Student Research Projects

This page showcases a selection of Bachelor’s theses and Master’s dissertations proposed and supervised within the Geosense Laboratory. Our goal is to encourage students to explore innovative research directions in Earth Observation, geospatial analysis, environmental monitoring, and AI-assisted data interpretation.

Each topic is designed to combine scientific relevance with practical applications, often linked to ongoing research projects or open datasets developed by our team.

Students are welcome to adapt existing topics, propose new ideas, or collaborate with Geosense researchers on interdisciplinary challenges.

Bachelor's theses

1. Reconstruction of Spectral Index Time Series (NDVI, EVI, NDWI) from Sentinel-2 Imagery Using GAN Networks

Objective.To implement a spatio-temporal Generative Adversarial Network (GAN) that integrates spectral, spatial, and temporal information from Sentinel-2 image series in order to generate continuous and realistic vegetation index values. The model aims to ensure temporal consistency between successive observations and spatial coherence within each patch, contributing to the creation of a uniform time series suitable for phenological analyses and precision agriculture applications.

Expected results. The project aims to achieve realistic reconstruction of spectral index time series, robust to missing observations and cloud coverage, resulting in a quasi-uniform, spatio-temporally coherent sequence. The outputs are valuable for phenological studies, agricultural productivity estimation, and vegetation monitoring in the context of climate change. (Coordonator. Sl. Andreea GRIPARIS)

2. Prediction of Vegetation Indices (NDVI, EVI, NDWI) Based on Seasonal Dynamics Using LSTM and Transformer Networks.

Objective. A short- and medium-term prediction approach for vegetation indices will be developed using sequential neural networks (LSTM) and Transformer architectures. These models are designed to capture nonlinear temporal relationships and the seasonal dynamics of vegetation. The input time series will be irregularly sampled; therefore, temporal encoding mechanisms will be employed to handle uneven observation intervals.

Expected results.
The proposed model will provide accurate and consistent predictions of vegetation indices, adapted to seasonal dynamics and robust to missing observations. ( Coordonator Conf. Daniela Faur)

3. Integration of Retrieval-Augmented Generation (RAG) Methodology in Satellite Image Analysis

Objective.
This work aims to integrate the Retrieval-Augmented Generation (RAG) methodology into satellite image analysis by combining visual, textual, and metadata information associated with Earth Observation (EO) data. The goal is to develop a multimodal framework that enables contextualized querying of EO datasets and the generation of explanatory, data-driven insights based on both imagery and auxiliary sources. A RAG architecture adapted for satellite imagery will be implemented, combining a retrieval module—responsible for semantic search across image–caption–metadata collections—with a generative Large Language Model (LLM) capable of producing coherent and context-aware explanations. In the preprocessing stage, Sentinel-2 images will be annotated with semantic captions and spatio-temporal metadata, which will be used for contextual indexing and retrieval. The system’s performance will be tested in two application domains: precision agriculture and urban monitoring. Evaluation criteria include the accuracy and contextual relevance of the generated responses, as well as the model’s ability to link visual and linguistic information effectively.

Expected Results.
Integrating RAG into satellite image analysis will enable the automatic extraction of contextualized insights from EO data, facilitating the semantic interpretation of satellite scenes and supporting the development of intelligent assistants for Earth Observation data analysis and explanation. ( Coordonator Conf. Daniela FAUR, Dr. Teodor COSTĂCHIOIU)

4. Temporal Semantic Segmentation of Agricultural Crops at Pixel Level Using Deep Convolutional and Spatio-Temporal Networks

Objective. To develop and evaluate deep learning models for semantic segmentation of agricultural crops at the pixel level, using Sentinel-2 multispectral time series from the AgriSen-COG dataset. The study aims to extend and test the methodology on regions in Romania, analyzing the transferability of models across different European agro-climatic zones.

Three advanced architectures for temporal semantic segmentation will be implemented and compared:
U-Net — for spatial segmentation with a convolutional encoder–decoder structure; ConvLSTM — for modeling temporal dynamics at the pixel level and ConvStar — a deep spatio-temporal network capable of integrating spectral, spatial, and temporal information.

Expected results. Accurate and temporally consistent crop classification maps are expected, demonstrating the potential of spatio-temporal deep learning methods for scaling AgriSen-COG models to national-level applications. (Coordonator Conf. Daniela FAUR)

5. Classification of Parcel-Level Aggregated Time Series Using LSTM, Transformer, and TempCNN Models

Objective. To implement and compare sequential deep learning models for crop-type classification based on time series of vegetation indices and climatic variables aggregated at the parcel level. The study uses data from the AgriSen-COG dataset and explores model transferability to agricultural parcels in Romania.

Three sequence-oriented architectures will be evaluated: LSTM (Long Short-Term Memory) — to capture temporal dependencies and crop seasonality; Transformer — using attention mechanisms and temporal encoding for learning long-range relationships; TempCNN (Temporal Convolutional Neural Network) — for automatic feature extraction from multivariate time series. The input data will include vegetation index time series and climatic parameters (temperature, precipitation), aggregated by parcel.

Expected results.
The study will identify models with the strongest spatio-temporal generalization capability and determine the most agronomically relevant variable combinations. The results will support the development of a robust, transferable crop-type classification system for monitoring crop dynamics and seasonal vegetation patterns at the national scale. (Coordonator Conf. Daniela FAUR)

Dissertations

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut elit tellus, luctus nec ullamcorper mattis, pulvinar dapibus leo.