BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//talks.osgeo.org//foss4g-2026//speaker//3WMEC3
BEGIN:VTIMEZONE
TZID:JST
BEGIN:STANDARD
DTSTART:20000101T000000
RRULE:FREQ=YEARLY;BYMONTH=1
TZNAME:JST
TZOFFSETFROM:+0900
TZOFFSETTO:+0900
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-foss4g-2026-AMUSYL@talks.osgeo.org
DTSTART;TZID=JST:20260903T113000
DTEND;TZID=JST:20260903T120000
DESCRIPTION:Accurate land-cover classification maps are essential geospatia
 l resources that support a wide range of societal applications\, including
  urban development\, environmental monitoring\, and disaster risk manageme
 nt. With the growing availability of satellite remote sensing data\, pixel
 -wise classification methods\, which assign a single land-cover class to e
 ach image pixel\, have become widely adopted. However\, when medium-resolu
 tion satellite imagery\, such as Sentinel-2\, is used\, a single pixel oft
 en contains more than one land-cover type. This phenomenon\, commonly refe
 rred to as the mixed-pixel problem\, leads to the loss of information abou
 t minority classes within a pixel\, making it difficult to accurately char
 acterize fine-scale surface conditions and limiting the practical utility 
 of classification results. One promising solution to this issue is composi
 tional classification\, which estimates the proportion of each land-cover 
 class present within a pixel rather than assigning a single label. This ap
 proach retains information about all classes\, including those that occupy
  only a small fraction of a pixel. Compositional data are subject to two m
 athematical constraints: all values must be non-negative\, and the proport
 ions across all classes must sum to one. Yet\, the implementation of such 
 compositional classification is challenging\, as machine learning or deep 
 learning model architectures do not consider these characteristics properl
 y. This study proposes a deep learning framework for compositional land-co
 ver estimation at 10 m spatial resolution. We explore combinations of inpu
 t features\, model architectures\, and loss functions suited to compositio
 nal outputs. Rather than seeking a single definitive solution\, we aim to 
 provide practical insights into how these design choices interact and infl
 uence estimation quality. \n\nAs an experiment\, we examined two inputs\, 
 three models\, and two loss functions. Two types of 10-m level input featu
 res were examined. The first was spectral reflectance data from 10 multisp
 ectral bands (B2-8\, B8A\, B11\, and B12) in Sentinel-2\, which directly c
 apture surface characteristics across visible and infrared wavelengths. Th
 e second was Embedding V1\, a 64-dimensional feature vector generated by A
 lphaEarth Foundations\, a geospatial embedding model that integrates multi
 ple data streams\, including optical\, radar\, and LiDAR observations from
  multi-temporal satellite imagery. We used OpenEarthMap (OEM)\, a publicly
  available global dataset with high-resolution eight land-cover labels at 
 0.25–0.5 m spatial resolution\, as reference data. To generate training 
 targets at medium resolution\, we aggregated all OEM pixels within each 10
  m pixel and computed the fractional coverage of each land-cover class. Th
 is aggregation produced pixel-level composition vectors that served as gro
 und truth for supervised learning\, establishing a direct correspondence b
 etween medium-resolution inputs and high-resolution reference proportions.
  Three deep-learning model architectures were tested: a multilayer percept
 ron (MLP)\, a two-dimensional convolutional neural network (2D-CNN)\, and 
 a three-dimensional convolutional neural network (3D-CNN). All models cons
 isted of two fully connected layers followed by a Softmax activation funct
 ion to ensure outputs formed valid compositional proportions. The CNN-base
 d models were included to capture spatial context from neighboring pixels\
 , while the MLP was used as a simpler baseline that processes each pixel i
 ndependently. The dataset was split into 70% for training\, 10% for valida
 tion\, and 20% for testing. We compared two loss functions: Mean Absolute 
 Error (MAE)\, a widely used metric in regression tasks\, and the Aitchison
  distance\, a measure derived from compositional data analysis. The Aitchi
 son distance evaluates the relative differences between components rather 
 than their absolute errors by applying a logarithmic transformation and ce
 ntering to each proportion. To eliminate redundancy associated with the co
 mpositional constraint\, we applied a linear transformation using an ortho
 normal basis\, projecting the data into a D−1 dimensional space. This me
 tric enables stable distance computation while reflecting the geometric st
 ructure and constraints inherent to compositional data. This combination o
 f architectures and loss functions allowed for a systematic comparison acr
 oss multiple design dimensions. We evaluated the performance of each combi
 nation using MAE and the Aitchison distance to identify the most effective
  framework for compositional land-cover classification mapping. All proces
 sing and training are implemented with open-source Python geospatial/ML to
 ols\, and we will release code and experiment configurations to support re
 producibility.\n\nThe experimental results revealed clear differences base
 d on the combination of inputs and model architecture. When using only Sen
 tinel-2 spectral bands as input\, 3D-CNN models outperformed the MLP (MAE 
 of 0.1126 versus 0.1185)\, indicating that spatial context from convolutio
 nal operations benefits models when input features are limited to raw refl
 ectance. In contrast\, when using Embedding V1 as input\, the MLP achieved
  the best performance (MAE of 0.0989) among all tested models. This sugges
 ts that Embedding V1 already encodes rich spatial and semantic information
  through its pretraining process\, making simpler models better suited to 
 leverage these representations without overfitting to their internal struc
 ture. Regarding the choice of loss function\, the resulting maps showed a 
 meaningful difference between MAE and Aitchison distance. Models trained w
 ith MAE tended to produce smoothed predictions\, with estimated proportion
 s concentrated near intermediate values and rarely approaching 0 or 1. Thi
 s smoothing effect is a known limitation of MAE in compositional contexts\
 , where extreme values are physically meaningful. When the Aitchison dista
 nce was used instead\, the estimated proportion maps showed sharper contra
 sts and more realistic spatial patterns that better matched the actual dis
 tribution of land-cover classes. These results suggest that using a loss f
 unction aligned with the mathematical properties of compositional data lea
 ds to more credible and interpretable outputs. \n\nThe findings demonstrat
 e that the combination of input features\, model architecture\, and loss f
 unction substantially impacts the quality of fractional land-cover estimat
 ion. Under our experimental conditions\, the MLP model using Embedding V1 
 and the Aitchison distance achieved the highest accuracy\, providing pract
 ical insights into the design of compositional land cover classification f
 rameworks in remote sensing.
DTSTAMP:20260717T225802Z
LOCATION:Cosmos2
SUMMARY:Aitchison-Loss Training with Geospatial Embeddings Sharpens Composi
 tional Land-Cover Maps - Narumasa Tsutsumida\, Ayato Kanno
URL:https://talks.osgeo.org/foss4g-2026/talk/AMUSYL/
END:VEVENT
END:VCALENDAR