Brick Kilns Across India: A Temporal Analysis of Brick Kiln Distribution and Technology Types

Overview The MuTBriK dataset contains multispectral Sentinel-2 image patches prepared for multi-class semantic segmentation of brick kiln types. Each image consists of 8 channels: six original Sentinel-2 spectral bands and two derived indices (NDVI and NDBI). All bands are normalized. Each image has a corresponding pixel-wise annotation mask. The dataset has been used to generate a national-scale map of brick kiln distribution and technology types across India. Dataset Contents Input images: 8-band multispectral GeoTIFF patches Label images: Pixel-wise segmentation masks One-to-one correspondence between each image and its label Image sizes: Training images: 263 × 263 pixels Validation/Test images: 256 × 256 pixels Spectral Bands Each image contains the following channels: Band Description Resolution -------- -------------------------------- -------------------------- B2 Blue 10 m B3 Green 10 m B4 Red 10 m B8 Near Infrared (NIR) 10 m B11 Shortwave Infrared 1 (SWIR1) 20 m → resampled to 10 m B12 Shortwave Infrared 2 (SWIR2) 20 m → resampled to 10 m Band 7 NDVI = (B8 − B4) / (B8 + B4) Derived Band 8 NDBI = (B11 − B8) / (B11 + B8) Derived Normalization All bands are normalized. Sentinel-2 reflectance values were scaled by dividing by 10,000. Classes and Labels Each label mask contains six classes: Class ID Class name ---------- ------------ 0 Background 1 FCBTK 2 Zigzag 3 CBTK 4 RBTK 5 DDK Labels are encoded as integer pixel values in the mask files. Preprocessing Steps - Sentinel-2 band selection (B2, B3, B4, B8, B11, B12) - Resampling of 20 m bands (B11, B12) to 10 m - Normalization (÷ 10,000) - Computation of NDVI and NDBI - Band stacking into 8-channel images - Patch extraction from larger Sentinel-2 tiles File Format Images: GeoTIFF (.tif) Labels: GeoTIFF (.tif) Fully compatible with open-source tools. Software Compatibility The dataset can be used with open-source software such as QGIS, Python (rasterio, NumPy, PyTorch, etc.) Geographic Coverage India Inference Results The full maps of India, showing all detected brick kilns and their corresponding technology classification, are available in the Maps_GeoJSON.zip file (2019-2025). Related Code Repository https://github.com/ssugandh01/MuTBriK# 数据集概述 MuTBriK数据集专为砖窑类型多分类语义分割任务构建，包含多光谱哨兵二号（Sentinel-2）图像块。每张图像包含8个通道：6个哨兵二号原始光谱波段，以及2个衍生指数：归一化差分植被指数（NDVI）与归一化差分建筑指数（NDBI）。所有波段均已完成归一化处理，每张图像均配有对应的逐像素标注掩码。 该数据集已用于生成印度全国范围的砖窑分布与技术类型地图。 # 数据集内容 输入图像：8波段多光谱GeoTIFF图像块 标签图像：逐像素分割掩码 图像与标签一一对应 # 图像尺寸 训练集图像：263 × 263 像素 验证集/测试集图像：256 × 256 像素 # 光谱波段 每张图像包含如下通道： 波段编号 波段描述 空间分辨率 -------- ------------------------------ -------------------------- B2 蓝光波段 10米 B3 绿光波段 10米 B4 红光波段 10米 B8 近红外波段（NIR） 10米 B11 短波红外1波段（SWIR1） 20米 → 重采样至10米 B12 短波红外2波段（SWIR2） 20米 → 重采样至10米 波段7 归一化差分植被指数（NDVI）=(B8−B4)/(B8+B4) 衍生波段 波段8 归一化差分建筑指数（NDBI）=(B11−B8)/(B11+B8) 衍生波段 # 归一化处理 所有波段均已完成归一化。哨兵二号的反射率数值通过除以10000进行缩放。 # 类别与标签 每个标签掩码包含6个类别： 类别ID 类别名称 -------- ---------- 0 背景 1 FCBTK 2 Zigzag 3 CBTK 4 RBTK 5 DDK 标签在掩码文件中以整数像素值进行编码。 # 预处理步骤 - 哨兵二号波段筛选（选取B2、B3、B4、B8、B11、B12） - 将20米分辨率波段（B11、B12）重采样至10米 - 归一化处理（除以10000） - 计算归一化差分植被指数（NDVI）与归一化差分建筑指数（NDBI） - 波段堆叠为8通道图像 - 从更大尺寸的哨兵二号影像瓦片提取图像块 # 文件格式 图像：GeoTIFF（.tif）格式 标签：GeoTIFF（.tif）格式 本数据集完全兼容各类开源工具。 # 软件兼容性 该数据集可与QGIS、Python（含rasterio、NumPy、PyTorch等库）等开源软件配合使用。 # 地理覆盖范围 印度 # 推理结果 包含所有检测到的砖窑及其技术分类的印度全境地图，可于Maps_GeoJSON.zip（2019-2025）文件中获取。 # 相关代码仓库 https://github.com/ssugandh01/MuTBriK