Segmentation of arbitrary features in very high resolution remote sensing imagery

TL;DR

EcoMapper introduces a scalable, automated pipeline for segmenting arbitrary features in very high resolution remote sensing imagery, addressing the transferability gap of context-specific DL models. It leverages MMSegmentation with Mask2Former in a CLI-based workflow that preprocesses geospatial data, trains models with class-weighted sampling and augmentations, and postprocesses predictions across overlapping tiles. A key contribution is the Cording Index, which links feature size to an optimal ground sampling distance, enabling a DL-informed field survey workflow and guiding data collection under resource constraints. On a real UAV dataset, EcoMapper achieves competitive segmentation performance without dataset-specific tuning and demonstrates how resolution, feature size, and survey extent influence results, informing practical deployments in biodiversity, precision agriculture, and sustainable land management.

Abstract

Very high resolution (VHR) mapping through remote sensing (RS) imagery presents a new opportunity to inform decision-making and sustainable practices in countless domains. Efficient processing of big VHR data requires automated tools applicable to numerous geographic regions and features. Contemporary RS studies address this challenge by employing deep learning (DL) models for specific datasets or features, which limits their applicability across contexts. The present research aims to overcome this limitation by introducing EcoMapper, a scalable solution to segment arbitrary features in VHR RS imagery. EcoMapper fully automates processing of geospatial data, DL model training, and inference. Models trained with EcoMapper successfully segmented two distinct features in a real-world UAV dataset, achieving scores competitive with prior studies which employed context-specific models. To evaluate EcoMapper, many additional models were trained on permutations of principal field survey characteristics (FSCs). A relationship was discovered allowing derivation of optimal ground sampling distance from feature size, termed Cording Index (CI). A comprehensive methodology for field surveys was developed to ensure DL methods can be applied effectively to collected data. The EcoMapper code accompanying this work is available at https://github.com/hcording/ecomapper .

Figures (18)

• Figure 1: A high level overview of EcoMapper's architecture. Data pre- and post-processing, as well as model training, evaluation, and inference are fully automated.
• Figure 2: Overview of the labeling process. Top row: QGIS labeling. (a) Partial view of a Chayote plantation in the Sto. Niño region; (b) an overlay of the manually drawn label map for Chayote, labels were palettized for visualization. Blue indicates "Chayote", red indicates "Border" (uncertainty). Bottom row: CVAT labeling. (c) Input image; (d) points are placed indicating the feature to label; (e) the label (cyan) is generated automatically by CVAT.
• Figure 3: Methods of resolution degradation. The original image (top left) can be downsized (A), downsized and upscaled to the original tile dimensions (B), or the orignal orthomosaic can be downsized and split into tiles anew, yielding fewer tiles that cover more spatial distance and appear "zoomed out" (C).
• Figure 4: Training, validation, and test sets after manually labeling and splitting the Sto. Niño datset. Ground truth labels are indicated, showing a favorable distribution of classes in all three sets for both features. Spacing was introduced between the train and evaluation sets to reduce spatial leakage.
• Figure 5: EcoMapper performance in different tasks. (a) Image splitting using a stride of 0.5; (b) prediction merging, input size indicates total size of all tiles; (c) GPU training with PyTorch via MMSegmentation.