Crop Field Segmentation

Motivation

Crop field segmentation is relevant for planning crop production and better managing the areas dedicated to agriculture. While satellite imagery has greatly improved our capabilities in this task, we still need to effectively investigate which plantations are more sensitive to which hyperspectral information. Moreover, the investigation of which plantations are sensitive to which spectral bands were not exploited, particularly since we do not know if certain crops are easier than others. The time series of images was not investigated to determine which images were more relevant to identifying the crops.

Task

1. Investigate the relationship between crops and bands
2. Investigate a time series of images, which are the most significant ones
3. Investigate if certain crops are easier than others
4. Investigate if the models are biased toward certain crop characteristics

Reference

KAN You See It? KANs and Sentinel for Effective and Explainable Crop Field Segmentation

Hyperspectral Canopy Height Maps

Motivation

Estimating canopy height and height change at meter resolution from satellite imagery has numerous applications, such as monitoring forest health, logging activities, wood resources, and carbon stocks. While using LiDAR sensors remains a costly but effective solution, new solutions are now being designed to provide reliable and cheaper solutions. RGB imagery was already investigated, and also depth estimation foundational models. Light conditions can strongly affect RGB, while other frequencies do not suffer. However, depth estimation models (like depth anything) are trained on natural RGB imagery.

Task

Adapt Depth Estimation models to hyperspectral imagery without starting from scratch.

Reference

1. Open-Canopy: A Country-Scale Benchmark for Canopy Height Estimation at Very High Resolution
2. Depth Any Canopy: Leveraging Depth Foundation Models for Canopy Height Estimation

Better Losses Better Language Models

Motivation