Linnea flowers, graphic.

Doctoral project: AI in administration of agricultural subsidies

We want to design, implement and evaluate systems based on artificial intelligence that supports our customer with the administration of agricultural subsidies. Therefore, we focus on computer vision in order to analyze aerial photos and on natural language processing in order to get information form several documents according to our customers needs.

Project information

Doctoral student
Niels Gundermann
Supervisor
Professor Welf Löwe, Linnaeus University
Assistant supervisors
Professor Johan Fransson, Linnéuniversitetet, Erika Olofsson, senior lecturer, Linnéuniversitetet, Prof. Dr.-Ing. Andreas Wehrenpfennig, Hochschule Neubrandenburg - University of Applied Science
Participant organizations
Data Experts, Linnaeus University
Financiers
Data Experts
Timetable
September 2023 – September 2028
Subject
Computer and information science (Department of Computer Science and Media Technology, Faculty of Technology)
Research group
Data Intensive Software Technologies and Applications (DISTA)
Linnaeus University Centre
Linnaeus University Centre for Data Intensive Sciences and Applications (DISA)

More about the project

Data experts support state authorities in the administration of agricultural subsidies. They have various artificial intelligence (AI) based computer vision, text classification and interpretation tasks that the thesis project researches, develops, and adopts. Computer vision tasks include

1. Recognition of agriculture parcels: Every year, the authorities manually compare the existing record of all agricultural areas (reference parcel) with current aerial photographs. A decision is made as to which of the reference parcels must be adjusted because they no longer fit the geographical conditions. The idea is to support this comparison task with AI. There are several levels of support possible:

  • Recognition of parcels where probably no changes need to be made.
  • Recognition of parcels where a change is (very) likely needed.
  • Propose changes for reference parcels to be adjusted.

2. Control tasks connected to applicant order:

a. Recognition of agriculture-photovoltaic systems: Areas are subsidized that are partially covered by photovoltaic systems but allow further agricultural use. These areas/systems could be recognized by AI.

b. Recognition of flower strips: For control purposes, flower strips can be recognized by aerial photography. The idea is to recognize them using an AI.

c. Crop rotation analysis: From 2024, crop rotation should be carried out according to certain criteria. Whether the crop rotation is taking place can be seen on aerial photographs. The idea is to use an AI to recognize the crop and, hence, the correct implementation of rotation.

d. Detection of tree harvesting: Within the forest support programs, the harvest could be monitored by means of regularly taken aerial photographs.

e. Validation of planting: The applicant submits pictures as evidence of the planting of the arable land. The authority uses this image to check whether the requested use can be subsidized. The identification of the planting in the picture and other relevant criteria can be checked by an AI. In addition, the verification of the location at which the respective photo was taken poses a problem. The verification means currently used (geo-tagging) can be manipulated. One could try to compare static objects visible on the submitted photo (e.g., masts, high buildings, etc.) with aerial photos or corresponding register (cadastres).

f. Stripe detection: Rows of trees, hedges, etc. are particularly eligible for funding. They can be located in the middle of the area and then are hard to recognize. Based on aerial photos, these stripes can be detected. Text classification and interpretation tasks include:

3. Evaluation of documents/evidence to be submitted: If the applicant submits a document that is not formalized (certificates, handwritten letters, emails), this document must be categorized manually and assigned to a processor. The idea is to support this process with an AI that is trained based on the properties of the document. Examples of submitted non-formalized documents are: Invoices, Evidence documents, Planning documents, Legal Remedies, Inquiries, and even erroneously received documents (e.g. incorrectly addressed).

4. Validation of certain information: Documents contain several information for the administration and calculation of subsidies. Actually, humans collect this information manually. The idea is to support this process of information retrieval with an AI.

The doctoral project is performed within the industry graduate school Data Intensive Applications (DIA) Data Intensive Software Technologies and Applications (DISTA) and Linnaeus University Centre for Data Intensive Sciences and Applications (DISA)