\documentclass[12pt,a4paper]{article} % ---------- Packages ---------- \usepackage[top=1.45in,bottom=1in,left=1in,right=1in,headheight=70pt,headsep=0.05in]{geometry} \usepackage{times} \usepackage{setspace} \usepackage{graphicx} \usepackage{caption} \usepackage{amsmath} \usepackage{amssymb} \usepackage{multirow} \usepackage{array} \usepackage{xcolor} \usepackage{hyperref} \hypersetup{colorlinks=true, urlcolor=blue, citecolor=black, linkcolor=blue} \usepackage{enumitem} \usepackage{titlesec} \usepackage{fancyhdr} \usepackage{booktabs} \usepackage{tikz} \usetikzlibrary{shapes.geometric, arrows.meta, positioning, calc, fit, backgrounds} % ---------- Header: banner image on every page ---------- \pagestyle{fancy} \fancyhf{} \renewcommand{\headrulewidth}{0pt} \fancyhead[C]{\includegraphics[width=\textwidth]{image.png}} % ---------- Section heading style ---------- \titleformat{\section}{\fontsize{12}{14}\bfseries}{}{0pt}{} \titlespacing*{\section}{0pt}{12pt}{6pt} \setlength{\parindent}{0pt} \setlength{\parskip}{8pt} % ---------- TikZ pipeline styles ---------- \tikzset{ src/.style = {cylinder, shape border rotate=90, draw, fill=orange!25, minimum width=1.1cm, minimum height=1.0cm, align=center, font=\scriptsize}, proc/.style = {trapezium, trapezium left angle=70, trapezium right angle=110, draw, fill=yellow!35, align=center, font=\scriptsize, minimum height=0.7cm}, band/.style = {rectangle, draw, fill=white, align=center, font=\scriptsize, minimum width=1.6cm, minimum height=0.42cm}, stack/.style = {rectangle, draw, fill=cyan!18, align=center, font=\scriptsize, minimum width=2.1cm, minimum height=0.7cm}, outp/.style = {rectangle, draw, fill=green!18, align=center, font=\scriptsize, minimum width=1.7cm, minimum height=0.6cm}, ar/.style = {-{Latex[length=2mm]}, thick} } \begin{document} % ---------- Title ---------- \begin{center} {\fontsize{14}{17}\selectfont\bfseries A Multi-Modal Deep-Learning Benchmark for Rock Glacier Mapping\\ in the Sikkim Himalaya at 10\,m Resolution} \end{center} \vspace{6pt} \begin{center} {\fontsize{12}{14}\selectfont Nasiruddin Kabir$^{1*}$} \end{center} \vspace{6pt} \begin{center} {\fontsize{11}{13}\selectfont $^{1}$Department of Civil Engineering, Indian Institute of Technology Bombay, Mumbai, India\par} \end{center} \vspace{6pt} \begin{center} {\fontsize{11}{13}\selectfont \href{mailto:nasiruddinkabir0123@gmail.com}{nasiruddinkabir0123@gmail.com}\quad(*Corresponding author)} \end{center} \vspace{10pt} % ---------- Abstract ---------- \noindent {\fontsize{12}{15}\selectfont\bfseries\itshape Abstract:}\enspace {\fontsize{11}{15}\selectfont\itshape Rock glaciers are climatically resilient stores of ground ice and key indicators of mountain permafrost, yet mapping them at scale is hampered by their spectral similarity to surrounding talus and debris. We present an end-to-end deep-learning segmentation workflow for rock glaciers in the Sikkim Himalayas from freely available 10\,m multi-modal satellite data, structured around \textbf{three contributions}. \textbf{First}, a data-collection and preparation pipeline: driven by a recent InSAR-derived rock-glacier inventory for Sikkim (626 polygons; Kumar et al., 2026), it composites Sentinel-2 optical imagery over a cloud-free post-monsoon window, adds Copernicus DEM terrain derivatives and a Sentinel-1 SAR band, and clips a multi-band feature stack for each glacier; the polygons are simultaneously rasterised into binary masks, so every glacier yields a paired feature--mask patch for supervised segmentation. Cloud-probability masking over this window removes the cloud/snow no-data that contaminates naive annual composites --- the single largest source of accuracy gain in our study. \textbf{Second}, a separability-guided band-selection study that identifies a compact, high-performing band subset and shows that \emph{terrain} morphology contributes roughly twice the discriminative power of optical reflectance. \textbf{Third}, a controlled benchmark of four architectures (U-Net, Attention U-Net, SegFormer, Swin-UNet) against a classical Random-Forest baseline under an identical protocol, with test-time augmentation, ensembling, and cross-validation. The Swin-UNet is the best single model (Dice~0.653, IoU~0.485), exceeding the Random Forest by +0.052~Dice and U-Net by +0.043~Dice; a three-model ensemble reaches Dice~0.668 and IoU~0.501, and 5-fold cross-validation confirms stability at Dice~$0.583\pm0.014$. Notably, the Random Forest matches the convolutional baselines, showing that deep learning helps here only when paired with the right inductive bias. The workflow and curated 10\,m Himalayan rock-glacier dataset provide a reproducible, transferable basis for periglacial inventory mapping.} \vspace{6pt} \noindent {\fontsize{12}{14}\selectfont\itshape Keywords:}\quad {\fontsize{12}{14}\selectfont\itshape Rock-glaciers; Segmentation; Permafrost; Himalaya; Sentinel; Swin-UNet; Multi-modal; Deep-learning} \vspace{10pt} % ---------- Body ---------- \fontsize{12}{18}\selectfont \section*{1.\quad Introduction} Rock glaciers are lobate or tongue-shaped accumulations of angular debris with interstitial ice, predominantly found in the periglacial zones of high-altitude mountain ranges. They act as long-term freshwater reservoirs and are sensitive indicators of permafrost distribution and climate change (Haeberli et al., 2006). In the Hindu Kush--Himalayan region, where glacial and periglacial landforms supply water to millions downstream, accurate mapping of rock glaciers is essential for hydrological modelling, hazard assessment, and climate-impact studies (Bolch et al., 2012). Traditional rock-glacier inventories rely on manual digitisation from optical imagery---a labour-intensive, subjective process that is difficult to scale. The availability of free 10\,m Sentinel imagery and modern deep-learning segmentation methods opens the way to automation. However, rock glaciers are difficult targets: their spectral signatures overlap with talus, debris-covered ice, and moraine, and their sizes range from small cirque lobes ($\sim$0.4\,km) to valley-filling tongues ($>$2\,km). Purely spectral classifiers therefore struggle, and the quality of the input data (seasonal snow, cloud contamination) strongly conditions achievable accuracy. This paper makes \textbf{three contributions}. \textbf{(i)}~A reproducible, cloud-free multi-modal data-collection and preparation pipeline that pairs, for each glacier, a 28-band 10\,m feature stack with a rasterised binary mask and eliminates the no-data contamination of naive composites. \textbf{(ii)}~A rigorous band-selection study, based on Cohen's~$d$ separability and cumulative retraining, that identifies an optimal 20-band subset and quantifies the contribution of each feature modality. \textbf{(iii)}~A controlled benchmark of four architectures (U-Net, Attention U-Net, SegFormer, Swin-UNet) and a classical Random-Forest baseline under an identical protocol, with honest held-out test evaluation, test-time augmentation, model ensembling, and 5-fold cross-validation. \section*{2.\quad Study Area and Reference Data} The study area encompasses the Sikkim Himalayas, India, spanning approximately 88.07\textdegree--88.95\textdegree\,E and 27.32\textdegree--28.08\textdegree\,N. The terrain is rugged and high-altitude; rock glaciers occur predominantly above 4{,}500\,m on north- and east-facing cirques and valleys. A polygon inventory of 626 active rock glaciers---produced for Sikkim by SBAS-InSAR time-series detection with Google Earth Pro optical refinement (Kumar et al., 2026)---was used as reference. After removing invalid geometries, the inventory was rasterised at 10\,m into binary masks in UTM Zone~45N (EPSG:32645). One 128$\times$128\,px patch (1.28\,km $\times$ 1.28\,km) was extracted per glacier, centred on the glacier centroid, plus a small fraction of glacier-free background patches. The resulting dataset contains 671 patches (613 glacier-bearing, 58 background), split at the patch level into 469 training, 101 validation, and 101 test patches (seed~=~42). Crucially, because all glacier masks are rasterised jointly, neighbouring glaciers within one patch are labelled correctly rather than being treated as background. \section*{3.\quad Data-Collection Pipeline} All features were generated on Google Earth Engine and downloaded directly as per-glacier NumPy patches. The key design choice is the \emph{temporal window}: instead of an all-season multi-year median (which retains cloud shadow, haze and seasonal snow, leaving $\sim$20\% of optical pixels as no-data at these altitudes), we composite only October--November scenes from 2020--2023 and mask with the Sentinel-2 cloud-probability product ($p_{\text{cloud}}<35\%$). This post-monsoon, pre-winter window yields a cloud-free, snow-minimised surface with \emph{0\%} optical no-data in the final patches. Terrain derivatives are computed from the Copernicus GLO-30 DEM at native resolution and resampled to 10\,m; a Sentinel-1 VV backscatter composite over the same window supplies a weather-independent structural cue. Figure~1 summarises the pipeline. \begin{figure}[h] \centering \resizebox{\textwidth}{!}{% \begin{tikzpicture}[ every node/.style={font=\scriptsize}, src/.style ={cylinder, shape border rotate=90, draw, fill=orange!25, align=center, minimum width=1.05cm, minimum height=0.7cm, aspect=0.28}, win/.style ={rectangle, rounded corners=2pt, draw, very thick, fill=violet!18, align=center, minimum width=2.9cm, minimum height=1.0cm}, proc/.style={trapezium, trapezium left angle=74, trapezium right angle=106, draw, fill=yellow!35, align=center, minimum height=0.85cm, text width=3.0cm, inner sep=2.5pt}, feat/.style={rectangle, rounded corners=1pt, draw, fill=cyan!22, align=center, minimum width=2.1cm, minimum height=1.0cm}, mask/.style={rectangle, rounded corners=1pt, draw, fill=red!16, align=center, minimum width=2.1cm, minimum height=1.0cm}, pair/.style={rectangle, rounded corners=1pt, draw, fill=gray!18, align=center, minimum width=2.1cm, minimum height=1.1cm}, outp/.style={rectangle, draw, fill=green!22, align=center, minimum width=1.5cm, minimum height=0.6cm}, ar/.style ={-{Latex[length=2mm]}, thick}] % ===== STEP 1: input inventory -> per-glacier window (top row) ===== \node[src] (poly) at (0, 2.1) {Glacier polygon\\inventory\\(626, shapefile)}; \node[win] (win) at (4.1, 2.1) {Per-glacier $128{\times}128$ window\\(centred on centroid;\\neighbouring polygons kept)}; % ===== STEP 2a: multi-sensor feature extraction over each window ===== \node[src] (s2) at (0, 0) {Sentinel-2\\MSI}; \node[src] (dem) at (0,-2.3) {Copernicus\\GLO-30 DEM}; \node[src] (s1) at (0,-4.6) {Sentinel-1\\GRD (VV)}; \node[proc] (f2) at (4.1, 0) {Oct--Nov cloud-prob\\composite $\rightarrow$ 10 optical\\$+$ 7 indices}; \node[proc] (fd) at (4.1,-2.3) {Terrain derivatives\\(slope, TPI, TRI, TWI,\\rough., curv., hillshade)\\$\rightarrow$ 10 terrain}; \node[proc] (fs) at (4.1,-4.6) {Oct--Nov VV\\median $\rightarrow$ 1 SAR}; \node[feat] (feat) at (7.9,-2.3) {28-band feature\\patch\\$128^2\times28$}; % ===== STEP 2b: mask from polygons in the window ===== \node[proc] (raster) at (7.9, 2.1) {Rasterise all polygons\\within window}; \node[mask] (mask) at (11.3, 2.1) {Binary mask\\patch\\$128^2\times1$}; % ===== STEP 3: pair -> split ===== \node[pair] (pair) at (11.3,-2.3) {671 paired\\(feature, mask)\\patches}; \node[outp] (train) at (14.2,-1.4) {Train (469)}; \node[outp] (val) at (14.2,-2.3) {Val (101)}; \node[outp] (test) at (14.2,-3.2) {Test (101)}; % ===== arrows ===== \draw[ar](poly)--(win); \draw[ar](win.south) to[out=-90,in=90] (feat.north); % window drives feature extent \draw[ar](win.east)--(raster.west); % window drives masking \draw[ar](s2)--(f2); \draw[ar](dem)--(fd); \draw[ar](s1)--(fs); \draw[ar](f2.east)--(feat.north west); \draw[ar](fd.east)--(feat.west); \draw[ar](fs.east)--(feat.south west); \draw[ar](raster)--(mask); \draw[ar](feat.east)--(pair.west); \draw[ar](mask.south)--(pair.north); \draw[ar](pair.east)--(train.west); \draw[ar](pair.east)--(val.west); \draw[ar](pair.east)--(test.west); % ===== dashed frame + title ===== \draw[orange!75, dashed, very thick, rounded corners] (-1.5, 3.7) rectangle (15.7, -5.9); \node[anchor=west, font=\normalsize\bfseries, red!70!black] at (-1.35, 3.3) {Data preparation}; \end{tikzpicture}} \caption{Data-collection pipeline. The glacier polygon inventory (626 polygons) drives the process: each glacier defines a $128\times128$ (1.28\,km) window centred on its centroid, within which \emph{neighbouring} polygons are retained. For every window, (i)~a cloud-free 28-band feature patch is extracted from three sensors --- Sentinel-2 (post-monsoon Oct--Nov, cloud-probability-masked; 10 optical $+$ 7 index bands), Copernicus GLO-30 DEM (10 terrain bands), and Sentinel-1 SAR (1 band) --- and (ii)~all polygons within the window are rasterised into a binary mask. The paired (feature, mask) patches ($n=671$) are split at the patch level into train, validation, and test sets.} \end{figure} The 28 bands comprise: \emph{ten optical} (B2, B3, B4, B8, B8A, B11, B12, and red-edge B5/B6/B7); \emph{seven indices} (NDVI, NDWI, NDSI, NDMI, BSI, Tasselled-Cap Brightness, Iron-Oxide ratio); \emph{ten terrain} derivatives (Elevation, Slope, Aspect, Plan/Profile Curvature, TPI, TRI, TWI, Surface Roughness, Hillshade); and \emph{one SAR} band (Sentinel-1 VV). All bands were normalised per channel using z-score statistics computed on the training patches only. \section*{4.\quad Band Selection} \subsection*{4.1\quad Separability ranking} To rank the 28 candidate bands we computed Cohen's~$d$ effect size between glacier and background pixel populations for each band across the training and validation patches: $d = |\mu_{g}-\mu_{b}| / \sigma_{\text{pooled}}$. Table~1 reports the ranking. The four highest-ranked bands are all \emph{terrain} descriptors (Surface Roughness, Slope, TPI, TRI), confirming that rock glaciers are, first and foremost, a morphological rather than a spectral target. Sentinel-1 SAR and Elevation also rank highly, while several optical bands and the curvature terms carry very little class-separating information ($d<0.05$). \begin{table}[h] \centering \caption{Band separability by Cohen's~$d$ (glacier vs.\ background), measured on patches\_v2.} \small \begin{tabular}{clc@{\hskip 1.2cm}clc} \toprule \textbf{Rank} & \textbf{Band} & \textbf{Cohen's $d$} & \textbf{Rank} & \textbf{Band} & \textbf{Cohen's $d$} \\ \midrule 1 & Surface Roughness & 0.496 & 15 & Red-Edge 3 & 0.161 \\ 2 & Slope & 0.489 & 16 & Hillshade & 0.154 \\ 3 & TPI & 0.424 & 17 & Red-Edge 2 & 0.141 \\ 4 & TRI & 0.361 & 18 & B11 (SWIR-1) & 0.123 \\ 5 & NDVI & 0.304 & 19 & Brightness (TC) & 0.081 \\ 6 & B12 (SWIR-2) & 0.275 & 20 & Red-Edge 1 & 0.080 \\ 7 & TWI & 0.261 & 21 & Aspect & 0.077 \\ 8 & S1\_VV (SAR) & 0.237 & 22 & B2 (Blue) & 0.076 \\ 9 & Elevation & 0.237 & 23 & B3 (Green) & 0.064 \\ 10 & NDWI & 0.234 & 24 & B4 (Red) & 0.051 \\ 11 & NDMI & 0.228 & 25 & NDSI & 0.031 \\ 12 & BSI & 0.212 & 26 & Iron-Oxide & 0.008 \\ 13 & B8A (Narrow-NIR) & 0.184 & 27 & Plan Curvature & 0.008 \\ 14 & B8 (NIR) & 0.167 & 28 & Profile Curvature & 0.005 \\ \bottomrule \end{tabular} \end{table} \subsection*{4.2\quad Cumulative band-count experiment} Starting from the Cohen's~$d$ ranking, we trained an Attention U-Net (identical protocol) on the top-$k$ bands for $k \in \{8,12,16,18,20,22,24,28\}$ and recorded validation Dice (Table~2). Performance rises with more bands up to $k{=}20$, then \emph{declines}: adding the four lowest-ranked bands ($k{=}24,28$) injects noise and lowers Dice. The 20-band subset is therefore adopted for all subsequent experiments. It comprises the ten terrain derivatives excluding the two curvature terms and Aspect, the SAR band, and the higher-ranked optical bands and indices. \begin{table}[h] \centering \caption{Effect of band count (Attention U-Net, validation Dice). Best at 20 bands.} \small \begin{tabular}{lcccccccc} \toprule \textbf{\# Bands} & 8 & 12 & 16 & 18 & \textbf{20} & 22 & 24 & 28 \\ \midrule \textbf{Val Dice} & 0.5288 & 0.5438 & 0.5448 & 0.5355 & \textbf{0.5700} & 0.5540 & 0.5497 & 0.5514 \\ \bottomrule \end{tabular} \end{table} \subsection*{4.3\quad Comprehensive band-configuration study} Beyond the monotone $k$-sweep, we conducted a total of \textbf{thirteen} band-configuration experiments across two rounds to ensure the 20-band choice is robust and not an artefact of a single selection strategy. Round~1 varied the number of top-ranked bands ($k = 8$ to $28$); Round~2 tested targeted hypotheses: dropping the SAR band (``Top-20~$-$~SAR''), optical-only and terrain-only subsets, a compact terrain$+$SAR set, and a hand-curated 18-band set that removes the three lowest-ranked terrain terms (Aspect, Plan/Profile Curvature). All experiments used the identical Attention U-Net training protocol; Table~3 ranks them by validation Dice. The 20-band configuration is the global optimum across every metric except recall, and no alternative selection strategy surpasses it. \begin{table}[h] \centering \caption{All thirteen band-configuration experiments (Attention U-Net, validation set), ranked by Dice. The top-20 subset is the overall optimum.} \small \begin{tabular}{clccccc} \toprule \textbf{Rank} & \textbf{Configuration} & \textbf{\# Bands} & \textbf{Dice} & \textbf{IoU} & \textbf{Prec.} & \textbf{Rec.} \\ \midrule 1 & \textbf{Top-20 (terrain+optical+SAR)} & \textbf{20} & \textbf{0.5700} & \textbf{0.4001} & 0.5029 & 0.6642 \\ 2 & Top-22 & 22 & 0.5540 & 0.3855 & 0.4605 & 0.6984 \\ 3 & All bands & 28 & 0.5514 & 0.3835 & 0.4636 & 0.6863 \\ 4 & Top-24 & 24 & 0.5497 & 0.3807 & 0.5010 & 0.6135 \\ 5 & Top-16 & 16 & 0.5448 & 0.3765 & 0.4366 & 0.7272 \\ 6 & Top-12 & 12 & 0.5438 & 0.3748 & 0.5258 & 0.5744 \\ 7 & Top-18 & 18 & 0.5355 & 0.3667 & 0.4795 & 0.6178 \\ 8 & Top-20 $-$ SAR & 19 & 0.5341 & 0.3670 & 0.4565 & 0.6583 \\ 9 & Top-8 & 8 & 0.5288 & 0.3616 & 0.4716 & 0.6096 \\ 10 & Curated-18 (no curvature/aspect) & 18 & 0.5143 & 0.3488 & 0.4854 & 0.5484 \\ 11 & Terrain + SAR & 8 & 0.5086 & 0.3466 & 0.5129 & 0.5104 \\ 12 & Terrain only & 10 & 0.4738 & 0.3113 & 0.4285 & 0.5494 \\ 13 & Optical + indices only & 12 & 0.4465 & 0.2883 & 0.4483 & 0.4468 \\ \bottomrule \end{tabular} \end{table} \subsection*{4.4\quad Feature-modality ablation} Isolating the modality-only rows of Table~3 quantifies the contribution of each data source (Table~4). Terrain-only already reaches Dice~0.474, whereas optical-only reaches only 0.447; adding optical to the full stack contributes +0.062~Dice while terrain contributes +0.124 --- roughly twice the discriminative power. The SAR band adds a further +0.036. This confirms that rock-glacier segmentation is dominated by morphological signal, with spectral and radar features providing complementary refinement. \begin{table}[h] \centering \caption{Feature-modality ablation (validation Dice) and marginal contribution to the 20-band model.} \small \begin{tabular}{lcc} \toprule \textbf{Feature subset} & \textbf{Val Dice} & \textbf{Marginal contribution} \\ \midrule Optical + indices only (12 bands) & 0.4465 & --- \\ Terrain only (10 bands) & 0.4738 & --- \\ Terrain + SAR (8 bands) & 0.5086 & --- \\ \midrule Full top-20 (terrain + optical + SAR) & \textbf{0.5700} & terrain +0.124 / optical +0.062 / SAR +0.036 \\ \bottomrule \end{tabular} \end{table} \section*{5.\quad Segmentation Models and Training} As a non-deep reference, a \textbf{Random-Forest} pixel classifier (300 trees, max depth 20, balanced class weights) was trained on the same 20 bands using a class-balanced sample of $\sim$1.2M training pixels. Four deep architectures were then compared under an identical protocol on the 20-band input: \textbf{(i)}~a plain U-Net (Ronneberger et al., 2015) with a 64--128--256--512 encoder; \textbf{(ii)}~an Attention U-Net (Oktay et al., 2018) adding soft attention gates on each skip connection; \textbf{(iii)}~SegFormer (Xie et al., 2021), a hierarchical transformer encoder (MiT-B0) with a lightweight all-MLP decoder; and \textbf{(iv)}~a Swin-UNet (Liu et al., 2021; Cao et al., 2022), a U-shaped network with shifted-window self-attention blocks that capture long-range context while retaining locality. The Swin-UNet architecture (Figure~2) follows a residual encoder--decoder: four residual-convolution encoder stages (16, 32, 64, 128 channels) progressively halve the spatial resolution to an $8\times8$ bottleneck, where two shifted-window self-attention blocks operate over the 64 spatial tokens to model long-range context; a symmetric decoder with skip connections then restores full resolution, and a $1\times1$ convolution with sigmoid activation produces the glacier probability map. \begin{figure}[h] \centering \begin{tikzpicture}[ enc/.style ={rectangle, draw, fill=blue!12, align=center, font=\scriptsize, minimum width=1.7cm, minimum height=0.62cm}, dec/.style ={rectangle, draw, fill=green!14, align=center, font=\scriptsize, minimum width=1.7cm, minimum height=0.62cm}, bn/.style ={rectangle, rounded corners, draw, fill=orange!22, align=center, font=\scriptsize, minimum width=3.0cm, minimum height=0.95cm}, io/.style ={rectangle, draw, fill=gray!14, align=center, font=\scriptsize, minimum width=1.6cm, minimum height=0.6cm}, dn/.style ={-{Latex[length=1.7mm]}, thick}, sk/.style ={-{Latex[length=1.7mm]}, dashed, gray!70}] \node[io] (in) at (0,4.5) {Input\\$128^2\times20$}; \node[enc] (e1) at (0,3.55) {Res-Conv 16\\$128^2$}; \node[enc] (e2) at (0,2.55) {Res-Conv 32\\$64^2$}; \node[enc] (e3) at (0,1.55) {Res-Conv 64\\$32^2$}; \node[enc] (e4) at (0,0.55) {Res-Conv 128\\$16^2$}; \node[bn] (bn) at (2.9,-0.55){Res-Conv 256 $+\;2\times$Swin\\$8^2$ (64 tokens)}; \node[dec] (d4) at (5.8,0.55) {Res-Conv 128\\$16^2$}; \node[dec] (d3) at (5.8,1.55) {Res-Conv 64\\$32^2$}; \node[dec] (d2) at (5.8,2.55) {Res-Conv 32\\$64^2$}; \node[dec] (d1) at (5.8,3.55) {Res-Conv 16\\$128^2$}; \node[io] (out) at (5.8,4.5) {Output\\$128^2\times1$}; \draw[dn] (in)--(e1); \draw[dn] (e1)--(e2) node[midway,left,font=\tiny]{pool $\downarrow$2}; \draw[dn] (e2)--(e3) node[midway,left,font=\tiny]{pool $\downarrow$2}; \draw[dn] (e3)--(e4) node[midway,left,font=\tiny]{pool $\downarrow$2}; \draw[dn] (e4)-|(bn); \draw[dn] (bn)|-(d4); \draw[dn] (d4)--(d3) node[midway,right,font=\tiny]{up $\uparrow$2}; \draw[dn] (d3)--(d2) node[midway,right,font=\tiny]{up $\uparrow$2}; \draw[dn] (d2)--(d1) node[midway,right,font=\tiny]{up $\uparrow$2}; \draw[dn] (d1)--(out); \draw[sk] (e1)--(d1) node[midway,above,font=\tiny,black]{skip}; \draw[sk] (e2)--(d2); \draw[sk] (e3)--(d3); \draw[sk] (e4)--(d4); \end{tikzpicture} \caption{Swin-UNet architecture. A residual convolutional encoder (blue) downsamples the 20-band $128\times128$ input to an $8\times8$ bottleneck (orange), where two shifted-window self-attention (Swin) blocks model long-range context over the 64 spatial tokens. A symmetric decoder (green) with skip connections (dashed) restores full resolution; a final $1\times1$ convolution yields the glacier probability map.} \end{figure} All models were trained with the AdamW optimiser (weight decay $10^{-4}$), batch size~32, for 250 epochs, using a cosine-annealing learning-rate schedule decaying from $10^{-4}$ to $10^{-6}$. Cosine annealing was found to outperform step-decay scheduling by allowing the model to keep improving late in training. The loss combines a label-smoothed binary cross-entropy with a Focal-Tversky term: \[ \mathcal{L} = 0.5\,\mathcal{L}_{\text{BCE}} + \left(1 - \frac{TP}{TP + \alpha\,FP + \beta\,FN}\right)^{\gamma}, \] with $\alpha{=}0.4$, $\beta{=}0.6$ (mild recall bias), and $\gamma{=}4/3$. Online augmentation comprised random flips, 90\textdegree\ rotations, and radiometric jitter. At inference, the final predictor is an \emph{ensemble} of three independently trained Swin-UNet models, each evaluated with 8-way test-time augmentation (TTA; four rotations $\times$ two flips); the resulting probability maps are averaged and thresholded (Figure~3). The operating threshold was selected by sweeping 0.30--0.70 on the validation set. All metrics are computed over valid (non-padded) pixels. \begin{figure}[h] \centering \begin{tikzpicture}[node distance=0.45cm and 0.85cm, mdl/.style={rectangle, rounded corners, draw, fill=cyan!15, align=center, font=\scriptsize, minimum width=1.9cm, minimum height=0.8cm}, op/.style ={rectangle, draw, fill=yellow!30, align=center, font=\scriptsize, minimum width=1.5cm, minimum height=0.7cm}, io/.style ={rectangle, draw, fill=green!15, align=center, font=\scriptsize, minimum width=1.7cm, minimum height=0.7cm}, a/.style ={-{Latex[length=2mm]}, thick}] \node[io] (in) {20-band\\128$\times$128\\patch}; \node[mdl, right=of in, yshift=1.1cm] (m1) {Swin-UNet\\(seed 1)}; \node[mdl, right=of in] (m2) {Swin-UNet\\(seed 2)}; \node[mdl, right=of in, yshift=-1.1cm] (m3) {Swin-UNet\\(seed 3)}; \node[op, right=1.0cm of m1] (t1) {8-way\\TTA}; \node[op, right=1.0cm of m2] (t2) {8-way\\TTA}; \node[op, right=1.0cm of m3] (t3) {8-way\\TTA}; \node[op, right=of t2] (avg) {average\\prob.}; \node[op, right=of avg] (th) {threshold\\$t^{*}$}; \node[io, right=of th] (out) {glacier\\mask}; \draw[a] (in) -- (m1); \draw[a] (in) -- (m2); \draw[a] (in) -- (m3); \draw[a] (m1) -- (t1); \draw[a] (m2) -- (t2); \draw[a] (m3) -- (t3); \draw[a] (t1) -- (avg); \draw[a] (t2) -- (avg); \draw[a] (t3) -- (avg); \draw[a] (avg) -- (th); \draw[a] (th) -- (out); \end{tikzpicture} \caption{Ensemble inference pipeline. A 20-band patch is passed through three independently trained Swin-UNet models; each applies 8-way test-time augmentation (four rotations $\times$ two flips). The resulting probability maps are averaged and thresholded at the validation-optimal $t^{*}$ to yield the final binary glacier mask.} \end{figure} \section*{6.\quad Results} \subsection*{6.1\quad Architecture comparison (held-out test)} Table~5 reports test-set performance for a classical Random-Forest baseline and all four deep architectures under identical inputs (20 bands). The Random Forest---a per-pixel classifier trained on the same features---reaches Dice~0.601, which is already \emph{competitive with the convolutional baselines} (U-Net 0.610, Attention U-Net 0.604) and in fact exceeds SegFormer (0.555). This is an important sanity check: on a small dataset, a strong classical method is not trivially beaten, and only an architecture with the right inductive bias offers a decisive gain. The Swin-UNet provides that gain, exceeding the Random Forest by +0.052~Dice and the plain U-Net by +0.043~Dice and +0.046~IoU. Notably, SegFormer---despite the strongest \emph{validation} score---generalises worst on the test set, indicating that its global efficient attention over-adapts to the training scenes; the Swin-UNet's shifted-window attention is a better inductive bias for this small-data, terrain-dominated regime. \begin{table}[h] \centering \caption{Test-set performance of a classical baseline and four deep architectures on 101 held-out patches (identical 20-band input; deep models use 8-way TTA). Best model in \textbf{bold}.} \small \begin{tabular}{lccccc} \toprule \textbf{Model} & \textbf{Dice} & \textbf{IoU} & \textbf{Precision} & \textbf{Recall} & \textbf{Thr.} \\ \midrule Random Forest (classical) & 0.6005 & 0.4291 & 0.5190 & 0.7125 & 0.60 \\ \midrule Plain U-Net & 0.6104 & 0.4392 & 0.6072 & 0.6135 & 0.30 \\ Attention U-Net & 0.6035 & 0.4322 & 0.5910 & 0.6166 & 0.40 \\ SegFormer (MiT-B0) & 0.5547 & 0.3838 & 0.5056 & 0.6145 & 0.35 \\ \textbf{Swin-UNet} & \textbf{0.6529} & \textbf{0.4847} & \textbf{0.6256} & \textbf{0.6827} & 0.35 \\ \bottomrule \end{tabular} \end{table} \subsection*{6.2\quad Ensemble and cross-validation} A three-model Swin-UNet ensemble (different random seeds), combined with 8-way TTA, improves the test Dice to 0.668 and pushes IoU past 0.50 (Table~6), with a large recall gain (+0.057) at a small cost in precision. A 5-fold cross-validation---in which each fold is validated on a genuinely held-out slice never seen during that fold's training---yields a Dice of $0.583\pm0.014$, confirming the stability of the result (standard deviation of only 0.014 across folds). \begin{table}[h] \centering \caption{Swin-UNet single model vs.\ ensemble on the held-out test set, and 5-fold cross-validation.} \small \begin{tabular}{lcccc} \toprule \textbf{Configuration} & \textbf{Dice} & \textbf{IoU} & \textbf{Precision} & \textbf{Recall} \\ \midrule Swin-UNet (single) & 0.6529 & 0.4847 & 0.6256 & 0.6827 \\ Swin-UNet (3-model ensemble + TTA)& \textbf{0.6679} & \textbf{0.5014} & 0.6087 & \textbf{0.7398} \\ \midrule 5-fold CV (mean $\pm$ std) & \multicolumn{4}{c}{Dice $= 0.583 \pm 0.014$} \\ \bottomrule \end{tabular} \end{table} We evaluated five ensembling strategies to select the final predictor (Table~7). The three-seed ensemble and the larger eight-model ``mega'' ensemble tie for the best Dice (0.668) and IoU (0.501); however, the three-seed ensemble attains this with far fewer models and the \emph{highest recall} (0.740), which is preferable for inventory completeness, where a missed glacier is costlier than slight boundary over-segmentation. The five-fold cross-validation ensemble is marginally lower, and an ensemble of models trained with copy-paste augmentation did not improve over the plain seeds. We therefore adopt the parsimonious three-seed ensemble as the final predictor. \begin{table}[h] \centering \caption{Ensemble-strategy comparison on the held-out test set (all use the Swin-UNet, 20 bands, 8-way TTA). The three-seed ensemble is selected: best Dice/IoU with the fewest models and highest recall.} \small \begin{tabular}{lccccc} \toprule \textbf{Ensemble strategy} & \textbf{\# Models} & \textbf{Dice} & \textbf{IoU} & \textbf{Precision} & \textbf{Recall} \\ \midrule Single model (no ensemble) & 1 & 0.6529 & 0.4847 & 0.6256 & 0.6827 \\ \textbf{3-seed ensemble (chosen)} & \textbf{3} & \textbf{0.6679} & \textbf{0.5014} & 0.6087 & \textbf{0.7398} \\ 5-fold cross-validation ensemble & 5 & 0.6629 & 0.4957 & 0.6278 & 0.7021 \\ 8-model ``mega'' ensemble & 8 & 0.6679 & 0.5014 & 0.6482 & 0.6889 \\ 5-model copy-paste ensemble & 5 & 0.6528 & 0.4845 & 0.6328 & 0.6741 \\ \bottomrule \end{tabular} \end{table} \subsection*{6.3\quad Training dynamics and qualitative results} Figure~4 shows the training and validation Dice/loss curves for the Swin-UNet under the cosine-annealing schedule; the model improves smoothly throughout training without abrupt plateaus. Figure~5 presents detailed qualitative results for the five best-segmented held-out test patches, comparing (left to right) the RGB composite, ground-truth mask, model prediction, predicted probability, and a colour-coded error map. Figure~6 provides a nine-patch gallery in which the predicted glacier outline (cyan) is overlaid on the ground-truth outline (yellow). All panels are georeferenced (WGS-84 longitude/latitude). The model delineates well-defined glacier bodies accurately, with most residual error confined to the gradational glacier--talus boundary. \begin{figure}[h] \centering \includegraphics[width=\textwidth]{paper_figures/training_history.png} \caption{Training dynamics of the Swin-UNet with cosine-annealing learning rate over 250 epochs: (a) Dice and (b) loss for the training and validation sets.} \end{figure} \begin{figure}[h] \centering \includegraphics[width=0.98\textwidth]{paper_figures/top5_detailed.png} \caption{Detailed qualitative results on the five best-segmented held-out test patches. Columns: RGB composite (with 500\,m scale bar and north arrow), ground truth, predicted mask, predicted probability, and error map (green = true positive, red = false positive, blue = false negative). Axes are WGS-84 longitude/latitude.} \end{figure} \begin{figure}[h] \centering \includegraphics[width=0.82\textwidth]{paper_figures/montage_best.png} \caption{Gallery of representative test patches. The Swin-UNet prediction (cyan) is overlaid on the ground-truth glacier outline (yellow) over the RGB composite; per-patch Dice is annotated. Axes are WGS-84 longitude/latitude.} \end{figure} \section*{7.\quad Discussion} Three findings stand out. First, \emph{data quality dominates}. Replacing a naive all-season composite (with $\sim$20\% cloud/snow no-data) by an October--November cloud-probability composite (0\% optical no-data) was the single largest source of improvement in this study, underlining that in high-mountain segmentation the compositing strategy matters as much as the model. Second, \emph{rock glaciers are a terrain problem}: the Cohen's~$d$ ranking and modality ablation both place terrain derivatives---surface roughness, slope, topographic position---above all optical features, with terrain contributing roughly twice the discriminative power of reflectance. Third, \emph{architecture matters, but with the right inductive bias}: a classical Random Forest already matches the convolutional baselines (Dice~0.601), so the value of deep learning here is not automatic---only the Swin-UNet's shifted-window attention yields a decisive margin (+0.052~Dice over the Random Forest, +0.043 over U-Net), while a global-attention transformer (SegFormer) actually underperforms the classical baseline. The Swin-UNet's advantage is realised without catastrophic overfitting (5-fold CV std of 0.014). The main limitation is dataset size: with one patch per glacier (671 total), the achievable ceiling is bounded by label quantity, and standard augmentations (CutOut, copy-paste, MixUp) did not yield consistent gains beyond the ensemble in our experiments. The moderate precision (0.63) reflects the gradual glacier--talus boundary, where even expert labels are uncertain at 10\,m. Future work should therefore (i)~expand the dataset with multiple offset patches per glacier and additional Himalayan sub-ranges; (ii)~explore self-training on unlabelled terrain; and (iii)~incorporate multi-temporal SAR interferometric coherence to exploit the diagnostic surface creep of active rock glaciers. \section*{8.\quad Conclusion} We presented an end-to-end deep-learning framework for rock-glacier segmentation in the Sikkim Himalayas from free 10\,m multi-modal satellite data. A cloud-free October--November collection pipeline, a separability-driven 20-band selection, and a Swin-UNet architecture together achieve a held-out test Dice of 0.653 (single model) and 0.668 (3-model ensemble with TTA), with a 5-fold cross-validated Dice of $0.583\pm0.014$. The Swin-UNet outperforms U-Net, Attention U-Net, and SegFormer under identical training. Beyond the model, the study contributes a reproducible pipeline, a curated Himalayan rock-glacier dataset, and clear evidence that terrain morphology is the primary discriminator for automated rock-glacier mapping. The approach is directly transferable to other periglacial regions of the Himalaya. \section*{References} \noindent Bolch, T., Kulkarni, A., K\"a\"ab, A., Huggel, C., Paul, F., Cogley, J.~G., \ldots\ \& Bajracharya, S. (2012). The state and fate of Himalayan glaciers. \textit{Science, 336}(6079), 310--314. \vspace{4pt}\noindent Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., \& Wang, M. (2022). Swin-Unet: Unet-like pure transformer for medical image segmentation. \textit{ECCV 2022 Workshops}, 205--218. \vspace{4pt}\noindent Haeberli, W., Hallet, B., Arenson, L., Elconin, R., Humlum, O., K\"a\"ab, A., \ldots\ \& Vonder M\"uhll, D. (2006). Permafrost creep and rock glacier dynamics. \textit{Permafrost and Periglacial Processes, 17}(3), 189--214. \vspace{4pt}\noindent Kumar, A., Malviya, A., \& Singh, G. (2026). Multi-sensor-based rock glacier detection over Sikkim Himalaya. \textit{iScience, 29}, 115983. \url{https://doi.org/10.1016/j.isci.2026.115983} \vspace{4pt}\noindent Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., \ldots\ \& Guo, B. (2021). Swin Transformer: Hierarchical vision transformer using shifted windows. \textit{ICCV 2021}, 10012--10022. \vspace{4pt}\noindent Oktay, O., Schlemper, J., Folgoc, L.~L., Lee, M., Heinrich, M., Misawa, K., \ldots\ \& Rueckert, D. (2018). Attention U-Net: Learning where to look for the pancreas. \textit{MIDL 2018}. \vspace{4pt}\noindent Robson, B.~A., Bolch, T., MacDonell, S., H\"olbling, D., Rastner, P., \& Schaffer, N. (2020). Automated detection of rock glaciers using deep learning and object-based image analysis. \textit{Remote Sensing of Environment, 250}, 112033. \vspace{4pt}\noindent Ronneberger, O., Fischer, P., \& Brox, T. (2015). U-Net: Convolutional networks for biomedical image segmentation. \textit{MICCAI 2015}, 234--241. \vspace{4pt}\noindent Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.~M., \& Luo, P. (2021). SegFormer: Simple and efficient design for semantic segmentation with transformers. \textit{NeurIPS 2021}, 12077--12090. \end{document}