如何将XGBoost回归的残差导出为单个栅格？

修复残差计算错误 #

"未定义sb对象"这类报错，本质是引用了未初始化的变量（比如模型对象名、数据框名拼写错误）。先确保模型训练流程正确，再按以下步骤计算残差：

1. 从数据框分离自变量与因变量：

X <- df[, c("pop", "tirs", "agbh")]
y_true <- df$ntl

2. 用训练好的模型生成预测值并计算残差：

y_pred <- predict(xgb_model, newdata = X)
residuals <- y_true - y_pred
# 将残差绑定回原数据框
df$residuals <- residuals

将残差导出为栅格 #

要生成带空间参考的残差栅格，必须保留原始栅格的像元索引或空间坐标，以下是完整流程：

前提：确保数据框包含像元索引 #

如果你的数据框是从栅格提取而来，提取时需保留像元ID：

# 加载原始栅格堆栈
raster_stack <- stack("pop.tif", "tirs.tif", "agbh.tif", "ntl.tif")
# 提取像元值及对应ID
df <- as.data.frame(extract(raster_stack, 1:ncell(raster_stack), cellnumbers = TRUE))
# 重命名列匹配你的数据集
colnames(df) <- c("cell_id", "pop", "tirs", "agbh", "ntl")

创建并导出残差栅格 #

1. 选取任意原始栅格作为模板（保证空间参数一致）：

template_raster <- raster_stack$pop

2. 初始化残差栅格并填充值：

residual_raster <- template_raster
# 将残差赋值到对应像元位置
residual_raster[df$cell_id] <- df$residuals

3. 导出为GeoTIFF格式：

writeRaster(residual_raster, filename = "xgb_residuals.tif", format = "GTiff", overwrite = TRUE)

完整可运行示例 #

library(caret)
library(xgboost)
library(raster)

# 加载数据（替换为你的数据路径）
df <- read.csv("your_dataset.csv")
# 加载原始栅格堆栈（用于生成栅格模板）
raster_stack <- stack("pop.tif", "tirs.tif", "agbh.tif", "ntl.tif")

# 用最优参数训练XGBoost回归模型
train_control <- trainControl(method = "cv", number = 5)
xgb_optimal_grid <- expand.grid(
  nrounds = 150,  # 替换为你的最优参数
  max_depth = 5,
  eta = 0.08,
  gamma = 0.1,
  colsample_bytree = 0.7,
  min_child_weight = 2,
  subsample = 0.9
)

xgb_model <- train(
  x = df[, c("pop", "tirs", "agbh")],
  y = df$ntl,
  method = "xgbTree",
  trControl = train_control,
  tuneGrid = xgb_optimal_grid,
  metric = "RMSE"
)

# 计算残差
df$y_pred <- predict(xgb_model, newdata = df[, c("pop", "tirs", "agbh")])
df$residuals <- df$ntl - df$y_pred

# 生成并导出残差栅格
template <- raster_stack$pop
residual_raster <- template
residual_raster[df$cell_id] <- df$residuals

writeRaster(residual_raster, "xgb_residuals.tif", format = "GTiff", overwrite = TRUE)

内容的提问来源于stack exchange，提问作者Nikos