between −0.99 and +0.99. The exception was the period between February and December 2017, when moderately wet conditions were experienced, and the period between January 2015 and December 2016 when moderately dry and severely dry conditions were prevalent.

During this period, only 1 month received normal rainfall, namely November 2015. The 2014–2016 drought was widely reported across southern Africa (Archer et al., 2017). This explains why the lowest SPEI values in the orchard were recorded in 2016. The 2014–2016 drought was attributed to the El Nino effect (Rembold et al., 2015; Blamey et al., 2018; Swemmer et al., 2018).

Blamey et al. (2018) estimated the return period of this extreme early summer drought to be greater than 200 years.

Tables 5, 6 provide summary outputs that resulted from the analysis for Cluster 0 and Cluster 1 models, respectively. The models that resulted from the analysis are as follows:

log ( NDVI ) Cluster 0 = − 1.3279 + ( − 1.7205 ) · ( NDWI Cluster 0 ) + 0.00086 · ( Temp _ min ) + 0.00043 · ( Temp _ max ) + 0.00025 · ( Precipitation ) + 0.00758 · ( SPEI _ 12 ) (1)

log NDVI Cluster 1 = − 1.21 − 1.84 ( NDWI Cluster 1 ) + 0.0072 ( Temp _ min ) − 0.0080 ( Temp _ max ) + 0.000029 ( Precipitation ) + 0.0089 ( SPEI _ 12 ) (2)

Table 5. Summary for cluster 0 model output.

Table 6. Summary for cluster 1 model output.

In both models, the Breusch–Pagan test indicated no evidence of heteroscedasticity, confirming that the residuals have constant variance, and the reported standard errors (SEs) are valid. The robust SE are noted in Tables 7a,b

Table 7. Robust residuals for cluster 0 model.

5.1 Model NDVI cluster 0

Log NDVI Cluster 0 regression model was significant, F(5,65) = 142.84, p < 0.001, explaining 91.7% of the variance in the dependent variable (Adjusted R² = 0.910). NDWI Cluster 0 was a strong negative predictor (β = −1.72, SE = 0.076, t = −22.79, p < 0.001). The intercept was also significant (β = −1.33, SE = 0.114, t = −11.63, p < 0.001). Other predictors (Temp_min, Temp_max, Precipitation, SPEI_12) were not statistically significant (all p > 0.23). Diagnostic tests indicated that the residuals were normally distributed (JB = 2.28, p = 0.32) and homoscedastic (BP = 1.73, p ≈ 0.88), confirming the validity of the regression assumptions.”

5.2 Model NDVI cluster 1

Log NDVI Cluster 1 regression model is significant, F(5,65) = 91.32, p < 0.001, explaining 87.5% of the variance in the dependent variable (Adjusted R² = 0.866). NDWI Cluster 1 was a strong negative predictor (β = −1.84, SE = 0.11, t = −17.18, p < 0.001). The intercept was also significant (β = −1.21, SE = 0.11, t = −10.81, p < 0.001). Other predictors (Temp_min, Temp_max, Precipitation, SPEI_12) were not statistically significant (all p > 0.09). Diagnostic tests indicated homoscedastic residuals (BP ≈ 0, p ≈ 1.0) but some deviation from normality (JB = 8.80, p = 0.012). White’s robust SEs had to be used when drawing inferences from the model. Table 7 provides a summary of the results based on these robust SEs.

An interesting finding from this study is that while SPEI and precipitation are used to assess environmental conditions such as floods and drought they did not emerge as good predictors for vegetation health. Instead, they were superseded by NDWI, which contributed the most to plant health in UFF. This knowledge is critical for the management of own food production systems such as orchards, where moisture stress needs to be closely monitored and addressed early. Such knowledge is important for irrigation scheduling.

Within the context of this study, one could argue that though GEE is popularly used for assessing and evaluating the ecological quality of urban environments (Yang and Su, 2023), it could also be used to monitor environmental indices like NDVI and NDWI, as well as climatic variables such as maximum and minimum temperature in a manner that supports the monitoring of the productivity of food systems in UFF and promotion of urban food security.

Careful monitoring of UFF based food systems is important for two main reasons. Firstly, as the city grows, there is loss of agricultural space due to acquisition of land for developments taking place in the peri-urban environments. Secondly, the distances over which food must be hauled from the countryside into the city are increasing. Consequently, food prices are expected to continue to increase in the future, thus rendering many poor urban households’ food insecure. Climate change exacerbates this situation in Harare where the policy environment does not acknowledge the value of urban farming. In the case of Zimbabwe, climate change will continue to impact food systems in urban areas where many households rely on their own food production sources. Zimbabwe National Statistics Agency (2017) notes that in Zimbabwean urban areas 38% of all forms of food produced by employed urban dwellers are for own household consumption.

As part of UFF monitoring, regression models could be updated when new images are added to the GEE collection on a daily basis, thus making real-time monitoring of environmental variables such NDVI, NDWI, precipitation and temperature possible. This could make the management of UFF production systems easier and more efficient, thus enhancing the resilience and adaptive capacity of urban households against climate change. As climate change and land use and land cover changes continue to alter the dynamics of the urban landscape, food insecurity is expected to worsen. Micro spaces in the urban landscape should be used to create opportunities for urban households to manage their own food systems, including those based on irrigated backyard gardens and orchards. Since orchards and other forms of UFF food outlets depend on scarce water resources such as groundwater, whose scarcity is likely to worsen due to climate change and whose pumping costs are likely to continue to rise as water tables recede, optimal water use should be ensured. Improved scheduling of irrigation will prevent excessive watering or deficits in supply of water. In this regard, monitoring of NDVI and / or NDWI could enhance water use efficiency (Singh et al., 2023). It should be argued that urban farming must be considered as an integral component of modern cities, while the use of smart technologies like remote sensing and GIS to monitor UFFs in the city should be encouraged. These technologies make it easy to capture, manage, store, retrieve and analyze huge quantities of data efficiently and timeously. This will improve the quality of decisions made about food availability, access, stability, utilization, agency and sustainability, all of which contribute to food security (HLPE, 2020).

6 Conclusion

The climate of Harare has been changing during the last decade. While existing early warning systems rely on NDVI for monitoring vegetation and crop health, this study reveals that conditions characterizing NDWI constitute the underlying factors that drive changes in NDVI. From the foregoing discussion it can be concluded that the influence of SPEI_12, Precipitation, Maximum temperature and Minimum temperature on NDVI is minimal, since they only explain a small fraction of the variance exhibited by the regression models. This knowledge is critical for the management of urban forests, where water is directly conveyed to plants by artificial means such as irrigation, especially in an environment where drought and dry spells are persistent due to climate change. Since NDWI is negatively correlated with vegetation health (NDVI), and NDWI values are persistently low in the UFF, we can conclude that the prevailing aridity is the outcome of worsening soil moisture stress resulting from moisture abstraction by plants, as well as climate change. This is supported by the pattern of SPEI_12 variability. It can also be concluded that climate change induces water shortage in the urban environment and affect urban food forestry in a manner that undermines vegetation health and food security. Hence, the use of robust smart technologies to track changes in environmental indices such as NDVI and NDWI could enhance the adaptive capacity and resilience of urban households against climate change and transform the current fragile food systems into robust urban food systems.

7 Limitations of study

The study faced a few limitations. First, the study relied on Sentinel 2 images, whose spatial resolution of 10 m was too coarse for species detection. Hence, it was not possible to model NDVI for individual species. Second, the unavailability of data for the pre-2015 period and SPEI_12 data from the Global SPEI database (SPEIbase) for the post-2023 period imposed constraints on the period of study. Hence, the study was confined to the 2015–2023 period when all datasets were available. Third, there was no UFF based crop productivity data for validating the results of the models generated from this study. Fourth, the scope of this study was limited to environmental indices and climate variables. The influence of other factors such as edaphic, slope, aspect, underlying geological conditions, as well as depth of the water table and socio-economic factors on plant health was not investigated. Lastly, though technically robust and efficient, Remote Sensing and GIS based analyses employed in this study cannot easily be applied for monitoring UFFs due to technical challenges associated with their application. They are too sophisticated for adoption by most urban farmers, especially in the developing world.

8 Recommendations of the study

As demonstrated in this study, the modeling of plant health in UFFs is a tedious and complicated endeavor involving the application of Remote Sensing, GIS and regression analyses, which are too sophisticated for adoption by most urban farmers. Consequently, it is recommended that future studies should focus on development of simpler applications, including mobile based applications, that will make it easy for urban farmers to monitor NDVI and NDWI, as well as productivity in their UFFs. This will make their operations smarter and more efficient. The study also recommends that future studies should investigate the influence of geomorphological, hydrogeological and socio-economic factors on plant health and productivity in UFFs, since these were beyond the scope of the current study.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Bethuel Sibongiseni Ngcamu, University of South Africa, South Africa

Reviewed by: Nesrine Kadri, National Research Institute for Rural Engineering, Water and Forestry (INRGREF), Tunisia

John Beteri, University of Dar es Salaam Institute of Development Studies, Tanzania