From Global Noise to Local Accuracy: An Abstaining Classifier Approach for Robust Forest Mapping with Noisy Global Data

Accurate forest mapping is frequently hindered by the label noise inherent in large-scale global land cover products and the scarcity of high-quality local ground-truth data. This paper presents a novel AI-driven framework that effectively addresses this challenge by synergistically leveraging the broad coverage of noisy global datasets with a small, trusted set of clean annotations. Our approach utilizes a DeepLabV3+ architecture with Sentinel-2 multispectral imagery and derived vegetation indices as input.

The core of our methodology lies in a hybrid data strategy and a specialized composite loss function. We employ strategic batch sampling to prioritize learning from the clean dataset (85% of each batch) while still benefiting from the contextual coverage of the noisy data. Our composite loss function, integrating Dice Loss, Categorical Focal Loss, and a Deep Abstaining Classifier (DAC) Loss, is explicitly designed to manage label noise and boundary uncertainty by empowering the model to abstain from predictions on highly uncertain pixels.

The framework's efficacy was validated on the complex forested landscapes of North Africa, a region under-represented in many global training datasets. Our model achieved a validation F1-score of 91% and an Intersection over Union (IoU) of 83% on the held-out clean data. Critically, it achieved a recall of 95.4%, substantially minimizing the omission of forest areas compared to a baseline U-Net. This work provides a reliable, scalable methodology for refining large-scale, imperfect datasets, offering a robust solution for forest monitoring in data-challenged regions worldwide.