Forecasting vegetation condition for drought early warning systems in pastoral communities in Kenya

Highlights

• We present a forecast of commonly-used drought indicators in Kenya.

• For pastoralist regions, two separate methods were developed to forecast NDVI and VCI.

• Landsat Gaussian Processes models provide a 4-week forecast with a hit rate of 87%.

• Similar results were found for linear autoregression models using MODIS.

• These models can help disaster risk managers act early to reduce impact of drought.

Abstract

Droughts are a recurring hazard in sub-Saharan Africa, that can wreak huge socioeconomic costs. Acting early based on alerts provided by early warning systems (EWS) can potentially provide substantial mitigation, reducing the financial and human cost. However, existing EWS tend only to monitor current, rather than forecast future, environmental and socioeconomic indicators of drought, and hence are not always sufficiently timely to be effective in practice. Here we present a novel method for forecasting satellite-based indicators of vegetation condition. Specifically, we focused on the 3-month Vegetation Condition Index (VCI3M) over pastoral livelihood zones in Kenya, which is the indicator used by the Kenyan National Drought Management Authority (NDMA). Using data from MODIS and Landsat, we apply linear autoregression and Gaussian process modelling methods and demonstrate high forecasting skill several weeks ahead. As a bench mark we predicted the drought alert marker used by NDMA (VCI3M<35). Both of our models were able to predict this alert marker four weeks ahead with a hit rate of around 89% and a false alarm rate of around 4%, or 81% and 6% respectively six weeks ahead. The methods developed here can thus identify a deteriorating vegetation condition well and sufficiently in advance to help disaster risk managers act early to support vulnerable communities and limit the impact of a drought hazard.

Introduction

Droughts are a major threat globally as they can cause substantial damage to society, especially in regions that depend on rain-fed agriculture. They particularly impact food security by significantly reducing agricultural production (Lesk et al., 2016) and raising food prices (Nelson et al., 2014; Brown and Kshirsagar, 2015), which often leads to increased levels of malnutrition, migration, disease, and other health concerns (Piguet et al., 2011; Stanke et al., 2013). The majority of droughts occur in sub-Saharan Africa (EM-DAT, 2019) where many communities rely on predictable rainfall patterns for their livelihood.

In East Africa, the main economic activity in the arid and semi-arid lands (ASAL) is subsistence rain-fed agriculture, as well as livestock farming using pastures and grasslands as the main source of fodder. As a result, the pastoral and agro-pastoral communities who live in these drylands are particularly vulnerable to drought (Nyong et al., 2007; Orindi et al., 2007), especially since their existing coping strategies have been compromised by population growth and land use change in recent years (Galvin et al., 2001). Governments and donor agencies in the region have thus developed several tools and early warning systems (EWS) to mitigate the impact of droughts on pastoralists.

Most EWS tend to monitor current key biophysical and socio-economic factors to assess the possible exposure of vulnerable people to specific hazards. However, once the impacts are visible, it may be too late to mitigate the consequences (Kogan et al., 2013). Hence there is growing interest in moving toward a proactive humanitarian approach to disasters by developing preparedness actions based on climate forecasts (Coughlan de Perez et al., 2015; Lopez et al., 2018; Wilkinson et al., 2018). Additionally, it is estimated that being better prepared before a drought hits significantly reduces the costs and losses from these disasters (Venton et al., 2012). Hence, EWS now increasingly include expert knowledge and qualitative assessments of seasonal climate forecasts to assess the future development of food security, and define actions to mitigate possible losses (Coughlan de Perez et al., 2015; Tozier de la Poterie and Baudoin, 2015). However for drought conditions, a meteorological drought does not always lead to negative agricultural outputs (Bhuiyan et al., 2006). There is thus a growing interest to include forecasts of the impacts of these hazards (WMO, 2015; Sai et al., 2018; Sutanto et al., 2019).

In Kenya, following several periods of intense drought, the government established the National Drought Management Authority (NDMA) in 2016, to set up and operate a drought EWS, as well as to establish drought preparedness strategies and contingency plans. The NDMA provides monthly bulletins assessing food security in the 23 ASAL regions using current biophysical (e.g., rainfall, vegetation condition) and socio-economic (production, access, and utilisation) factors. One key biophysical indicator used by the NDMA drought phase classification is based on the Vegetation Condition Index (VCI) (Kogan, 1995; Klisch and Atzberger, 2016; Rulinda et al., 2011; Rojas et al., 2011).

The VCI, which expresses the Normalized Difference Vegetation Index (NDVI) in terms of where it currently lies within its expected range for the given pixel, is one of a number of satellite-based indicators that have been developed to detect and monitor drought (Zargar et al., 2011). While there is little agreement between VCI and precipitation-based meteorological drought indicators (Bhuiyan et al., 2006; Quiring and Ganesh, 2010), it is strongly linked to agricultural production and widely used to identify drought onset, intensity, duration, and impact (Jiao et al., 2016). The NDMA uses the 3-month averaged VCI (VCI3M) in its operational EWS (Klisch and Atzberger, 2016). Once the VCI3M goes below a threshold of 35, the NDMA triggers a rapid food security assessment and has access to the National Drought Contingency Fund in order to implement its preparedness strategies and contingency plans.

The main goal of this paper is to explore machine-learning techniques to forecast the vegetation indices that are commonly used in the pastoral areas of Kenya to monitor droughts. In order to provide useful information to drought risk managers, we aim to identify the right balance between forecast lead time and uncertainty. To this end, we evaluated the performance of our approaches up to ten weeks ahead.

Based on NDMA's experience, we particularly focused on the pastoral livelihood zones as the VCI3M is more reliable in identifying drought condition for grazing and browsing in the more arid regions of the country. Several studies have developed statistical and machine-learning approaches (Udelhoven et al., 2009; Meroni et al., 2014; Zambrano et al., 2018; Vrieling et al., 2016) to predict end-of-season crop, forage and biomass production. Recently, Matere et al. (2019) developed a decision support tool based on a mechanistic model to estimate 6-monthly forecasts of forage condition. Here, we specifically focus on Gaussian Process (GP) modelling (Rasmussen and Williams, 2006), and linear autoregressive (AR) modelling (Hamilton, 1994) to forecast NDVI and VCI3M, which are derived from both Landsat (every 16 days at 30 m resolution) and the MODerate resolution Imaging Spectroradiometer (MODIS - daily data at 500 m resolution). GP modelling uses kernel-based non-parametric Bayesian inference on the structure of correlations between observations, and is widely applied to classification, interpolation, change detection and forecasting problems (Brahim-Belhouari and Vesin, 2001; Chandola and Vatsavai, 2011; Camps-Valls et al., 2016; Upreti et al., 2019). Linear AR is the regression of future observations on past observations, assuming a linear dependence. Previously it has been performed on monthly (i.e. temporally more sparse) NDVI data, see for example (Asoka and Mishra, 2015; Papagiannopoulou et al., 2017), with mixed results in terms of forecasting potential (R²-scores between 0 and 0.4 at a lead time of one month).