Magnifier: A Multi-grained Neural Network-based Architecture for Burned Area Delineation

TL;DR

The paper addresses the challenge of burned area delineation with limited labeled data by introducing Magnifier, a dual-encoder, multi-granularity architecture that fuses global and local features before a shared decoder. Magnifier is architecture-agnostic and can be layered on top of existing encoder-decoder models, delivering consistent IoU and F1 improvements while maintaining lower complexity than large single models. Extensive experiments on CaBuAr, Europe, and Indonesia datasets show Magnifier providing notable gains (e.g., +2.65% average IoU) and robustness across CNN and transformer backbones, with favorable computational efficiency. The authors also analyze data efficiency, transfer learning across continents, and demonstrate Magnifier’s potential to generalize to other remote sensing tasks, underscoring its practical impact for resource-constrained, real-world disaster monitoring. Overall, Magnifier offers a principled, data-efficient route to improve semantic segmentation in EO where labeled data are scarce and computational resources are limited.

Abstract

In crisis management and remote sensing, image segmentation plays a crucial role, enabling tasks like disaster response and emergency planning by analyzing visual data. Neural networks are able to analyze satellite acquisitions and determine which areas were affected by a catastrophic event. The problem in their development in this context is the data scarcity and the lack of extensive benchmark datasets, limiting the capabilities of training large neural network models. In this paper, we propose a novel methodology, namely Magnifier, to improve segmentation performance with limited data availability. The Magnifier methodology is applicable to any existing encoder-decoder architecture, as it extends a model by merging information at different contextual levels through a dual-encoder approach: a local and global encoder. Magnifier analyzes the input data twice using the dual-encoder approach. In particular, the local and global encoders extract information from the same input at different granularities. This allows Magnifier to extract more information than the other approaches given the same set of input images. Magnifier improves the quality of the results of +2.65% on average IoU while leading to a restrained increase in terms of the number of trainable parameters compared to the original model. We evaluated our proposed approach with state-of-the-art burned area segmentation models, demonstrating, on average, comparable or better performances in less than half of the GFLOPs.

Figures (8)

• Figure 1: Average Mean IoU vs. Number of parameters. The Average Mean IoU has been computed considering all datasets. Architecture Type refers to the family of base networks employed for segmentation. On average, the Magnifier backbone achieves IoU improvements compared to MobileNetV3 Small and Large, ResNet-18 and 101, and MiT-B0 and B1 without increasing the number of parameters too much.
• Figure 2: RGB samples taken from the three datasets with the corresponding binary ground truth.
• Figure 3: Distribution of wildfires in the analyzed datasets. In (a) the areas covered by wildfires in red. In (b), the location of the wildfires, where each color represents the fold it belongs to. In (c), the areas covered by the dataset are in red squares.
• Figure 4: Magnifier architecture. In the lower branch, (i) the image is cropped in smaller patches (as shown in \ref{['fig:cropping']}), giving each patch to an encoder. (ii) The encodings are concatenated by putting each one in the original position in the image (as shown in \ref{['fig:recompose']}). In the upper branch (iii), the entire image is given to an encoder. (iv) The two encodings are concatenated along the channel axis, and (v) they are given to the decoder to get the final prediction.
• Figure 5: Cropping procedure. The image is cropped in patches, and each of them keeps the original position associated.