refactor(pipeline): 路径直接传输 — 统一 ctx 字段名/panel key/step 形参名

src/core/pipeline/context.py

@@ -20,23 +20,28 @@ class PipelineContext:
     """流水线运行上下文（在 14 个 step 之间传递的内存字典）
 
     字段命名约定：
-      - 主路径字段统一 `_path` 后缀（如 water_mask_path）
-      - 目录类字段无 `_path` 后缀（如 models_dir）
+      - 路径类字段名 = panel key 名 = step 形参名（全链路无翻译）
+      - 训练/产物 CSV 用 `_path` 后缀（如 training_csv_path / water_mask_path）
+      - 入参影像/CSV 沿用 panel 原名（img_path / csv_path），无 `_path` 后缀
+      - 目录类字段无 `_path` 后缀（如 models_dir / prediction_dir）
       - 元信息字段无后缀（如 user_config / status / log）
     """
 
-    # ── 9 步主路径（按 step 输出顺序排列） ──
-    raw_img_path: Optional[str] = None            # Step 1 入参：原始影像
+    # ── 11 个 step 的入参/产物（按 step 顺序排列；字段名 = panel key = step 形参） ──
+    img_path: Optional[str] = None                # Step 1/2/3 入参：原始影像
     water_mask_path: Optional[str] = None         # Step 1 出 → Step 2/3/7 入
     glint_mask_path: Optional[str] = None         # Step 2 出 → Step 3/7 入
     deglint_img_path: Optional[str] = None        # Step 3 出 → Step 5/7 入
-    raw_csv_path: Optional[str] = None            # Step 4 入：原始 CSV
+    csv_path: Optional[str] = None                # Step 4/5/6_5/6_75 入参：原始/训练 CSV
     processed_csv_path: Optional[str] = None      # Step 4 出 → Step 5 入
-    training_spectra_path: Optional[str] = None   # Step 5 出 → Step 6 入
+    training_csv_path: Optional[str] = None       # Step 5 出 → Step 5_5/6/6_5/6_75 入
+    boundary_path: Optional[str] = None           # Step 5 入参：边界 SHP（panel step5 名）
     indices_path: Optional[str] = None            # Step 5.5 出
-    sampling_csv_path: Optional[str] = None       # Step 7 出 → Step 8/9 入
-    prediction_csv_path: Optional[str] = None     # Step 8 出
+    sampling_csv_path: Optional[str] = None       # Step 7 出 → Step 8/8_5/8_75/9 入
+    prediction_csv_path: Optional[str] = None     # Step 8 出 → Step 9 入
     distribution_map_path: Optional[str] = None   # Step 9 出
+    boundary_shp_path: Optional[str] = None       # Step 9 入参：边界 SHP（panel step9 名）
+    formula_csv_path: Optional[str] = None        # Step 8_75 入参：公式 CSV
 
     # ── 目录类（命名不带 _path 以示区别） ──
     models_dir: Optional[str] = None

src/core/pipeline/runner.py

@@ -4,10 +4,8 @@ PipelineRunner：基于 StepSpec 声明式调度 14 个 step。
 
 设计要点：
   - StepSpec 声明 requires（ctx 字段名列表）+ produces（ctx 字段名列表）
-  - 默认约定：ctx 字段名去掉 `_path` 后缀 = step 方法形参名
-    例：ctx.water_mask_path → 形参 water_mask
-    例：ctx.raw_img_path → 形参 raw_img
-  - 可被 spec.parameter_map 覆盖
+  - 命名约定：ctx 字段名 == panel key 名 == step 形参名（全链路无翻译）
+  - 保留 spec.parameter_map 字段骨架供极少数特例覆盖（默认空 dict）
   - 调度顺序：按 PIPELINE_STEPS 列表顺序，requires 缺则 skip
   - 软取消：在每个 step 前检查 ctx.is_cancelled()
   - duck-typed pipeline：runner 只调 getattr(pipeline, method_name)，不强依赖类层级
@@ -48,101 +46,76 @@ class StepSpec:
 PIPELINE_STEPS: List[StepSpec] = [
     StepSpec(
         step_id="step1", method_name="step1_generate_water_mask",
-        requires=["raw_img_path"], produces=["water_mask_path"],
-        # ctx.raw_img_path → 形参 img_path（老 step1 形参名是 img_path，不是 raw_img）
-        parameter_map={"raw_img_path": "img_path"},
+        requires=["img_path"], produces=["water_mask_path"],
         description="水域掩膜生成（NDWI 或 SHP）",
     ),
     StepSpec(
         step_id="step2", method_name="step2_find_glint_area",
-        requires=["raw_img_path", "water_mask_path"], produces=["glint_mask_path"],
-        # raw_img_path→img_path；water_mask_path 不变
-        parameter_map={"raw_img_path": "img_path"},
+        requires=["img_path", "water_mask_path"], produces=["glint_mask_path"],
         description="耀斑区域检测",
     ),
     StepSpec(
         step_id="step3", method_name="step3_remove_glint",
-        requires=["deglint_img_path"], produces=["deglint_img_path"],
-        # deglint_img_path→img_path（老 step3 形参名是 img_path）
-        # 注意：glint_mask_path 不在 requires 中——step3 形参表无该参数，内部走 self.glint_mask_path 回退
-        parameter_map={"deglint_img_path": "img_path"},
+        requires=["img_path", "water_mask_path", "glint_mask_path"],
+        produces=["deglint_img_path"],
         description="耀斑去除",
     ),
     StepSpec(
         step_id="step4", method_name="step4_process_csv",
-        requires=["raw_csv_path"], produces=["processed_csv_path"],
-        # raw_csv_path→csv_path（老 step4 形参名是 csv_path）
-        parameter_map={"raw_csv_path": "csv_path"},
+        requires=["csv_path"], produces=["processed_csv_path"],
         description="CSV 异常值清洗",
     ),
     StepSpec(
         step_id="step5", method_name="step5_extract_training_spectra",
-        requires=["deglint_img_path", "processed_csv_path"], produces=["training_spectra_path"],
-        # processed_csv_path→csv_path（老 step5 形参名是 csv_path）；deglint_img_path 不变
-        parameter_map={"processed_csv_path": "csv_path"},
+        requires=["deglint_img_path", "csv_path", "boundary_path", "glint_mask_path"],
+        produces=["training_csv_path"],
         description="实测样本点光谱提取",
     ),
     StepSpec(
         step_id="step5_5", method_name="step5_5_calculate_water_quality_indices",
-        requires=["training_spectra_path"], produces=["indices_path"],
-        # 老 step5.5 形参是 training_spectra_path；ctx 字段同名，无需映射
-        parameter_map={},
+        requires=["training_csv_path"], produces=["indices_path"],
         description="水质光谱指数计算（optional）",
     ),
     StepSpec(
         step_id="step6", method_name="step6_train_models",
-        requires=["training_spectra_path"], produces=["models_dir"],
-        # training_spectra_path→training_csv_path（老 step6 形参名是 training_csv_path）
-        parameter_map={"training_spectra_path": "training_csv_path"},
+        requires=["training_csv_path"], produces=["models_dir"],
         description="ML 建模（GridSearchCV / AutoML）",
     ),
     StepSpec(
         step_id="step6_5", method_name="step6_5_non_empirical_modeling",
-        requires=["training_spectra_path"], produces=["models_dir"],
-        # training_spectra_path→csv_path（老 step6.5 形参名是 csv_path）
-        parameter_map={"training_spectra_path": "csv_path"},
+        requires=["training_csv_path"], produces=["models_dir"],
         description="非经验统计回归",
     ),
     StepSpec(
         step_id="step6_75", method_name="step6_75_custom_regression",
-        requires=["training_spectra_path"], produces=["models_dir"],
-        # training_spectra_path→csv_path（老 step6.75 形参名是 csv_path）
-        parameter_map={"training_spectra_path": "csv_path"},
+        requires=["training_csv_path"], produces=["models_dir"],
         description="自定义回归分析",
     ),
     StepSpec(
         step_id="step7", method_name="step7_generate_sampling_points",
         requires=["deglint_img_path", "water_mask_path"], produces=["sampling_csv_path"],
-        # 老 step7 形参是 deglint_img_path / water_mask_path；ctx 字段同名
-        parameter_map={},
         description="整景密集采样点生成 + 光谱提取",
     ),
     StepSpec(
         step_id="step8", method_name="step8_predict_water_quality",
         requires=["sampling_csv_path", "models_dir"], produces=["prediction_csv_path"],
-        parameter_map={},
         description="ML 模型预测（采样点）",
     ),
     StepSpec(
         step_id="step8_5", method_name="step8_5_predict_with_non_empirical_models",
-        requires=["sampling_csv_path"], produces=["prediction_dir"],
-        parameter_map={},
+        requires=["sampling_csv_path", "models_dir"], produces=["prediction_dir"],
         description="非经验模型预测",
     ),
     StepSpec(
         step_id="step8_75", method_name="step8_75_predict_with_custom_regression",
-        requires=["sampling_csv_path"], produces=["prediction_dir"],
-        parameter_map={},
+        requires=["sampling_csv_path", "models_dir", "formula_csv_path"],
+        produces=["prediction_dir"],
         description="自定义回归预测",
     ),
     StepSpec(
         step_id="step9", method_name="step9_generate_distribution_map",
-        requires=["prediction_csv_path"],
+        requires=["prediction_csv_path", "boundary_shp_path"],
         produces=["distribution_map_path"],
-        # 老 step9 形参是 prediction_csv_path / boundary_shp_path；ctx 字段同名
-        # 注意：sampling_csv_path / water_mask_path 不在 requires 中——step9 形参表无该参数，
-        # 内部走 self.sampling_csv_path / self.water_mask_path 回退
-        parameter_map={},
         description="克里金插值成图",
     ),
 ]
@@ -157,7 +130,7 @@ class PipelineRunner:
 
     用法：
         runner = PipelineRunner(pipeline_instance)
-        ctx = PipelineContext(raw_img_path=..., ...)
+        ctx = PipelineContext(img_path=..., ...)
         result_ctx = runner.run(ctx)
     """
 

src/core/prediction/automl_trainer.py

@@ -0,0 +1,544 @@
+# -*- coding: utf-8 -*-
+"""
+Optuna + 智能子采样 AutoML 训练器（路线 B 防爆引擎）。
+
+为什么需要这个：
+  - 老路径：11 预处理 × 4 模型 × 3 划分 = 132 组 GridSearchCV
+    对中小数据集 10 分钟+，对大数据集 5w+ 行 直接 OOM
+  - AutoML 路径：1 预处理 × N 模型（Optuna 调超参），用智能子采样避开 OOM
+    再用最优超参在**全量数据**上 refit，最终保存单一模型
+
+设计要点：
+  - 入口 train_with_automl(csv, feature_start_column, model_names, ...)
+  - AutoMLResult dataclass 返回（每个目标列一份）
+  - smart_subsample：N > max_samples 时随机下采样
+  - 失败兜底：optuna 未装 / 全 trial 失败 → fallback 到 WaterQualityModelingBatch
+  - 文件命名规范：{target}_{preprocess}_{model}_AUTOML.joblib
+  - save_data["metadata"]["automl"] = True 标记
+
+调用：
+    from src.core.prediction.automl_trainer import train_with_automl
+    results = train_with_automl(
+        training_csv_path=".../training_spectra.csv",
+        feature_start_column="374.285004",
+        model_names=["RF", "SVR", "Ridge"],
+        n_trials=20,
+        timeout_sec=300,
+    )
+"""
+
+from __future__ import annotations
+
+import json
+import time
+from dataclasses import asdict, dataclass, field
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Optional, Tuple
+
+import numpy as np
+import pandas as pd
+
+
+# ============================================================
+# 常量
+# ============================================================
+
+# AutoML 寻优阶段允许的最大样本数（避免 OOM）
+# 5000 样本对 RF/SVR/Ridge 的 Optuna 寻优足够给出稳定 CV
+DEFAULT_MAX_SAMPLES = 5000
+
+# 单次 Optuna trial 的默认超时（秒）
+DEFAULT_TIMEOUT = 300.0
+
+# 默认 trial 数
+DEFAULT_N_TRIALS = 20
+
+# AutoML 输出目录名后缀
+AUTOML_DIR_SUFFIX = "_AutoML"
+
+
+# ============================================================
+# 数据类
+# ============================================================
+
+@dataclass
+class AutoMLResult:
+    """单个目标列的 AutoML 训练结果"""
+    success: bool = False
+    model_path: Optional[str] = None
+    cv_score: float = -float("inf")
+    best_params: Optional[Dict[str, Any]] = None
+    target_column: str = ""
+    preprocessing: str = ""
+    model_name: str = ""
+    n_trials_done: int = 0
+    n_samples_used: int = 0
+    fallback_used: bool = False
+    elapsed_sec: float = 0.0
+    error: Optional[str] = None
+    metadata: Dict[str, Any] = field(default_factory=dict)
+
+
+# ============================================================
+# 智能子采样
+# ============================================================
+
+def smart_subsample(
+    X: np.ndarray,
+    y: np.ndarray,
+    max_samples: int = DEFAULT_MAX_SAMPLES,
+    random_state: int = 42,
+) -> Tuple[np.ndarray, np.ndarray, bool]:
+    """当 N > max_samples 时随机下采样；否则原样返回。
+
+    Returns:
+        (X_sub, y_sub, was_subsampled)
+    """
+    n = X.shape[0]
+    if n <= max_samples:
+        return X, y, False
+    rng = np.random.default_rng(random_state)
+    idx = rng.choice(n, size=max_samples, replace=False)
+    return X[idx], y[idx], True
+
+
+# ============================================================
+# 模型工厂
+# ============================================================
+
+def _build_model(model_name: str, random_state: int = 42):
+    """根据英文模型键名构造 sklearn-compatible 模型实例（factory）。"""
+    from sklearn.ensemble import (
+        AdaBoostRegressor, ExtraTreesRegressor, GradientBoostingRegressor,
+        RandomForestRegressor,
+    )
+    from sklearn.linear_model import (
+        ElasticNet, Lasso, LinearRegression, Ridge,
+    )
+    from sklearn.neighbors import KNeighborsRegressor
+    from sklearn.neural_network import MLPRegressor
+    from sklearn.svm import SVR
+    from sklearn.tree import DecisionTreeRegressor
+
+    factory = {
+        "RF": lambda **kw: RandomForestRegressor(random_state=random_state, n_jobs=1, **kw),
+        "ET": lambda **kw: ExtraTreesRegressor(random_state=random_state, n_jobs=1, **kw),
+        "GradientBoosting": lambda **kw: GradientBoostingRegressor(random_state=random_state, **kw),
+        "AdaBoost": lambda **kw: AdaBoostRegressor(random_state=random_state, **kw),
+        "Ridge": lambda **kw: Ridge(**kw),
+        "Lasso": lambda **kw: Lasso(max_iter=5000, **kw),
+        "ElasticNet": lambda **kw: ElasticNet(max_iter=5000, **kw),
+        "LinearRegression": lambda **kw: LinearRegression(**kw),
+        "SVR": lambda **kw: SVR(**kw),
+        "KNN": lambda **kw: KNeighborsRegressor(n_jobs=1, **kw),
+        "MLP": lambda **kw: MLPRegressor(max_iter=500, random_state=random_state, **kw),
+        "DecisionTree": lambda **kw: DecisionTreeRegressor(random_state=random_state, **kw),
+        "PLS": None,   # sklearn.cross_decomposition.PLSRegression 暂未集成
+    }
+    builder = factory.get(model_name)
+    if builder is None:
+        return None
+    return builder
+
+
+# ============================================================
+# Optuna 超参 search space
+# ============================================================
+
+def _get_search_space(model_name: str, trial) -> Dict[str, Any]:
+    """按模型名返回 Optuna 超参 search space。"""
+    sp: Dict[str, Any] = {}
+    if model_name == "RF":
+        sp["n_estimators"] = trial.suggest_int("n_estimators", 50, 300, step=50)
+        sp["max_depth"] = trial.suggest_int("max_depth", 3, 20)
+        sp["min_samples_split"] = trial.suggest_int("min_samples_split", 2, 10)
+        sp["min_samples_leaf"] = trial.suggest_int("min_samples_leaf", 1, 5)
+    elif model_name == "ET":
+        sp["n_estimators"] = trial.suggest_int("n_estimators", 50, 300, step=50)
+        sp["max_depth"] = trial.suggest_int("max_depth", 3, 20)
+    elif model_name == "GradientBoosting":
+        sp["n_estimators"] = trial.suggest_int("n_estimators", 50, 300, step=50)
+        sp["max_depth"] = trial.suggest_int("max_depth", 3, 8)
+        sp["learning_rate"] = trial.suggest_float("learning_rate", 0.01, 0.3, log=True)
+    elif model_name == "SVR":
+        sp["C"] = trial.suggest_float("C", 0.1, 100.0, log=True)
+        sp["epsilon"] = trial.suggest_float("epsilon", 0.001, 1.0, log=True)
+        sp["kernel"] = trial.suggest_categorical("kernel", ["rbf", "linear"])
+    elif model_name == "KNN":
+        sp["n_neighbors"] = trial.suggest_int("n_neighbors", 3, 20)
+        sp["weights"] = trial.suggest_categorical("weights", ["uniform", "distance"])
+    elif model_name in ("Ridge", "Lasso", "ElasticNet"):
+        sp["alpha"] = trial.suggest_float("alpha", 0.01, 100.0, log=True)
+        if model_name == "ElasticNet":
+            sp["l1_ratio"] = trial.suggest_float("l1_ratio", 0.0, 1.0)
+    elif model_name == "MLP":
+        sp["hidden_layer_sizes"] = trial.suggest_categorical(
+            "hidden_layer_sizes", [(50,), (100,), (50, 50), (100, 50)]
+        )
+        sp["alpha"] = trial.suggest_float("alpha", 1e-5, 1e-1, log=True)
+        sp["learning_rate_init"] = trial.suggest_float("learning_rate_init", 1e-4, 1e-2, log=True)
+    elif model_name == "DecisionTree":
+        sp["max_depth"] = trial.suggest_int("max_depth", 3, 20)
+        sp["min_samples_split"] = trial.suggest_int("min_samples_split", 2, 10)
+    elif model_name == "AdaBoost":
+        sp["n_estimators"] = trial.suggest_int("n_estimators", 30, 200, step=30)
+        sp["learning_rate"] = trial.suggest_float("learning_rate", 0.01, 1.0, log=True)
+    else:
+        sp["n_estimators"] = trial.suggest_int("n_estimators", 50, 200, step=50)
+    return sp
+
+
+def _make_objective(model_name: str, X: np.ndarray, y: np.ndarray,
+                    cv_folds: int, random_state: int):
+    """构造 Optuna objective（5 折 CV R²）。"""
+    from sklearn.model_selection import KFold, cross_val_score
+
+    def objective(trial):
+        params = _get_search_space(model_name, trial)
+        try:
+            builder = _build_model(model_name, random_state=random_state)
+            if builder is None:
+                return -1.0
+            model = builder(**params)
+            kf = KFold(n_splits=cv_folds, shuffle=True, random_state=random_state)
+            scores = cross_val_score(model, X, y, cv=kf, scoring="r2", n_jobs=1)
+            return float(np.mean(scores))
+        except Exception:
+            return -1.0
+
+    return objective
+
+
+def _refit_full(model_name: str, best_params: Dict[str, Any],
+                X: np.ndarray, y: np.ndarray, random_state: int):
+    """用 best params 在**全量数据**上 refit。"""
+    builder = _build_model(model_name, random_state=random_state)
+    if builder is None:
+        return None
+    model = builder(**best_params)
+    model.fit(X, y)
+    return model
+
+
+# ============================================================
+# 失败兜底（回退到老 GridSearchCV 路径）
+# ============================================================
+
+def _fallback_train(
+    training_csv_path: str,
+    feature_start_column,
+    preprocessing: str,
+    model_name: str,
+    split_method: str,
+    cv_folds: int,
+    output_dir: Path,
+    target_column: str,
+) -> AutoMLResult:
+    """AutoML 失败时调老 WaterQualityModelingBatch。
+
+    返回的 AutoMLResult.fallback_used=True。
+    """
+    try:
+        from src.core.modeling.modeling_batch import WaterQualityModelingBatch
+    except ImportError as e:
+        return AutoMLResult(
+            success=False, error=f"fallback 导入失败: {e!r}", fallback_used=True,
+            target_column=target_column, preprocessing=preprocessing, model_name=model_name,
+        )
+
+    try:
+        out_dir = output_dir / preprocessing
+        out_dir.mkdir(parents=True, exist_ok=True)
+        modeler = WaterQualityModelingBatch(str(out_dir))
+        modeler.train_models_batch(
+            csv_path=training_csv_path,
+            feature_start_column=feature_start_column,
+            preprocessing_methods=[preprocessing],
+            model_names=[model_name],
+            split_methods=[split_method],
+            cv_folds=cv_folds,
+        )
+        # 找产出
+        candidates = list(out_dir.rglob(f"{target_column}_{preprocessing}_{model_name}.joblib"))
+        model_path = str(candidates[0]) if candidates else None
+        return AutoMLResult(
+            success=model_path is not None,
+            model_path=model_path,
+            target_column=target_column, preprocessing=preprocessing, model_name=model_name,
+            fallback_used=True,
+            metadata={"source": "WaterQualityModelingBatch"},
+        )
+    except Exception as e:
+        return AutoMLResult(
+            success=False, error=f"fallback 失败: {e!r}", fallback_used=True,
+            target_column=target_column, preprocessing=preprocessing, model_name=model_name,
+        )
+
+
+# ============================================================
+# 主入口
+# ============================================================
+
+def train_with_automl(
+    training_csv_path: str,
+    feature_start_column,
+    preprocessing_methods: Optional[List[str]] = None,
+    model_names: Optional[List[str]] = None,
+    split_methods: Optional[List[str]] = None,
+    cv_folds: int = 5,
+    output_dir: Optional[str] = None,
+    n_trials: int = DEFAULT_N_TRIALS,
+    timeout_sec: float = DEFAULT_TIMEOUT,
+    max_samples: int = DEFAULT_MAX_SAMPLES,
+    random_state: int = 42,
+    callback: Optional[Callable[[str, str, str], None]] = None,
+) -> List[AutoMLResult]:
+    """用 Optuna + 子采样跑 AutoML。失败时自动回退到 GridSearchCV。
+
+    Args:
+        training_csv_path: 训练用 CSV（Step 5 产物 training_spectra.csv）
+        feature_start_column: 特征起始列名或索引（之前所有列视为目标 y）
+        preprocessing_methods: 候选预处理列表（**仅用第 1 个**，避免笛卡尔爆炸）
+        model_names: 候选模型列表（每个都会跑一遍 Optuna）
+        split_methods: 候选数据划分列表（AutoML 仅用第 1 个）
+        cv_folds: 交叉验证折数
+        output_dir: 输出目录（默认 <models_dir>_AutoML）
+        n_trials: 单模型 Optuna trial 数
+        timeout_sec: 单模型超时（秒），到时强制停止
+        max_samples: 寻优阶段允许的最大样本数
+        callback: 状态回调 callback(step_name, status, message)
+
+    Returns:
+        List[AutoMLResult]，每个目标列一份结果
+    """
+    def notify(status: str, msg: str = "") -> None:
+        if callback:
+            callback("步骤6_AutoML", status, msg)
+
+    # ---- 1) 参数默认值 ----
+    if preprocessing_methods is None:
+        preprocessing_methods = ["MMS"]
+    if model_names is None:
+        model_names = ["RF", "SVR", "Ridge"]
+    if split_methods is None:
+        split_methods = ["spxy"]
+
+    # 决策：仅用第一个预处理 + 第一个划分，避免笛卡尔爆炸
+    preproc = preprocessing_methods[0]
+    split_method = split_methods[0]
+
+    if output_dir is None:
+        output_dir = "./7_Supervised_Model_Training_AutoML"
+    out_dir = Path(output_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    preproc_dir = out_dir / preproc
+    preproc_dir.mkdir(parents=True, exist_ok=True)
+
+    # ---- 2) 加载数据 ----
+    notify("start", f"AutoML 训练开始 (n_trials={n_trials}, timeout={timeout_sec}s, max_samples={max_samples})")
+    if not Path(training_csv_path).exists():
+        return [AutoMLResult(success=False, error=f"训练 CSV 不存在: {training_csv_path}")]
+
+    df = pd.read_csv(training_csv_path)
+
+    # 提取目标列（feature_start_column 之前所有数值列）
+    if isinstance(feature_start_column, int):
+        y_cols = [c for c in df.columns[:feature_start_column]
+                  if pd.api.types.is_numeric_dtype(df[c])]
+    else:
+        try:
+            idx = list(df.columns).index(feature_start_column)
+            y_cols = [c for c in df.columns[:idx]
+                      if pd.api.types.is_numeric_dtype(df[c])]
+        except ValueError:
+            y_cols = []
+
+    if not y_cols:
+        notify("error", "AutoML: 未识别出目标列（feature_start_column 之前的所有数值列）")
+        return [AutoMLResult(success=False, error="未识别出目标列")]
+
+    feat_cols = [c for c in df.columns if c not in y_cols]
+    X_all = df[feat_cols].values.astype(np.float64)
+
+    # ---- 3) 预处理（仅第一项） ----
+    if preproc != "None":
+        try:
+            from src.preprocessing.spectral_Preprocessing import Preprocessing
+            processed = Preprocessing(preproc, df[feat_cols])
+            if isinstance(processed, pd.DataFrame):
+                X_all = processed.values.astype(np.float64)
+            else:
+                X_all = np.asarray(processed, dtype=np.float64)
+        except Exception as e:
+            notify("warning", f"预处理 {preproc} 失败: {e!r}，改用 None")
+            preproc = "None"
+
+    # ---- 4) 检查 Optuna 是否可用 ----
+    try:
+        import optuna
+        optuna.logging.set_verbosity(optuna.logging.WARNING)
+        optuna_available = True
+    except ImportError:
+        optuna_available = False
+        notify("warning", "optuna 未安装，全目标列回退到 GridSearchCV（pip install \"optuna>=3.6\"）")
+
+    # ---- 5) 逐 target 跑 ----
+    results: List[AutoMLResult] = []
+    total = len(y_cols)
+    per_model_timeout = max(10.0, timeout_sec / max(1, len(model_names)))
+
+    for ti, tgt in enumerate(y_cols, 1):
+        t0 = time.time()
+        yv = df[tgt].values.astype(np.float64)
+        mask = ~np.isnan(yv)
+        X_t = X_all[mask]
+        y_t = yv[mask]
+
+        if X_t.shape[0] < cv_folds * 2:
+            notify("warning", f"目标 {tgt}: 有效样本 {X_t.shape[0]} 不足，跳过")
+            results.append(AutoMLResult(
+                success=False, target_column=tgt, error=f"样本不足({X_t.shape[0]})",
+                preprocessing=preproc,
+            ))
+            continue
+
+        X_sub, y_sub, was_sub = smart_subsample(X_t, y_t, max_samples=max_samples, random_state=random_state)
+        if was_sub:
+            notify("info", f"目标 {tgt}: {X_t.shape[0]} 样本 → 子采样 {X_sub.shape[0]}（寻优用）")
+
+        best_overall = AutoMLResult(success=False, target_column=tgt, preprocessing=preproc)
+
+        if not optuna_available:
+            # 全目标列一次性 fallback
+            best_overall = _fallback_train(
+                training_csv_path, feature_start_column, preproc, model_names[0], split_method,
+                cv_folds, out_dir, tgt,
+            )
+        else:
+            for model_name in model_names:
+                try:
+                    builder = _build_model(model_name, random_state=random_state)
+                    if builder is None:
+                        notify("warning", f"模型 {model_name} 暂不支持 AutoML 寻优")
+                        continue
+
+                    study = optuna.create_study(
+                        direction="maximize",
+                        sampler=optuna.samplers.TPESampler(seed=random_state),
+                    )
+                    study.optimize(
+                        _make_objective(model_name, X_sub, y_sub, cv_folds, random_state),
+                        n_trials=n_trials,
+                        timeout=per_model_timeout,
+                        show_progress_bar=False,
+                    )
+
+                    if study.best_value is None or study.best_value <= -1.0:
+                        notify("warning", f"{tgt}/{model_name}: 全部 trial 失败（CV 全部 <= -1）")
+                        continue
+
+                    # refit on FULL
+                    final_model = _refit_full(model_name, study.best_params, X_t, y_t, random_state)
+                    if final_model is None:
+                        continue
+
+                    # 保存
+                    import joblib
+                    fname = f"{tgt}_{preproc}_{model_name}_AUTOML.joblib"
+                    fpath = preproc_dir / fname
+                    joblib.dump({
+                        "model": final_model,
+                        "target_column_name": tgt,
+                        "preprocess_method": preproc,
+                        "model_name": model_name,
+                        "metadata": {
+                            "automl": True,
+                            "best_params": study.best_params,
+                            "cv_score": float(study.best_value),
+                            "n_trials_done": len(study.trials),
+                            "n_samples_used_full": int(X_t.shape[0]),
+                            "n_samples_used_for_search": int(X_sub.shape[0]),
+                            "was_subsampled": was_sub,
+                            "split_method": split_method,
+                        },
+                    }, fpath)
+
+                    cand = AutoMLResult(
+                        success=True,
+                        model_path=str(fpath),
+                        cv_score=float(study.best_value),
+                        best_params=study.best_params,
+                        target_column=tgt,
+                        preprocessing=preproc,
+                        model_name=model_name,
+                        n_trials_done=len(study.trials),
+                        n_samples_used=int(X_sub.shape[0]),
+                        metadata={"refit_on_full": True, "n_samples_full": int(X_t.shape[0])},
+                    )
+                    if cand.cv_score > best_overall.cv_score:
+                        best_overall = cand
+                except Exception as e:
+                    notify("warning", f"目标 {tgt} / 模型 {model_name} 失败: {e!r}")
+                    continue
+
+            if not best_overall.success:
+                notify("warning", f"目标 {tgt} 全部 Optuna trial 失败，回退 GridSearchCV")
+                best_overall = _fallback_train(
+                    training_csv_path, feature_start_column, preproc, model_names[0], split_method,
+                    cv_folds, out_dir, tgt,
+                )
+
+        best_overall.elapsed_sec = time.time() - t0
+        results.append(best_overall)
+        notify("info", f"AutoML 目标 {tgt} 完成 ({ti}/{total}) cv={best_overall.cv_score:.4f}")
+
+    # ---- 6) 汇总 json ----
+    summary_path = out_dir / "automl_summary.json"
+    try:
+        with open(summary_path, "w", encoding="utf-8") as f:
+            json.dump([asdict(r) for r in results], f, ensure_ascii=False, indent=2, default=str)
+    except Exception as e:
+        notify("warning", f"写 automl_summary.json 失败: {e!r}")
+
+    success_n = sum(1 for r in results if r.success)
+    fallback_n = sum(1 for r in results if r.fallback_used)
+    notify("completed", f"AutoML 训练完成 {success_n}/{len(results)} 成功（{fallback_n} 走 fallback），汇总 {summary_path}")
+    return results
+
+
+# ============================================================
+# CLI 自测
+# ============================================================
+
+if __name__ == "__main__":
+    import argparse
+
+    p = argparse.ArgumentParser(description="AutoML 训练器 CLI 自测")
+    p.add_argument("--csv", required=True, help="训练用 CSV（feature_start_column 之前的列为目标 y）")
+    p.add_argument("--feature-start", default="0", help="特征起始列名或索引（默认 0）")
+    p.add_argument("--n-trials", type=int, default=DEFAULT_N_TRIALS)
+    p.add_argument("--timeout", type=float, default=DEFAULT_TIMEOUT)
+    p.add_argument("--max-samples", type=int, default=DEFAULT_MAX_SAMPLES)
+    p.add_argument("--out", default="./7_Supervised_Model_Training_AutoML")
+    args = p.parse_args()
+
+    # 智能推断 feature_start_column 类型
+    fsc: Any = args.feature_start
+    try:
+        fsc = int(fsc)
+    except ValueError:
+        pass
+
+    res = train_with_automl(
+        training_csv_path=args.csv,
+        feature_start_column=fsc,
+        n_trials=args.n_trials,
+        timeout_sec=args.timeout,
+        max_samples=args.max_samples,
+        output_dir=args.out,
+    )
+    print(f"\n训练完成 {len(res)} 个目标")
+    for r in res:
+        marker = "✓" if r.success else "✗"
+        fb = " [fallback]" if r.fallback_used else ""
+        print(f"  {marker} {r.target_column}: cv={r.cv_score:.4f} path={r.model_path}{fb}")

src/core/steps/data_preparation_step.py

@@ -126,7 +126,7 @@ class DataPreparationStep:
 
     @staticmethod
     def calculate_water_quality_indices(
-        training_spectra_path: Optional[str] = None,
+        training_csv_path: Optional[str] = None,
         formula_csv_file: Optional[str] = None,
         formula_names: Optional[List[str]] = None,
         output_file: Optional[str] = None,
@@ -153,8 +153,8 @@ class DataPreparationStep:
             notify("skipped", "跳过水质指数计算")
             return None
 
-        if training_spectra_path is None:
-            raise ValueError("必须提供 training_spectra_path 参数")
+        if training_csv_path is None:
+            raise ValueError("必须提供 training_csv_path 参数")
         if formula_csv_file is None:
             raise ValueError("必须提供 formula_csv_file 参数")
 
@@ -170,7 +170,7 @@ class DataPreparationStep:
 
         from src.utils.band_math import BandMathCalculator
 
-        calculator = BandMathCalculator(training_spectra_path)
+        calculator = BandMathCalculator(training_csv_path)
         result_df = calculator.process_formulas_from_csv(
             formula_csv_file=formula_csv_file,
             formula_names=formula_names,

src/core/water_quality_inversion_pipeline_GUI.py

@@ -173,7 +173,7 @@ class WaterQualityInversionPipeline:
         self.interpolated_img_path = None  # 存储插值后的影像路径
         self.deglint_img_path = None
         self.processed_csv_path = None
-        self.training_spectra_path = None
+        self.training_csv_path = None
         self.indices_path = None
         self.custom_regression_path = None
         
@@ -511,7 +511,7 @@ class WaterQualityInversionPipeline:
                            left_shoulder_wave: Optional[float] = None,
                            valley_wave: Optional[float] = None,
                            right_shoulder_wave: Optional[float] = None,
-                           water_mask: Optional[Union[str, np.ndarray]] = None,
+                           water_mask_path: Optional[Union[str, np.ndarray]] = None,
                            interpolate_zeros: bool = False,
                            interpolation_method: str = 'nearest',
                            enabled: bool = True,
@@ -546,7 +546,7 @@ class WaterQualityInversionPipeline:
                 left_shoulder_wave=left_shoulder_wave,
                 valley_wave=valley_wave,
                 right_shoulder_wave=right_shoulder_wave,
-                water_mask=water_mask,
+                water_mask=water_mask_path,
                 interpolate_zeros=interpolate_zeros,
                 interpolation_method=interpolation_method,
                 enabled=enabled,
@@ -655,13 +655,13 @@ class WaterQualityInversionPipeline:
             water_mask_path=self.water_mask_path,
             output_dir=str(self.training_spectra_dir),
         )
-        self.training_spectra_path = result
+        self.training_csv_path = result
         self._record_step_time("步骤5: 提取训练样本点光谱", 0, 0)
         self._notify("completed", f"训练光谱数据已保存: {result}")
         return result
 
     def step5_5_calculate_water_quality_indices(self,
-                                               training_spectra_path: Optional[str] = None,
+                                               training_csv_path: Optional[str] = None,
                                                formula_csv_file: Optional[str] = None,
                                                formula_names: Optional[List[str]] = None,
                                                output_file: Optional[str] = None,
@@ -669,29 +669,29 @@ class WaterQualityInversionPipeline:
                                                skip_dependency_check: bool = False) -> str:
         """
         步骤5.5: 根据训练光谱计算水质光谱指数
-        
+
         使用band_math.py中的方法实现，支持从公式CSV文件中批量计算指定公式
-        
+
         Args:
-            training_spectra_path: 训练光谱数据CSV路径（如果为None，使用步骤5的结果）
+            training_csv_path: 训练光谱数据CSV路径（如果为None，使用步骤5的结果）
             formula_csv_file: 公式CSV文件路径，包含公式名称和具体公式
             formula_names: 要计算的公式名称列表，如果为None则计算所有公式
             output_file: 输出文件完整路径（支持绝对路径），如果为None则使用默认路径
-            
+
         Returns:
             包含计算结果的新CSV文件路径
         """
         # 参数解析（保留原逻辑）
-        if training_spectra_path is not None:
-            csv_path = training_spectra_path
-        elif self.training_spectra_path is not None:
-            csv_path = self.training_spectra_path
+        if training_csv_path is not None:
+            csv_path = training_csv_path
+        elif self.training_csv_path is not None:
+            csv_path = self.training_csv_path
         else:
             csv_path = None
 
         self._notify("started", "步骤5.5: 计算水质光谱指数")
         result = DataPreparationStep.calculate_water_quality_indices(
-            training_spectra_path=csv_path,
+            training_csv_path=csv_path,
             formula_csv_file=formula_csv_file,
             formula_names=formula_names,
             output_file=output_file,
@@ -727,8 +727,8 @@ class WaterQualityInversionPipeline:
         # 参数解析（保留原逻辑）
         if training_csv_path is not None:
             final_csv_path = training_csv_path
-        elif self.training_spectra_path is not None:
-            final_csv_path = self.training_spectra_path
+        elif self.training_csv_path is not None:
+            final_csv_path = self.training_csv_path
         else:
             final_csv_path = None
 
@@ -911,7 +911,7 @@ class WaterQualityInversionPipeline:
         print("="*80)
         
         if training_csv_path is None:
-            training_csv_path = self.training_spectra_path
+            training_csv_path = self.training_csv_path
         if training_csv_path is None:
             raise ValueError("请提供训练数据CSV路径，或先执行步骤5")
         
@@ -1033,7 +1033,7 @@ class WaterQualityInversionPipeline:
         print("="*80)
         
         if csv_path is None:
-            csv_path = self.training_spectra_path
+            csv_path = self.training_csv_path
         if csv_path is None:
             raise ValueError("请提供CSV文件路径，或先执行步骤5")
         
@@ -1506,7 +1506,7 @@ class WaterQualityInversionPipeline:
             if 'step5' in config:
                 self._notify("步骤5: 光谱提取", "start")
                 self.step5_extract_training_spectra(**config['step5'])
-                self._notify("步骤5: 光谱提取", "completed", f"(输出: {self.training_spectra_path})")
+                self._notify("步骤5: 光谱提取", "completed", f"(输出: {self.training_csv_path})")
             else:
                 self._notify("步骤5: 光谱提取", "skipped", "未配置")
 
@@ -1615,7 +1615,7 @@ class WaterQualityInversionPipeline:
             
             # 生成散点图
             if 'visualization' in config and config['visualization'].get('generate_scatter', True):
-                if self.training_spectra_path and self.models_dir.exists():
+                if self.training_csv_path and self.models_dir.exists():
                     try:
                         self._notify("可视化", "info", "生成模型评估散点图...")
                         scatter_config = config['visualization'].get('scatter_config', {})
@@ -1653,7 +1653,7 @@ class WaterQualityInversionPipeline:
             
             # 生成光谱曲线图
             if 'visualization' in config and config['visualization'].get('generate_spectrum', True):
-                if self.training_spectra_path:
+                if self.training_csv_path:
                     try:
                         self._notify("可视化", "info", "生成光谱曲线对比图...")
                         spectrum_paths = self.generate_spectrum_comparison_plots(
@@ -1701,7 +1701,7 @@ class WaterQualityInversionPipeline:
                 pipeline_info['step2'] = {'status': 'completed', 'output_file': str(self.glint_mask_path) if self.glint_mask_path else 'N/A'}
                 pipeline_info['step3'] = {'status': 'completed', 'output_file': str(self.deglint_img_path) if self.deglint_img_path else 'N/A'}
                 pipeline_info['step4'] = {'status': 'completed', 'output_file': str(self.processed_csv_path) if self.processed_csv_path else 'N/A'}
-                pipeline_info['step5'] = {'status': 'completed', 'output_file': str(self.training_spectra_path) if self.training_spectra_path else 'N/A'}
+                pipeline_info['step5'] = {'status': 'completed', 'output_file': str(self.training_csv_path) if self.training_csv_path else 'N/A'}
                 pipeline_info['step5_5'] = {'status': 'completed', 'output_file': str(self.indices_path) if self.indices_path else 'N/A'}
                 pipeline_info['step6'] = {'status': 'completed', 'output_file': str(self.models_dir)}
                 pipeline_info['step6_75'] = {'status': 'completed', 'output_file': str(self.custom_regression_path) if self.custom_regression_path else 'N/A'}
@@ -1784,8 +1784,8 @@ class WaterQualityInversionPipeline:
         # 参数解析（保留原逻辑）
         if csv_path is not None:
             final_csv_path = csv_path
-        elif self.training_spectra_path is not None:
-            final_csv_path = self.training_spectra_path
+        elif self.training_csv_path is not None:
+            final_csv_path = self.training_csv_path
         else:
             final_csv_path = None
 
@@ -2109,7 +2109,7 @@ def main():
             'interpolation_method': 'bilinear',  # 插值方法: 'nearest'(邻近), 'bilinear'(双线性), 
                                                 # 'spline'(样条), 'kriging'(克里金)
             # 水域掩膜参数（可选）:
-            'water_mask':r"D:\BaiduNetdiskDownload\yaobao\roi\roi.shp",  # None表示自动使用步骤1生成的掩膜，也可以提供：
+            'water_mask_path':r"D:\BaiduNetdiskDownload\yaobao\roi\roi.shp",  # None表示自动使用步骤1生成的掩膜，也可以提供：
             #                      # - numpy数组
             #                      # - 栅格文件路径(.dat/.tif)
             #                      # - shapefile路径(.shp)