Farmer Credit Worthiness (Identifying Farmer Hotspots)

FARMER CREDIT WORTHINESS

Unlike in urban areas, rural areas lack reliable credit rating facilities, which poses one of the biggest challenges for lending firms in disbursing financial assistance to smallholding farmers. Lack of parameters, which help lending firms in analyzing the credit worthiness of an individual, has not only deprived farmers in getting financial help, but has acted as a bottleneck for lending firms, who aim to expand business in India’s rural areas and reduce percentage of non-performing assets (NPAs). The two of the most important road blocks for a bank in this domain are:

• Prior to the disbursement of loans, the firms would like to verify and analyze the details of loan applicant, such as personal details, ownership of land, plot no., cropping pattern, crop intensity, acreage, yields etc. The entire process is time consuming and requires extensive manpower to travel from one village to another.

• Once the loan is disbursed, the institutes try to evaluate the performance of villages/plots of a given farmer – which also requires periodic visits to farm land to capture the crop progress.

Farmer Credit Worthiness solution by RMSI Cropalytics not only helps in determining the farmers’ worthiness in a given village, but also assists firms in understanding a village’s profile. The solution helps in mitigating the challenges faced by both the stakeholders (farmers) and lending firms.

APPROACH

RMSI Croplaytics’ approach is to conduct the analysis from village up to cadastral level. We have the following approach for credit rating worthiness utilized during pre-lending screening:

1. Model that is scalable and reliable
2. Making information available with the stakeholders well in time, to make lending decisions
3. The AI/ML model is objective (based on verifiable, measurable quantities) rather than being based on modelled values or estimations, to the extent possible.
4. Instead of quantifying estimated yield at the cadastre level, relative grading is taken up which is a sufficient assessment for this purpose.

Deep Machine Learning (ML) techniques were used to develop advanced predictive models for recognizing various features in geospatial imagery with greater efficiency and precision. Vegetative Indices (VIs) are considered as an indicator of crop health which is widely used in crop yield estimation models.

Detailed ML-based predictive modeling approach has been discussed below:

1. Data Treatment: To impute the missing data in the training data, we tried multiple imputation techniques, including mean and median, and found that the median approach led to the most accurate predictions.
2. Data Analysis: To compare the individual importance of vegetative indices, soil and weather components in the yield prediction, we obtained the yield correlation with various inputs using standard scaling followed by Pearson’s correlation coefficient
3. Build Predictive model: We developed the models using two crop growth stage inputs, one crop stage input, using three ML techniques separately.
4. Model fine tuning: In this step, the models’ performance was improved by fine tuning various hyper- parameters.
5. Model validation: To examine the yield prediction error for individual regions, we obtained prediction error across various CCE locations existed in the validation dataset. The prediction error was consistently low (RMSE) for most of locations. The overall model performance is evaluated using validation R-square and RMSE value.
6. Artificial Neural Networks (ANN) based modelling approach
7. Partial Least Squares Regression (PLSR) based modelling approach
8. Random Forests (RF) based modelling approach

Basis this, the risk scores are assigned to Cadastres/ villages/ tehsils.

The approach helps identify and classify the cadastre’s in different buckets, thereby allowing the lenders to evaluate whether they can disburse loans to farmers belonging to specific cadastres.

Loanee Farmer Verification

Every crop season, the banks/lenders/insurers have to verify a large number of farmer loan applications in a limited time window. This industry requirement led RMSI to develop an AI/ML based automated solution. In RMSI’s experience of verifying millions of policies against land records, RMSI found that there are several mismatches in policy applications and land record details. In certain cases, RMSI found up to 40% mismatches (of the entire policy dataset), including large mismatches in loanee applications. These mismatches are:

1. Village name not found in land record
2. Survey Number not found in land record
3. Applicant name not found – The application name is not part of the land record document.
4. Name matching partially – This is tricky – Last name is same but first name doesn’t match, or name order doesn’t match, or name form doesn’t match.

In cases of partial match, the Insurance Companies usually accept the loan application for insurance. This category is very large and presents the greatest challenge for an automated algorithm.

1. Excess Area – The sum of multiple applications against the same survey number results in application area higher than acreage shown in land records. Sometimes, even a single application has acreage higher than what is mentioned in the land records. This is also a very large category of mismatches. It happens because joint owners take separate policies against the same land, or the same owner takes multiple policies, or sometimes, loanee farmers take an additional insurance policy from the CSC.

There are many challenges in comparing millions of loanee farmer applications with land ownership records.

• No common link between loanee data and state land records data: In state land records data, numerical village codes have not been used, and village names are in the local language (Telugu, etc.). Policy database contains numerical village codes. To join the two databases, a master list needs to be interfaced with the village names in the land records to assign numerical village codes. Often, there are multiple villages with the same name in the same Tehsil which needs to be resolved.

• Even after joining the two databases state by state, the transliteration and comparison of names between the databases is the true challenge. It is relatively easy to achieve a reasonable transliteration, but the real challenge is in comparing the name on the policy applications with the names appearing in the land records, and deciding whether the name matches or not. In many cases, the name form is different, or the father’s name appears, or the name order is different or the spelling is different from the transliteration.

Proposed Solution:

In order to compare the loanee farmer database (English names) with Land Records (vernacular script), RMSI has already created comparison algorithms in the following languages: Marathi, Oriya, Telugu (AP, Telangana), Hindi (Rajasthan, Bihar, Jharkhand, UP, MP, Chhattisgarh, Haryana), Bengali (West Bengal, Tripura) and Assamese. (Tamil, Kannada and Gujarati are in progress). RMSI has applied for a patent for its comparison algorithm.

RMSI has now deployed AI/ML based models for rapid loanee farmer verification against land records, through which we are able to verify millions of farm loan / insurance applications within weeks.

-----------------------------------------------------------------------------------
Please refer to the detailed Solution Approach Note (with images), Solution Output and Presentation in the zipped folder for more information.