Leveraging Geospatial Technology for Smallholder Farmer Credit Scoring ()
1. Introduction
Globally, smallholder farmers are generally understood to be those that farm on small pieces of land, often taken as 2 ha or less, and largely for subsistence; however, this size threshold varies from country to country, depending on the prevailing ecological and demographic conditions. The Food and Agriculture Organization (FAO) estimates that there are about 500 million such farmers in the world, and they produce about 80% of the food consumed in their countries [1] ; their farming activities are therefore critical to the food security of their communities and the world in general, as the UN can hardly achieve Sustainable Goal no.2 (no hunger) without them. Smallholder farmers face the challenge of limited access to credit due to the fact that most of them farm on unregistered land (a common phenomenon in developing countries), which cannot be offered as collateral to lending institutions; those that farm on registered lands fear losing such land (often their only key family asset) that they should default on loan payments and hence don’t apply for credit; those that still do apply are still disadvantaged by low credit scores due to poor financial history and low valued collateral. The result is financial exclusion, inability to provide optimal farm inputs, depressed yields and food insecurities [2] . An alternative approach to the evaluation of such farmers for credit, such as giving weight to the health of their crops (as an indicator of expected yields, and hence ability to service credit) would go a long way in mitigating the credit challenge for smallholder farmers.
Credit scoring is a statistical analysis which is performed by either lenders or credit reference institutions to assess a person’s creditworthiness. For lenders, credit scoring is subsequently used to help decide whether to extend or deny credit to a borrower [3] . For credit reference institutions, credit scoring is described as a means of making a summary of information on a credit application so as to produce a credit score. In Kenya, specifically Migori County, most farmers are unable to access loans to finance their farming activities due to low credit scores. This is mainly attributed to the fact that they are financially excluded [2] .
The need for credit scoring arises out of the fact that lending is always associated with the risk that the borrower may default on paying back their loan; for example, it has been recently reported that over 60% of small scale borrowers in Kenya defaulted on their loans in a given month [4] . The lender, therefore, desires to separate loan applicants into high risk and low risk ones, so that the lower risk ones can receive priority in getting loans, thus minimizing bad debts for lenders; a credit score is able to achieve this binary separation.
The emergence of credit scoring can be traced back to 1941 when David Durand established a credit scoring system. Durand identified different variables that helped lenders distinguish between good and bad loans [5] . In 1946, E. F. Wonderlic developed a credit score guide that helped define and narrow the variables of good and bad loans [6] . The credit score guide helped to indicate the degree of risk associated with a customer. In the 1950’s credit scorecards were becoming a popular instrument used in creditworthiness assessment.
The Fair Isaac Corporation was founded in 1956 and introduced its first credit scoring system in 1958 (myFICO). These scorecards were models that helped to determine if a customer would default on their loan given their current financial position. The late 1960s to early 1970s brought about technology that allowed for credit scoring models to be developed further and automated [7] .
Today many lenders rely on the traditional 5Cs of credit to assess the credit risk of applicants. These are Character (e.g. reputation, credit history), Capacity (e.g. cash flow stability), Capital (e.g. net worth), Collateral (e.g. fixed asset that can be sold in case of default) and Condition (e.g. current economic condition) [8] . Smallholder farmers with no or little credit history, limited capital and collateral, inevitably fare poorly in this kind of traditional assessment.
In Kenya, credit scoring for farmers is mainly carried out by commercial banks, insurance companies, the Agricultural Finance Corporation and some civil society organizations such as the Syngenta Foundation [9] ; these organizations rely on the same 5Cs approach, to the detriment of smallholder farmers.
This study aimed to determine whether, by including a geospatial variable, specifically crop NDVI, in the credit evaluation model, a smallholder farmer’s chances of accessing credit could be improved. The contribution or novelty of this research could be considered threefold: It has demonstrated how geospatial technologies can be utilized to generate non-financial data as well as how such geospatial data can complement historical information used by financial institutions to determine more farmer-friendly credit scores. The study has also contributed knowledge on the application of artificial intelligence machine learning techniques to agriculture, especially African agriculture. Finally, the research has contributed to the global debate on the financial inclusion of vulnerable smallholder farmers whose farm outputs remain critical to the achievement of food security in the world.
2. The Study Area
This study was carried out in Migori County of Kenya, illustrated in Figure 1.
Migori is a county in the former Nyanza Province located in south-western Kenya, bordering Homabay, Narok, Tanzania, Kisii and Lake Victoria. Geographically, it lies between east longitudes 33˚55'42" and 34˚43'50" and latitudes −1˚39'06" and −1˚23'21". It has eight sub-counties, namely; Awendo, Uriri, Rongo, Kuria East, Kuria West, Suna East, Suna West and Nyatike. The main livelihood activities in Migori County include agro-farming and pastoral farming. Among these livelihood activities, crop farming forms the backbone of the economy in
the county. At least 70% of the people resident in Migori County are dependent on crop farming, and the farming is dependent on rain-fed agriculture.
Demographically, Migori County is the most diverse region of the former Nyanza province after Kisumu County. The main inhabitants are Suba people, Luo, Abakuria, Abagusii, and Abaluhya. Others are Somalis, Indians, Arabs, and Nubians. The population of Migori County according to the 2019 population census was 1.2 million people [10] .
Climate-wise, Migori County has two main rainy seasons. The first rainy season starts in March and ends in May, this season constitutes the long rains. The second rainy season starts in September and ends in November. The driest months are between December and February and June and September. The average daily temperature is usually a low of 24 degrees Celsius (74 F) and a high of 31 degrees Celsius (87 F). Rainfalls come in the afternoon and the heat is often dry and thus bearable. Migori County was chosen for this study due to its ease of access for field research, and also its prevalence for smallholder farmers who do not own the land they farm on.
3. Methodology
The methodology followed in this study as illustrated in Figure 2.
3.1. Data Collection
Data was collected in two phases. In phase 1, small scale farmers were identified through stratified random sampling and relevant data was collected from them, through a field questionnaire survey conducted in all the eight sub-counties of Migori County. The data collected from each respondent included personal information (such as name, identification, gender, presence or absence of bank account, etc.), occupation, any assets, types of crops grown and access to credit. 337 questionnaires were administered with 320 being completed and returned, giving a very good return rate of 95%. Interviews with key informants in the financial institutions and Credit reference bureaus in Kenya were also conducted, in order to get a feel of their awareness of and use of credit scoring.
In phase 2, Sentinel 2 imagery covering the study area was collected and LULC was classified using Google Earth Engine. The date of the imagery was November 2021. In addition, reference points for later use in image analysis were selected in a well-distributed pattern over the study area and positioned using hand-held GNSS. Phase 2 data collection started in Kuria sub-County, due to the rains and the poor road infrastructure there. 158 reference points were collected from seven sub-counties; Nyatike was excluded due to minimal farming there. To ensure even distribution of the points, it was decided to pick 5 points in each sub-County, representing the west, east, north, south and central areas of the sub-County; for example, in Suna East Sub-County, points were picked from Rabuor in the east, Nyarongi in the west, Godjope in the north, Witharaga in the south and Ngege in central. Two GNSS instruments were used during the reference data collection, a Trimble TDC100 mobile mapper and a Garmin Trex 10 hand-held. A research assistant was trained to operate the second instrument for the purpose of validation and backup of the data from the first instrument which was operated by the researcher herself. The image composite plus the collected reference points are shown in Figure 3.
3.2. Data Analysis
Data analysis was carried out in three steps, as indicated below.
3.2.1. Step 1: Identification of Smallholder Farms and Farmers
The smallholder farms and farmers that would later feature in the credit scoring analysis were identified according to the following criteria:
l Farms size was to be less or equal to 2 hectares.
l The farmers had to either have title to their farms or be farming on family land.
l For even distribution across Migori, 15 farms per sub-county were picked except for Nyatike sub-county where farming is minimal. In total, 101 farms were selected.
l The farms was to be in a cloud-free area in the imagery (as indicated by the corresponding reference data), to enable later generation of NDVI.
![]()
Figure 3. The imagery plus reference points.
3.2.2. Step 2: Image Classification and NDVI Generation
The image was classified using the supervised classification approach and the maximum likelihood classifier. The classification was based on the five classes shown in Table 1, and an overall classification accuracy of 0.861 was obtained. The classified image is shown in Figure 4.
Following the image classification, the classified image was masked to retain only Pixels covered by maize, and the masked image was then used to generate NDVI per pixel using the formula: NDVI = NIR − R/NIR + R.
The resulting NDVI image is shown in Figure 5.
On the assumption that any smallholder farm (≤2 ha) could fit within a 1 km × 1 km grid square, such a grid was overlaid onto the NDVI image and the average NDVI per grid square was computed; this would be the representative NDVI for any farm falling within that grid square. Table 2 shows an abstraction of the
![]()
Table 1. Image classification results (Classification accuracy = 0.861).
![]()
Table 2. Abstract of NDVI averages.
NDVI averages for maize farms within the study area as identified by the reference points that had been positioned within them.
3.2.3. Step 3: Development of a Score Model
This step involved the development of a score model that computes both the traditional score and the new score for any farmer, with the latter score computation including the geospatial variable NDVI. Machine learning using logistic regression was used. This is a predominant method in credit analysis and has become the benchmark method against the other methods for such credit analysis [11] .
This process involved modeling farmers’ data to come up with a credit scoring tool to assess the eligibility of a farmer accessing a loan, with additional emphasis on crop geospatial data/information. This was based on sample farmers from Migori County who were interviewed with regard to access to credit. Weight of Evidence (WOE), a statistical technique commonly used in credit scoring to evaluate the predictive power of various features or variables was employed. This method enabled the transformation of raw data into meaningful and informative predictors, providing a solid foundation for accurate credit risk assessment. Using the R programming language, the step-by-step processes which were undertaken in implementing credit scoring using WOE and the Generalized Linear Model are now outlined:
1) Start Variables
The start variables which were obtained from the questionnaire are shown in Table 3.
2) Correlation Testing
All the start variables were tested for correlation. This is because the predictor variables used in logistic regression should, ideally not be correlated. Correlation among such variables can cause model problems such as multi collinearity, leading to unstable and unreliable estimates of the regression coefficients. In such cases, the regression coefficient may change dramatically with minor changes to the data.
3) Final Variable Selection
Following the dropping of correlated variables, six final predictor variables were selected based on their information value (IV) and domain knowledge (i.e. knowledge of the credit scoring industry). IV is a tool in machine learning used to assess the predictive power of any variable for a given feature in a dataset and is often used in credit scoring. It quantifies the extent to which a variable can differentiate between different outcomes, such as default and non-default, and provides insights into the variable’s contribution to the predictive model’s performance. IV depends on a variable’s weight of evidence (WOE).
Statistically,
and IV = (% of non-events − % of events)
In the case of separating loan defaulters from non-defaulters, default represents an event while non-default represents a non-event.
Generally the higher the IV, the better the variable for the intended prediction; however the best IV values range between 0.3 and 0.5; although values between 0.1 and 0.3 may also be accepted for model development as medium predictors [11] .
Following the WOE/IV analysis, the variables (Land parcel, Collateral, Reasons and How long) were selected for traditional score model development. The NDVI would be deliberately introduced for computing the new score. NDVI (Normalized Difference Vegetation Index) is an important index used in remote sensing to assess and monitor vegetation health and vigour, and it is the geospatial variable
that was introduced into the model to indicate whether the farmer had healthy crops or not.
4) Data Modelling
After converting the original data points to WOE, a Generalized Linear Model was fitted based on the following logistic regression formula [12] .
The R link function used to generate results is the logit value.
Script 1: Data Modelling
The output of this model, which was in logit was converted to odds and then probability as follows:
Probability
After running the model, the coefficients achieved were in terms of log odds; which were converted to probabilities and then final scores (scaling). These scores were originally in the scale of 0 to 1000. The probabilities were converted to scores using the Scorecard package in R using the following approach (script 2).
Replace odds with logit.
Ts—target score.
To—target odds.
Pdo—slopes.
Script 2: Score Computation
The original scores were finally scaled to the FICO score, which is more universal and frequently used [13] .
A FICO score is a three-digit number, typically in a 300-850 range, that tells lenders how likely a consumer is to repay borrowed money based on their credit history. Only FICO Scores are created by the Fair Isaac Corporation and are used by over 90% of top lenders when making lending decisions.
The FICO score range is explained in Table 4.
Out of the 101 farmers and farms identified, 67 (about 2/3) were chosen for training the model, with the remaining 34 reserved for later testing of the model.
5) Model Evaluation
The performance of a binary classification model, such as the regression model used in this study, is often evaluated by plotting the Receiver Operating Characteristic (ROC) curve and determining the Area under the Curve (AUC). The curve checks how well the model is able to distinguish and separate events from non-events; it is a plot of the rate of true positives (events correctly predicted to be events) on the y-axis against the rate of false positives (non-events wrongly predicted to the events) on the x-axis. Generally the higher the AUC curve i.e. the bigger the area under the curve, the better the model, and 75% is the recommended acceptable minimum. Script 3 was used to determine this AUC. A related evaluation measure is the Gini coefficient defined as: Gini = (2 * AUC − 1).
Again, a higher Gini represents a better predictive model.
## A performance instance
## ‘Area under the ROC curve’
####Make prediction
Testing of the predictive power of the model on test data that had been earlier isolated is done.
## [1] “K-S Statistic = 0.846718005133847”
## [1] “Area under the curve = 0.961037770443711”
## [1] “Gini Coefficient = 0.922075540887423”
## (Intercept) Landparcels Landhold Collateral Creditrepayment
## −5.2637155148 −1.5550051782 −0.5318544217 0.4488074659 −0.6338934531
## Reasons Acquire Howlong Outstanding Set
## −2.6601538908 0.6818583697 0.3192265714 3.1808766951 −0.0006021895
## Pest Organic
## 0.0635936312 1.5129961303
Script 3: Evaluating the model by AUC
4. Results
The final model developed is shown in Table 5, with the indicated predictor variables.
The model was trained on 67 farms and farmers; an abstract of the results is shown in Table 6.
In respect of Table 6, it should be noted that the model was first run without NDVI, to generate the “traditional” score column. NDVI was then introduced, to generate the “with 1.0 NDVI” column.
For experimental purposes and in view of the central role that NDVI was to play in this whole arrangement, the weight of NDVI was deliberately biased to 1.2, 1.3, 1.4 and 1.5 the original weight and this generated the correspondingly labeled columns.
The performance of the model in this training was evaluated by plotting the ROC curve and determining the AUC; the resultant curve is shown in Figure 6.
![]()
Table 5. Credit scoring model statistics.
![]()
Table 6. Abstract of model training results.
From the curve, the AUC was found to be 95%, with a Gini Coefficient of 0.9.
Following these encouraging results from model training, the model was now tested on the remaining 34 farms and farmers, and the results are shown in Table 7.
5. Discussion
The credibility of the scoring model developed is evidenced by the ROC curve in Figure 6, whose high AUC indicates a good choice of predictors. The results in Table 7 show that out of the 34 farmers scored, 14 were poorly scored (below 580) and hence would not be recommended for credit. On introduction of the NDVI variable, nearly half of all farmers had their scores improved, and one farmer was able to move to a score of over 580 and hence become eligible for credit. On enhancing the weight of NDVI up to 1.5 NDVI, 5 more farmers were able to transition to eligible status. Therefore by the end of the NDVI experiment, 6 farmers out of the originally ineligible 14 had changed status and could now get credit; the level of non-eligibility had therefore changed from 41% to 24%. This is a promising result that can be built upon if the lending industry was to warm up to it.
However, since the model is developed using machine learning techniques, it needs fine tuning using a larger sample of farmers, and also testing in more geographic locations and agro climatic zones.
6. Conclusion
The results achieved from this study have demonstrated that NDVI can be a useful tool in improving smallholder farmers’ credit scoring. Lending institutions can adopt the model in order to improve the chances of smallholder farmers’ access to credit for the sustenance of their farming activities. However further research is needed in order to fine-tune the model using a larger data set.
Acknowledgements
The authors acknowledge the Gandhi Smarak Nidhi Fund (GSNF) and the Artificial Intelligence for Development (AI4D) programme (Funded by the IDRC and SIDA, and managed by the African Centre for Technology Studies (ACTS)) for providing the funding that enabled this research work.