Regional-Scale Analysis of Soil Moisture Content in Malawi Determined by Remote Sensing

Abstract

Soil moisture content is typically measured in situ using various instruments; however, due to the heterogeneous nature of soil, these measurements are only suitable at a very local scale. To overcome this limitation, earth observation satellite remote sensing data, particularly through the inversion of the closure phases of interferometric synthetic aperture radar (InSAR) observations, enables the determination of soil moisture content at regional to global scales. Here, we present, for the first time, a regional-scale study of soil moisture determined from remote sensing observations of Malawi, specifically, two areas of interest capturing arable and national parklands in Kasungu and Liwonde. We invert the closure phases of InSAR acquisitions from Sentinel-1 between 1 January 2023 and 31 May 2024 to measure the soil moisture content in the same time range. We show that soil moisture content is heavily influenced by local precipitation and highlight common trends in soil moisture in both regions. We suggest the difference in soil moisture observed inside and outside the national parks is a result of different overlying vegetation and conservation agriculture practices during the maize crop cycle in Malawi. Our results show the effectiveness and suitability of remote sensing techniques to monitor soil moisture at a regional scale. The upcoming additions to ESA’s fleet of earth observation satellites, in particular Sentinel-1C, will allow for higher-time-resolution soil moisture measurements.

1. Introduction

Malawi is a Least Developed Country [LDC] [1] located in southeastern Africa, where 85% of the population are employed in the agriculture sector [2], predominantly on small-holder farms [3]. The staple food of Malawi is maize (Zea mays L.), the vast majority (90%) of which is rainfed during the yearly rainy season from November to April [4]. The dependence on rainfed maize for employing and feeding the population makes Malawi particularly susceptible to food insecurity during droughts and floods. On 25 March 2024, the President of Malawi declared a national disaster as a result of a drought induced by El Niño (https://reliefweb.int/report/malawi/urgent-action-critical-malawi-faces-severe-drought, accessed on 27 October 2024). The frequency and severity of droughts are expected to increase due to higher mean global temperatures brought about by climate change [5]. It is thus essential to develop and improve ways of monitoring the indicators of oncoming drought. Remote sensing data offer timely, impartial measures of the impacts of climate change on biomass and agricultural systems [6], while monitoring soil moisture plays a key role in predicting the onset of droughts at a regional level and so contributes towards long-term food security.

Though many in situ methods for measuring soil moisture exist, they are only effective for very localised analysis owing to the heterogeneous nature of soil and the cost associated with installing in situ monitors being prohibitive for larger scale areas. However, regional-scale measurements of soil moisture are only possible using data from remote sensing earth observation satellites. Singh et al. [7] outlined the current popular methods of determining soil moisture between in situ, remote sensing, and machine learning algorithms. While passive sensing observations in optical and infrared wavelengths of the Normalised Difference Vegetation Index (NDVI) can be used to determine changes in soil moisture, they found that the most popular method in the literature of soil moisture estimation with remote sensing utilised radar backscatter from active sensing SAR observations. C-band SAR observations are particularly suited to estimating surface soil moisture content (defined as <5 cm below the surface). There also exist a small number of methods to measure soil moisture from Interferometric SAR (InSAR) observations, notably De Zan et al.’s [8] and De Zan and Gomba’s [9]. InSAR methods for determining soil moisture are still an active area of research [10], and existing implementations such as that developed by Karamvasis and Karathanassi [11] have not been fully explored in different agrifood contexts.

This paper explores two areas of interest (AOIs) near national parks in Malawi, Kasungu in the northwest and Liwonde in the southeast. The Kasungu AOI has an area of ∼1168 km², the western ∼334 km² of which is covered by Kasungu National Park. The remaining area is predominantly arable land with the exception of the town of Kasungu in the south east of the AOI. The Liwonde AOI is ∼911 km² in area and almost fully encompasses Liwonde National Park, which has an area of ∼438 km². The Shire river runs along the western edge of Liwonde National Park while the area outside the national park is arable land.

We use remote sensing data to determine the soil moisture content inside and outside the national parks in both regions. The remainder of this paper is structured as follows. Section 2 outlines the data used, Section 3 describes our methodology, and Section 4 highlights the results. We discuss the effectiveness and impact of our work in Section 5 and conclude with a summary in Section 6.

2. Data

The data used for this analysis were a combination of remote sensing satellite observations and meteorological data, which we briefly describe below.

2.1. Sentinel-1A

Sentinel-1A is a SAR satellite operated by the European Space Agency (ESA) as part of its Copernicus programme since 2015. The satellite operates in the C band with a ground resolution of 5 m × 20 m when operating in Interferometric Wide Swath (IW) mode. We obtained IW Single Look Complex (SLC) images from the two Sentinel-1A acquisition footprints marked in cyan in Figure 1 in the time period from 2 January 2023 to 27 May 2024, totalling 81 acquisitions. The data were accessed via the Copernicus Data Ecosystem (https://dataspace.copernicus.eu, accessed on 18 June 2024). SLC data contain both amplitude and phase information, which are crucial for performing the soil moisture analysis described in Section 3.

2.2. Moderate Resolution Imaging Spectroradiometer (MODIS)

The Moderate Resolution Imaging Spectroradiometer (MODIS) is an instrument onboard the Terra and Aqua Earth Observing System satellites operated by NASA. Data was accessed via the Google Earth Engine [12]. This analysis used the NDVI from the MODIS 16-day 500 m dataset in the time period from 1 January 2023 to 27 May 2024. Figure 1 shows a MODIS true colour image of Malawi (excluding Lake Malawi) on 11 July 2024.

2.3. Meteorological and Soil Composition Data

The soil moisture analysis made use of meteorological data obtained from the ERA5 hourly reanalysis dataset [13]. Precipitation and land surface temperature were acquired from the Copernicus Climate Change Service (https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-land, accessed on 18 June 2024).

Assumptions on the dielectric properties of soils for various compositions [14] are an integral part of the soil moisture inversion algorithm described by Karamvasis and Karathanassi [11]. We obtained the clay and sand compositions of the soil in our AOIs from the soilgrids dataset [15] (https://soilgrids.org/, accessed on 20 June 2024) at a 1000 m resolution.

3. Methods

Malawi has a generally tropical climate with an average annual rainfall that varies between ∼730 mm and ∼1440 mm. The vast majority of its yearly rainfall ($\sim 94\%$) occurs during the rainy season, which typically lasts from November to April. The predominant World Reference Base reference soil groups [16] are lixisols and luvisols, covering ∼47% of the land area of Malawi [17]. This, combined with the fact that only a small proportion of arable land is irrigated, makes Malawi an ideal place to test the De Zan and Gomba [9] soil moisture model as there should be a distinct seasonal change in soil moisture between the dry season and the rainy season. We aimed to verify the Karamvasis and Karathanassi [11] implementation of this model in two regions over an 18-month period. The time range from 1 January 2023 to 31 May 2024 was chosen in order to perform a year-on-year comparison between drought and non-drought conditions. This work sought to demonstrate the efficacy of this approach and thus provide the basis for following studies to expand and develop further applications for remote sensing measures of soil moisture with ultimately their inclusion in drought early-warning systems and in the optimisation of trends in food security strategies.

The analysis in this paper utilised recent developments in Interferometric SAR data processing to determine the soil moisture in the two AOIs described in Section 1. We chose these two locations because the south of Malawi was more adversely affected by the drought in March 2024 than the north, which allowed a comparison between two similar areas for different soil moisture conditions. These are shown in orange in Figure 1 as an area around Kasungu National Park towards the northern part of the country and around Liwonde National Park in the southern part. Each AOI encompasses an area of national park which acts as a distinct background compared to the surrounding cropland and grasslands.

Soil Moisture Monitoring

The surface soil moisture (SSM) analysis software INSAR4SM (commit 9caa75d) [11] makes use of the Sentinel-1A and meteorological data described above in order to determine the moisture content, in percent, in the top layer of soil (0–5 cm). It implements the De Zan and Gomba [9] model to invert interferometric closure phases and calculate the soil moisture content. In brief, De Zan and Gomba [9] suggested that the closure phase of SAR acquisitions l, m, and n,

$${\Phi }_{l,m,n}=arg\left({\displaystyle \frac{1/2}{j(k_l-k_m^{*})}}{\displaystyle \frac{1/2}{j(k_m-k_n^{*})}}{\displaystyle \frac{1/2}{j(k_n-k_l^{*})}}\right),$$

(1)

was proportional to the moisture level of the soil ($\theta$),

$${\Phi }_{l,m,n}\propto ({\theta }_l-{\theta }_m)({\theta }_m-{\theta }_n)({\theta }_n-{\theta }_l).$$

(2)

Here, k is the wavenumber for a given SAR acquisition, $k^{*}$ its complex conjugate, and j is the imaginary number. We outline the algorithm below; however, for a more comprehensive description we refer the reader to Karamvasis and Karathanassi [11], particularly Figure 1 therein.

• Identify driest SAR acquisition. In order to solve the ambiguity of the soil moisture order described by De Zan and Gomba [9], the meteorological data from Section 2 were used to determine SAR acquisition dates with little to no precipitation. From these, the SAR acquisition with the highest mean coherence was chosen as the driest.
• Calculate coherence and phase closures. An SSM estimation grid was constructed and the interferometric coherence and phase closures calculated for each grid cell.
• Identify “dry” SAR acquisitions. Using the calculated coherence, the remaining SAR acquisitions were ordered in “dryness” starting from the driest acquisition identified earlier. The first 30% of acquisitions were labelled as “dry”.
• Calculate coherence due to SSM. The coherence from the “dry” acquisitions were used to determine the coherence loss due to soil moisture for all SSM grid cells.
• Invert to find SSM. An inversion from phase closures to soil moisture level was performed for each grid cell using the De Zan and Gomba [9] model.

We optimised the software written by Karamvasis and Karathanassi [11] to drastically reduce the time taken in calculating the SSM inversion. This update, which vectorised some frequently used for loops, is available as a fork of their original github repository (https://github.com/murphp30/INSAR4SM, accessed on 6 August 2024). A rudimentary benchmark test of the time to run one inversion loop before and after the vectorisation is presented in

Table 1. Benchmark results for inversion loop inside INSAR4SM before and after vectorisation improvement.

	Before Vectorisation	After Vectorisation
Time	265.948 s	13.517 s

. The benchmark test was run on a local server with an AMD EPYC 7543P 32-core processor, with a maximum processor speed of 3.7 GHz.

Despite this optimisation, the memory intensive nature of this software still restricts the geographical area over which the soil moisture can be calculated. We used a 500 m grid size for the SSM estimation grid and limited our analysis to the AOIs outlined in orange in Figure 1. For each AOI, we coregistered a stack of SLCs [18] with the Interferometric synthetic aperture radar Scientific Computing Environment (https://github.com/isce-framework/isce2/, accessed on 14 July 2024) (ISCE) using the observation on 9 January 2023 as the reference image for the Liwonde AOI and 2 January 2023 for the Kasungu AOI. This coregistered stack, along with precipitation and skin temperature data from ERA5, was then used to calculate the soil moisture content for each observation. To determine the long-term change in soil moisture over time, we calculated the year-on-year difference between two soil moisture measurements one year apart to the nearest acquisition date. Short-term change was captured in the running difference between an observation and the one immediately prior to it.

Karamvasis and Karathanassi [11] found that their inversion algorithm resulted in a root-mean-square error of 0.031 m³/m³ when comparing to an in situ monitoring station. No such soil moisture monitoring station existed in either of our AOIs; thus, a quantitative measure of error was impossible to determine from our analysis. Furthermore, as we see in the following, uncertainties arising from the effect of crops on the phase delay of SAR acquisitions rendered obtaining a calibrated measure of soil moisture, bounded by uncertainties, an exceptionally difficult task. We proceeded with the goal of determining soil moisture changes relative to the initial SAR acquisition.

4. Results

As mentioned previously, both regions of interest in this study contained national parks. The parks acted as a source of “background signal” in terms of soil moisture compared to the surrounding arable land. For each AOI, we performed a soil moisture analysis and compared both inside and outside the national parks. Due to differences in geography and flora, an absolute comparison between both regions was not possible. We give the results for each AOI separately below and explore the differences and commonalities between both regions in our discussion in Section 5.

4.1. Kasungu

We made the reasonable assumption that the biggest contributor to soil moisture over a given area was the cumulative precipitation up to that point in time. Given that the ERA5 precipitation data were used in part to determine the driest SAR acquisitions, we expected the resulting soil moisture inversion estimate to be consistent with precipitation measurements. This is shown to be the case in Figure 2. The right-hand axis shows the soil moisture inversion (grey line and dots), while the left-hand axis shows the total cumulative precipitation per hour in mm. We averaged the daily and 12-day cumulative precipitation values to an hourly average. We chose a 12-day cumulative average to match the revisit time of the Sentinel-1A satellite. It is clear that soil moisture increased with the onset of the rainy season in November 2023. Precipitation in 2024 was higher than 2023, and this was also evident in the soil moisture content in each year.

The soil moisture over the entire region is shown in Figure 3 with the boundary of Kasungu National Park delineated in red (interested readers are directed to the movie in the Supplementary Materials, which shows a time lapse of the soil moisture estimate). There was an apparent difference between the area inside Kasungu National Park and the surrounding agricultural area, which we revisit in Section 5. An area around the town of Kasungu remained consistently dry but otherwise the seasonal change from wet to dry season and back again was reflected in the soil moisture over the AOI.

A year-on-year comparison within the AOI is shown in Figure 4. More negative values indicate a drier environment, while more positive values indicate a wetter one. Here, we see that 2024 was a wetter year than at the same time in 2023, with the exception being late December 2023 and January 2024. Once again, the boundary of Kasungu National Park is delineated in red, and a clear difference in soil moisture inside and outside the park is apparent.

Finally, we show the running difference in soil moisture in Figure 5. The negative and positive values in the colourbar have the same meanings as for the year-on-year difference. Each panel in Figure 5 represents the change in soil moisture over the course of 12 days. The running difference can therefore be used to determine short-term changes in the soil moisture. For example, there was a dramatic increase in soil moisture between 22 November 2023 and 4 December 2023. In contrast, little change was observed in soil moisture during the dry season.

Comparisons Inside and Outside the National Park

Previously, we mentioned a difference between the trends in soil moisture inside and outside Kasungu National Park. These are evident in Figure 3, Figure 4 and Figure 5. By spatially averaging of the soil moisture inside and outside Kasungu National Park, we can gain an insight into the different soil moisture histories of the two areas. This is shown in Figure 6. We see in the top panel that the soil moisture level inside (blue) the park remained higher than outside (orange) for the majority of 2023 until the onset of the rainy season in November. The bottom panel of Figure 6 shows the rate of change in soil moisture with rates of increasing moisture denoted with upward facing triangles and rates of decreasing moisture with downward facing triangles. There is an apparent lag in moisture rates near the onset of the rainy season by $\sim 36$ days and a much greater rate of moisture loss inside the park from April 2024 onwards.

We also investigated the relationship between soil moisture, NDVI, and interferometric coherence. Villarroya-Carpio et al. [19] showed that interferometric coherence was a good proxy for vegetation cover and was strongly anti-correlated with the NDVI. Figure 7 shows the soil moisture content (grey), NDVI (forest green), and interferometric coherence (blue). We posit that the difference between soil moisture inside and outside the national park can be explained by the difference in vegetation coverage, evidenced in both NDVI and coherence. From April to May 2024, we can see that the coherence was lower while the NDVI was higher inside the park compared to outside. This would suggest a higher vegetation coverage inside the national park. This is likely explained by the surrounding maize crop being harvested.

4.2. Liwonde

Similar to above, we show the average soil moisture level and cumulative precipitation over the Liwonde AOI in Figure 8. The axes and colours are the same as in Figure 2. Once again, the soil moisture level followed the 12-day average cumulative precipitation showing a decrease over the dry season and a rapid increase with each rain event during the rainy season.

This seasonal variation in soil moisture can clearly be seen in Figure 9. The decrease in soil moisture during the dry season from May to October is evident over the entire region. An area along the Shire river (along the western edge of Liwonde National Park) retained significantly more moisture throughout the dry season than the rest of the AOI. From April to May 2024, it is evident that the soil moisture content was significantly lower inside Liwonde National Park (outlined in red) than outside.

We compared the year-on-year difference in soil moisture in Figure 10. Here, we see that overall, the region was drier in 2024 than it was in 2023, notably within Liwonde National Park (outlined in red).

Figure 11 shows the running difference in soil moisture and again tracks the short-term changes. There was little change in soil moisture during the dry season, and the start of the rainy season was evident between 11 December 2023 and 23 December 2023.

Comparisons Inside and Outside the National Park

The average soil moisture inside (blue) and outside (orange) Liwonde National Park over time is shown in the top panel of Figure 12. Unlike the Kasungu region, there was no lag in soil moisture during the onset of the rainy season. The bottom panel of Figure 12 shows the moisture rate both inside and outside Liwonde National Park. Both spatial areas showed similar rates except from approximately February 2024. Once again, we see the area inside the national park had a more rapid decrease in soil moisture than that outside.

Similar to the Kasungu AOI, the difference in NDVI, coherence, and soil moisture inside and outside the park is shown in Figure 13.

5. Discussion

Despite the high accuracy of in situ methods for determining soil moisture, the spatial heterogeneity of the soils and their hydrological properties means that they cannot be used to extrapolate beyond a local scale. The ability to determine soil moisture at a regional scale is thus unique to remote sensing. Without in situ ground-truth data, the absolute value of the soil moisture content presented in our results were not calibrated, which is the reason why we have refrained from articulating specific values for soil moisture content. Our analysis, however, demonstrates that our results are self-consistent and in keeping with the ordering of “dry acquisitions” as determined by Karamvasis and Karathanassi [11]. This means that, while a specific value for soil moisture might not be correct, the overall trend and geographic distribution of soil moisture relative to this value are realistic. We discuss the rest of our results under this lens of self-consistency.

It is apparent from Figure 2 (and Figure 8) that the average soil moisture content over a region is strongly related to the cumulative precipitation. Whilst not unexpected, it is interesting to note how quickly the soil moisture content increases from the first precipitation events of the rainy season. The geographic extent of this is evident in the running-difference images (Figure 5 and Figure 11), where strong rainfall events between observations lead to an increase in soil moisture.

One particular note is the ability of SAR-derived soil moisture mapping to highlight anomalous drying or wetting over long time periods, as is evident in the yearly differences in soil moisture content (Figure 4 and Figure 10). The regional-scale analysis afforded by remote sensing can also pinpoint “hotspots” of soil moisture/drying such as the area inside Liwonde National Park in Figure 10.

De Zan and Gomba [9] noted that they observed different behaviours in their soil moisture inversion over different ground types, in particular forested and agricultural lands. Related to this, there is growing consensus that interferometric coherence is a good marker for crop growth seasons, e.g., [19,20,21]. Our results are in agreement with this, as they show a clear negative relationship between the NDVI and coherence (see Figure 7 and Figure 13).

We interpret the different drying rates between the national parks and the surrounding land as evidence of conservation agriculture (CA) methods being practised. It is not uncommon to see small-holder farms in Malawi covered with dry crop residue after the harvest [22]. This shields the underlying soil from the sun and thus retains soil moisture for a longer time. This could likely explain the lower moisture drying rate outside the national parks than inside from January to May 2024 in Figure 6 and Figure 12. We note that the same time period in 2023 shows drastically different drying rate in Figure 6. However, Figure 12 shows that the drying rate for January–May in both 2023 and 2024 was largely similar. This could suggest that CA is not strictly practised from one year to the next and/or from one region to another.

6. Conclusions

The use of synthetic aperture radar from earth observation remote sensing has enormous potential in being able to characterise cost effectively and at scale the soil moisture properties of those regions and communities most vulnerable to the adverse effects of climate change. In this study, we demonstrated its efficacy in capturing long- and short-term trends into two separate regions in Malawi that encompassed both arable and national parklands. In both cases, we found the soil moisture to closely match precipitation patterns and clearly denoted the start of the rainy season. From orbit, we could infer that the drying rate of soil inside the national parks was greater than that outside as might be expected. The NDVI and InSAR coherence signatures measured in both AOIs were closely related to the maize crop cycle and pointed to human intervention in order to reduce soil drying rates outside of the national parks, demonstrating our ability to identify Conservation Agriculture practices remotely.

The revisit rate of Sentinel-1A at the time of writing was 12 days, which hampered a rapid response to soil drying events. The launch of Sentinel-1C and Sentinel-1D in late 2024/2025 will effectively increase the cadence of soil moisture measurements to every four days. Such increased cadence will permit greater accuracy in monitoring soil moisture anomalies at regional scales using the methodologies such as those we have described. Given the commitment of the Copernicus Programme to guarantee unrestricted access to such SAR data products, and the growing repositories of open-source software to process and derive subsurface soil moisture estimates from such data, we anticipate a more widespread use of such EO-derived data products in crop management, in drought prediction, and in risk assessment associated with agrifinance/insurance, among other areas. This higher-cadence monitoring of short-term soil moisture changes at a regional scale can help in agricultural use cases and drought forecasting.