Step8/9 UX: feature_start 改 QComboBox + 多源 CSV 优先回退

_smoke_test_step8_9_csv_combo.py

@@ -0,0 +1,228 @@
+# -*- coding: utf-8 -*-
+"""
+Smoke test: step8_ml_train_panel + step9_ml_predict_panel 的新 CSV→QComboBox 联动
+& update_from_config 多源文件优先回退 逻辑。
+
+不真正 import 整个 panel（避免 osgeo 重依赖），只 import 类本身，验证：
+  1) Step8MlTrainPanel.feature_start 变成 QComboBox
+  2) Step8._on_training_csv_changed 正确读 CSV 表头 + 默认选中 374.285
+  3) Step8._resolve_training_csv_from_workdir 优先级（Step7 > 10_WaterIndex > Step6 > Step7 兜底）
+  4) Step8.get_config / set_config / get_training_params 用 QComboBox API 不崩
+  5) Step9._resolve_latest_wqi_test_csv 找到正确的 WQI 测试集
+  6) Step9.update_from_config 优先级（WQI CSV > Step4 raw）
+
+用 offscreen 模式避免弹窗，完整可重复跑。
+"""
+import os
+import sys
+import tempfile
+from pathlib import Path
+
+# 避免 panel 顶部 osgeo 导入阻塞
+os.environ.setdefault("QT_QPA_PLATFORM", "offscreen")
+
+# 让 src 目录可 import
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, HERE)
+
+import pandas as pd
+import numpy as np
+from PyQt5.QtWidgets import QApplication
+
+# 必须先 create QApplication 才能 new widget
+app = QApplication.instance() or QApplication(sys.argv)
+
+
+# ============================================================
+# 准备 fixture：模拟工作目录
+# ============================================================
+def make_fake_csv(path, n_cols_50_band, extra_wqi_cols=None, n_rows=10):
+    """生成模拟 CSV：50 个纯数字波段列 + 任意额外 WQI 列。"""
+    band_cols = [374.285 + i for i in range(n_cols_50_band)]
+    # 浮点列名会被 pandas 自动转成 374.285 / 374.285001 / ...
+    cols = band_cols + (extra_wqi_cols or [])
+    data = np.random.RandomState(42).rand(n_rows, len(cols))
+    df = pd.DataFrame(data, columns=cols)
+    df.to_csv(path, index=False)
+    return df
+
+
+with tempfile.TemporaryDirectory() as tmp:
+    work_dir = tmp
+    print(f"[fixture] work_dir = {work_dir}")
+
+    # 1) Step 6 原始光谱
+    os.makedirs(os.path.join(work_dir, "6_Spectral_Feature_Extraction"), exist_ok=True)
+    step6_csv = os.path.join(work_dir, "6_Spectral_Feature_Extraction", "training_spectra.csv")
+    make_fake_csv(step6_csv, 50)
+
+    # 2) Step 7 WQI 增强版
+    os.makedirs(os.path.join(work_dir, "7_Water_Quality_Indices"), exist_ok=True)
+    step7_csv = os.path.join(work_dir, "7_Water_Quality_Indices", "training_spectra_indices.csv")
+    wqi_names = [f"WQI_Chla_{i}" for i in range(10)] + [f"WQI_TP_{i}" for i in range(5)]
+    make_fake_csv(step7_csv, 50, wqi_names)
+
+    # 3) 10_WaterIndex_CSV 下的 training csv（较旧 mtime）
+    os.makedirs(os.path.join(work_dir, "10_WaterIndex_CSV"), exist_ok=True)
+    step10_csv = os.path.join(work_dir, "10_WaterIndex_CSV", "training_watercolor_summary.csv")
+    make_fake_csv(step10_csv, 50, wqi_names)
+    # 把 mtime 调成 3 年前，确认 7_Water_Quality_Indices 那个更新
+    old_time = (pd.Timestamp.now() - pd.Timedelta(days=365 * 3)).timestamp()
+    os.utime(step10_csv, (old_time, old_time))
+
+    # 4) Step 4 raw sampling
+    os.makedirs(os.path.join(work_dir, "4_sampling"), exist_ok=True)
+    step4_csv = os.path.join(work_dir, "4_sampling", "sampling_spectra.csv")
+    make_fake_csv(step4_csv, 50)
+
+    # 5) WQI 测试集（10_WaterIndex_CSV 下另一个 mtime 较新的 csv）
+    wqi_test_csv = os.path.join(work_dir, "10_WaterIndex_CSV", "sampling_watercolor_summary.csv")
+    make_fake_csv(wqi_test_csv, 50, wqi_names)
+    # 设为 1 小时前（mtime 比 step7_csv 新）
+    new_time = (pd.Timestamp.now() - pd.Timedelta(hours=1)).timestamp()
+    os.utime(wqi_test_csv, (new_time, new_time))
+
+    # 6) 9_ML_Prediction 输出目录
+    os.makedirs(os.path.join(work_dir, "9_ML_Prediction"), exist_ok=True)
+
+    # ============================================================
+    # 测试 1：Step8 加载 + QComboBox 化验证
+    # ============================================================
+    from src.gui.panels.step8_ml_train_panel import Step8MlTrainPanel
+    from PyQt5.QtWidgets import QComboBox
+
+    panel8 = Step8MlTrainPanel()
+    assert isinstance(panel8.feature_start, QComboBox), \
+        f"FAIL: feature_start 应为 QComboBox，实际为 {type(panel8.feature_start)}"
+    print(f"[OK] Step8.feature_start 类型: {type(panel8.feature_start).__name__}")
+
+    # ============================================================
+    # 测试 2：CSV 变更槽 — 模拟用户选 WQI CSV，自动填表头
+    # ============================================================
+    panel8._on_training_csv_changed(step7_csv)
+    assert panel8.feature_start.count() > 0
+    head_text = panel8.feature_start.currentText()
+    print(f"[OK] Step8 选中 step7_csv 后，feature_start.currentText() = {head_text!r}")
+    # 默认选中：第一个纯数字波段列 "374.285"
+    assert head_text.replace('.', '').lstrip('-').isdigit(), \
+        f"FAIL: 默认应选数字波段，实际 = {head_text!r}"
+    print(f"[OK] 默认选中数字波段列: {head_text!r}")
+
+    # 验证前 50 列已填充
+    n_items = panel8.feature_start.count()
+    assert n_items >= 50, f"FAIL: 应至少 50 项，实际 {n_items}"
+    print(f"[OK] Step8.feature_start 填充了 {n_items} 个候选列（前 50 列）")
+
+    # ============================================================
+    # 测试 3：CSV 不存在时，combo 回到占位项
+    # ============================================================
+    panel8._on_training_csv_changed("/nonexistent/path.csv")
+    assert panel8.feature_start.count() == 1
+    assert "请先选择训练 CSV" in panel8.feature_start.itemText(0)
+    print(f"[OK] Step8 CSV 不存在时，combo 回退到占位项: {panel8.feature_start.itemText(0)!r}")
+
+    # ============================================================
+    # 测试 4：get_config / set_config / get_training_params 用 QComboBox API 不崩
+    # ============================================================
+    panel8._on_training_csv_changed(step7_csv)
+    cfg = panel8.get_config()
+    assert 'feature_start_column' in cfg
+    assert isinstance(cfg['feature_start_column'], str)
+    assert cfg['feature_start_column'].replace('.', '').lstrip('-').isdigit()
+    print(f"[OK] Step8.get_config()['feature_start_column'] = {cfg['feature_start_column']!r}")
+
+    # set_config 试一下（先选个不存在的列名，应该走兜底 addItem 分支）
+    panel8.set_config({'feature_start_column': '999.999_unknown'})
+    # 重新触发 CSV 加载，验证回放不污染
+    panel8._on_training_csv_changed(step7_csv)
+    panel8.set_config({'feature_start_column': '374.285'})
+    assert panel8.feature_start.currentText() == '374.285'
+    print(f"[OK] Step8.set_config() 正确把 374.285 选中")
+
+    # get_training_params
+    params = panel8.get_training_params()
+    assert 'feature_start' in params
+    assert isinstance(params['feature_start'], float)
+    print(f"[OK] Step8.get_training_params()['feature_start'] = {params['feature_start']}")
+
+    # ============================================================
+    # 测试 5：_resolve_training_csv_from_workdir 优先级
+    # ============================================================
+    panel8.work_dir = work_dir
+    resolved = panel8._resolve_training_csv_from_workdir()
+    expected = step7_csv.replace('\\', '/')
+    assert resolved == expected, f"FAIL: 应优先 step7, 实际 {resolved!r}, 期望 {expected!r}"
+    print(f"[OK] Step8._resolve_training_csv_from_workdir() = {resolved}")
+    print(f"     优先级 1: Step 7 WQI 增强版 ✓")
+
+    # 模拟用户没跑过 Step 7（删掉 step7_csv 目录）— 但 10_WaterIndex_CSV 还在
+    import shutil
+    shutil.rmtree(os.path.join(work_dir, "7_Water_Quality_Indices"))
+    resolved = panel8._resolve_training_csv_from_workdir()
+    # 用户明确提到 10_WaterIndex_CSV 目录下的结果是有效来源，应作为第二优先级
+    expected_step10 = str(Path(step10_csv)).replace('\\', '/')
+    assert resolved == expected_step10, \
+        f"FAIL: 应取 10_WaterIndex_CSV/*training*, 实际 {resolved!r}, 期望 {expected_step10!r}"
+    print(f"[OK] Step 7 缺省时，回退到 10_WaterIndex_CSV/*training*: {resolved}")
+    print(f"     优先级 2: 10_WaterIndex_CSV/*training* ✓")
+
+    # 再删 10_WaterIndex_CSV 才回退到 Step 6
+    shutil.rmtree(os.path.join(work_dir, "10_WaterIndex_CSV"))
+    resolved = panel8._resolve_training_csv_from_workdir()
+    assert resolved == step6_csv.replace('\\', '/'), \
+        f"FAIL: 应回退 step6, 实际 {resolved!r}, 期望 {step6_csv!r}"
+    print(f"[OK] Step 7 / 10_WaterIndex_CSV 都缺省时，回退到 Step 6: {resolved}")
+    print(f"     优先级 3: Step 6 原始光谱 ✓")
+
+    # ============================================================
+    # 测试 6：Step9 加载
+    # ============================================================
+    from src.gui.panels.step9_ml_predict_panel import Step9MlPredictPanel
+
+    panel9 = Step9MlPredictPanel()
+    print(f"[OK] Step9MlPredictPanel 实例化成功")
+
+    # 重建 10_WaterIndex_CSV（WQI 测试集 fixture，test 5 中被删了）
+    os.makedirs(os.path.join(work_dir, "10_WaterIndex_CSV"), exist_ok=True)
+    wqi_test_csv = os.path.join(work_dir, "10_WaterIndex_CSV", "sampling_watercolor_summary.csv")
+    make_fake_csv(wqi_test_csv, 50, wqi_names)
+    new_time = (pd.Timestamp.now() - pd.Timedelta(hours=1)).timestamp()
+    os.utime(wqi_test_csv, (new_time, new_time))
+
+    # ============================================================
+    # 测试 7：Step9._resolve_latest_wqi_test_csv 优先 10_WaterIndex_CSV
+    # ============================================================
+    panel9.work_dir = work_dir
+    resolved = panel9._resolve_latest_wqi_test_csv()
+    # 应该是 sampling_watercolor_summary.csv（mtime 较新）
+    expected = wqi_test_csv.replace('\\', '/')
+    assert resolved == expected, \
+        f"FAIL: 应取 mtime 最新的 wqi 测试集, 实际 {resolved!r}, 期望 {expected!r}"
+    print(f"[OK] Step9._resolve_latest_wqi_test_csv() = {resolved}")
+    print(f"     优先级：10_WaterIndex_CSV/sampling_watercolor_summary.csv ✓")
+
+    # ============================================================
+    # 测试 8：Step9.update_from_config 优先级：WQI CSV > Step4 raw
+    # ============================================================
+    # 注：Step 4 回退路径依赖 main_window._panel_factory（offline 测试无 main_window，
+    # 故仅验证 WQI CSV 优先级主路径；Step 4 路径与原代码同源，逻辑没变）
+    # 有 WQI 时优先 WQI
+    os.makedirs(os.path.join(work_dir, "10_WaterIndex_CSV"), exist_ok=True)
+    wqi_test_csv2 = os.path.join(work_dir, "10_WaterIndex_CSV", "sampling_watercolor_summary.csv")
+    make_fake_csv(wqi_test_csv2, 50, wqi_names)
+    panel9.update_from_config(work_dir=work_dir, pipeline=None)
+    path = panel9.sampling_csv_file.get_path()
+    assert wqi_test_csv2.replace('\\', '/') in path, \
+        f"FAIL: 应优先 WQI CSV, 实际 {path!r}"
+    print(f"[OK] Step9.update_from_config() 优先 WQI CSV: {path}")
+
+    # ============================================================
+    # 测试 9：work_dir 不存在时，所有 helper 安全返回
+    # ============================================================
+    panel9.work_dir = ""
+    assert panel9._resolve_latest_wqi_test_csv() == ""
+    print(f"[OK] work_dir 空时 _resolve_latest_wqi_test_csv 安全返回空串")
+
+    print("\n" + "=" * 60)
+    print("ALL SMOKE TESTS PASSED")
+    print("=" * 60)

src/gui/panels/step8_ml_train_panel.py

@@ -14,10 +14,12 @@ if _HERE not in sys.path:
     sys.path.insert(0, _HERE)
 from src.gui.panels._step_path_resolver import get_step_output_path, resolve_step_widget, resolve_subdir
 
+import pandas as pd
+
 from PyQt5.QtWidgets import (
     QWidget, QVBoxLayout, QGroupBox, QFormLayout, QGridLayout,
     QHBoxLayout, QLabel, QLineEdit, QSpinBox, QCheckBox,
-    QPushButton, QFileDialog, QMessageBox, QSizePolicy,
+    QPushButton, QFileDialog, QMessageBox, QSizePolicy, QComboBox,
 )
 from PyQt5.QtCore import Qt
 
@@ -95,6 +97,9 @@ class Step8MlTrainPanel(QWidget):
         )
         layout.addWidget(self.training_csv_file)
 
+        # 训练 CSV 选定后，自动刷新"特征起始列"下拉框的候选列
+        self.training_csv_file.line_edit.textChanged.connect(self._on_training_csv_changed)
+
         # 机器学习模型页面
         self.ml_page = QWidget()
         self.create_ml_page()
@@ -144,19 +149,22 @@ class Step8MlTrainPanel(QWidget):
         params_group = QGroupBox("训练参数")
         params_layout = QFormLayout()
 
-        self.feature_start = QLineEdit()
+        self.feature_start = QComboBox()
-        self.feature_start.setText("374.285004")
+        self.feature_start.setMinimumWidth(180)
+        self.feature_start.setStyleSheet("""
+            QComboBox {
+                padding: 4px 8px;
+                border: 1px solid #C0C0C0;
+                border-radius: 4px;
+                min-height: 24px;
+            }
+        """)
+        # 初始空表：实际候选列由 _on_training_csv_changed 在用户选定 CSV 后填充；
+        # 此处先放一个占位项，防止未选 CSV 时下拉框完全空白（极端空状态视觉异常）
+        self.feature_start.addItem("（请先选择训练 CSV）", "")
+        self.feature_start.setCurrentIndex(0)
         params_layout.addRow("特征起始列:", self.feature_start)
 
-        # 特征起始列名提示：用记事本打开 training_spectra.csv 确认首个波长的精确表头
-        feature_start_hint = QLabel(
-            "提示：请使用记事本打开 training_spectra.csv 确认首个波长的精确表头名称"
-            "（如 374.285 或 374.285004）并在此填入，避免因浮点精度差异导致列名匹配失败。"
-        )
-        feature_start_hint.setWordWrap(True)
-        feature_start_hint.setStyleSheet("color: #666; font-size: 10px;")
-        params_layout.addRow(feature_start_hint)
-
         self.cv_folds = QSpinBox()
         self.cv_folds.setRange(2, 10)
         self.cv_folds.setValue(3)
@@ -318,6 +326,133 @@ class Step8MlTrainPanel(QWidget):
             return str(mw.work_dir)
         return ""
 
+    def _on_training_csv_changed(self, csv_path: str):
+        """训练 CSV 变更槽：自动读表头（前 50 列）填充 self.feature_start 下拉框。
+
+        触发场景：
+          - 用户点"浏览..."选了新 CSV
+          - 上游 panel_factory / set_config 注入路径
+
+        默认选中规则：
+          1) 优先 '374.285' 纯数字波段列（光谱起点；容忍 .285 / .285004 / .2850000001 等浮点变体）
+          2) 兜底：列表里第一个纯数字波段列
+          3) 仍找不到：保留占位项 "（请先选择训练 CSV）" 不动
+
+        注：nrows=1 即可读到全部表头，无需 load 整个文件
+        """
+        # 清空旧候选，保留占位策略
+        self.feature_start.blockSignals(True)
+        try:
+            self.feature_start.clear()
+            if not csv_path or not os.path.isfile(csv_path):
+                self.feature_start.addItem("（请先选择训练 CSV）", "")
+                self.feature_start.setCurrentIndex(0)
+                return
+
+            # 读取前 50 列表头
+            try:
+                df_head = pd.read_csv(csv_path, nrows=0)
+            except Exception as e:
+                print(f"[Step8] 读取 CSV 表头失败: {e}")
+                self.feature_start.addItem("（CSV 读取失败）", "")
+                self.feature_start.setCurrentIndex(0)
+                return
+
+            all_cols = list(df_head.columns)
+            head_cols = all_cols[:50]
+            if not head_cols:
+                self.feature_start.addItem("（CSV 无列）", "")
+                self.feature_start.setCurrentIndex(0)
+                return
+
+            # 填充前 50 列到下拉框
+            for col in head_cols:
+                self.feature_start.addItem(str(col), str(col))
+
+            # 默认选中规则：纯数字波段列（纯数字字符串视为波长）
+            default_idx = self._find_default_band_index(head_cols)
+            if default_idx is not None:
+                self.feature_start.setCurrentIndex(default_idx)
+            else:
+                self.feature_start.setCurrentIndex(0)
+        finally:
+            self.feature_start.blockSignals(False)
+
+    def _find_default_band_index(self, columns):
+        """在前 50 列中找到默认要选中的波段列索引。
+
+        优先级：
+          1) 列名以 '374.285' 开头（覆盖 374.285 / 374.285004 / 374.2850001）
+          2) 列名是纯数字（如 '443'）— 极少数仪器用整数波长
+          3) 第一个能被 float() 解析的纯数字列
+        """
+        # 1) 374.285 前缀
+        for i, col in enumerate(columns):
+            if str(col).startswith("374.285"):
+                return i
+        # 2) 纯数字列（无小数点）
+        for i, col in enumerate(columns):
+            if str(col).replace('.', '').lstrip('-').isdigit() and '.' not in str(col):
+                return i
+        # 3) 任意可被 float 解析的列
+        for i, col in enumerate(columns):
+            try:
+                float(str(col))
+                return i
+            except (ValueError, TypeError):
+                continue
+        return None
+
+    def _resolve_training_csv_from_workdir(self):
+        """根据工作目录智能挑选训练 CSV 路径。
+
+        优先级（从高到低）：
+          1) 7_Water_Quality_Indices/training_spectra_indices.csv（Step 7 WQI 增强版）
+          2) 10_WaterIndex_CSV/*training*  /  *training*indices*.csv（用户自定义带指数汇总）
+          3) 6_Spectral_Feature_Extraction/training_spectra.csv（Step 6 原始特征）
+          4) 7_Water_Quality_Indices/ 下任意 *training*.csv
+        """
+        work_dir = self._get_default_work_dir()
+        if not work_dir:
+            return ""
+
+        from pathlib import Path
+        wd = Path(work_dir)
+
+        # 1) Step 7 输出的训练 WQI 增强版
+        step7_csv = wd / "7_Water_Quality_Indices" / "training_spectra_indices.csv"
+        if step7_csv.is_file():
+            return str(step7_csv).replace('\\', '/')
+
+        # 2) 10_WaterIndex_CSV 下任何带 "training" 关键词的 csv（用户在 Step 10 跑过训练集）
+        idx_dir = wd / "10_WaterIndex_CSV"
+        if idx_dir.is_dir():
+            candidates = sorted(
+                idx_dir.glob("*training*.csv"),
+                key=lambda p: p.stat().st_mtime,
+                reverse=True,
+            )
+            if candidates:
+                return str(candidates[0]).replace('\\', '/')
+
+        # 3) Step 6 原始光谱特征
+        step6_csv = wd / "6_Spectral_Feature_Extraction" / "training_spectra.csv"
+        if step6_csv.is_file():
+            return str(step6_csv).replace('\\', '/')
+
+        # 4) Step 7 目录下任何 training*.csv（兜底）
+        step7_dir = wd / "7_Water_Quality_Indices"
+        if step7_dir.is_dir():
+            candidates = sorted(
+                step7_dir.glob("*training*.csv"),
+                key=lambda p: p.stat().st_mtime,
+                reverse=True,
+            )
+            if candidates:
+                return str(candidates[0]).replace('\\', '/')
+
+        return ""
+
     def browse_output_path(self):
         """浏览输出模型目录"""
         work_dir = getattr(self, 'work_dir', "")
@@ -342,7 +477,8 @@ class Step8MlTrainPanel(QWidget):
         ]
 
         config = {
-            'feature_start_column': self.feature_start.text(),
+            # QComboBox 适配：currentText() 拿用户选的列名（与原 QLineEdit.text() 语义等价）
+            'feature_start_column': self.feature_start.currentText(),
             'preprocessing_methods': preprocessing_methods if preprocessing_methods else ['None'],
             'model_names': model_names if model_names else ['SVR'],
             'split_methods': split_methods if split_methods else ['random'],
@@ -359,7 +495,15 @@ class Step8MlTrainPanel(QWidget):
     def set_config(self, config):
         """设置配置"""
         if 'feature_start_column' in config:
-            self.feature_start.setText(str(config['feature_start_column']))
+            # QComboBox 适配：先 findText 匹配候选；找不到时把值原样塞进去（兜底）
+            target = str(config['feature_start_column'])
+            idx = self.feature_start.findText(target)
+            if idx >= 0:
+                self.feature_start.setCurrentIndex(idx)
+            else:
+                # 配置回放时 CSV 可能尚未选定，候选列表为空；先插一项保留语义
+                self.feature_start.addItem(target, target)
+                self.feature_start.setCurrentIndex(self.feature_start.count() - 1)
         if 'cv_folds' in config:
             self.cv_folds.setValue(config['cv_folds'])
         if 'preprocessing_methods' in config:
@@ -393,20 +537,15 @@ class Step8MlTrainPanel(QWidget):
         else:
             self.work_dir = None
 
-        # 1. 强制读 Step 6 的 training_spectra.csv（光谱特征提取结果）
+        # 1. 智能挑选训练 CSV（不再"强制"读 Step 6，而是优先 WQI 增强版）
-        #    修复张冠李戴：原链路 STEP_DATA_SOURCE['training_spectra_csv'] → step5_clean_panel
+        #    优先级：Step 7 WQI > 10_WaterIndex_CSV/*training* > Step 6 原始光谱 > Step 7 兜底
-        #    错误地指向了 Step 5 的 processed_data.csv（纯清洗数据，不含光谱特征），
+        #    修复目标：用户跑过 Step 7 后再回到 Step 8，UI 默认应指向带指数的训练集，
-        #    实际 ML 训练需要的特征数据来自 Step 6 的 6_Spectral_Feature_Extraction/training_spectra.csv
+        #    否则训练好的模型有 95 维（50 波段 + 45 WQI），下次回放变成只 50 维训练，特征维数错位。
-        main_window = self.window()
         existing_training_csv = self.training_csv_file.get_path()
         if not existing_training_csv or not existing_training_csv.strip():
-            if self.work_dir:
+            candidate = self._resolve_training_csv_from_workdir()
-                step6_dir = resolve_subdir(self.work_dir, 'spectral_feature')
+            if candidate:
-                step6_training_csv = os.path.join(
+                self.training_csv_file.set_path(candidate)
-                    step6_dir, 'training_spectra.csv'
-                ).replace('\\', '/')
-                if step6_training_csv:
-                    self.training_csv_file.set_path(step6_training_csv)
 
         # 2. 自动填充输出目录为 8_Machine_Learning_Models
         if self.work_dir:
@@ -448,7 +587,9 @@ class Step8MlTrainPanel(QWidget):
         """获取模型训练参数"""
         return {
             'pipeline_type': 'machine_learning',
-            'feature_start': float(self.feature_start.text()),
+            # QComboBox 适配：currentText() 取列名；下拉项里就是"374.285004" 等纯数字波段名
+            # （与原 QLineEdit 中 "374.285004" 字符串保持完全一致，后端 float() 解析不变）
+            'feature_start': float(self.feature_start.currentText()),
             'cv_folds': self.cv_folds.value(),
             'preprocess_methods': [method for method, cb in self.preproc_checkboxes.items() if cb.isChecked()],
             'model_types': [model for model, cb in self.model_checkboxes.items() if cb.isChecked()],

src/gui/panels/step9_ml_predict_panel.py

@@ -8,6 +8,8 @@ import os
 import sys
 from pathlib import Path
 
+import pandas as pd
+
 # 路径归一化 helper（与 pipeline.get_step_output_dir 互为表里）
 _HERE = os.path.dirname(os.path.abspath(__file__))
 if _HERE not in sys.path:
@@ -333,24 +335,90 @@ class Step9MlPredictPanel(QWidget):
                     result[name] = self.external_models_dict[name]
         return result
 
+    def _resolve_latest_wqi_test_csv(self):
+        """在工作目录中智能挑选"最新生成的、含 WQI 指数的测试集 CSV"。
+
+        返回：找到则返回文件路径字符串；找不到返回 ""。
+
+        搜索策略（按优先级递减，命中即返回）：
+          1) 10_WaterIndex_CSV/*.csv  — Step 10 输出目录（用户在 Step 10 跑过的产品）
+          2) 7_Water_Quality_Indices/*sampling*.csv / *test*.csv  — 用户手动对采样点算过 WQI
+          3) work_dir 下任何 *indices*.csv / *wqi*.csv（不区分大小写）
+          4) work_dir 下任何 > 60 列的 csv（启发式：50 波段 + > 10 WQI 指数列）
+          5) 兜底空串，调用方回退到 Step 4 sampling_spectra.csv
+
+        多个候选时按 mtime 倒序选"最新生成的"。
+        """
+        work_dir = self._get_default_work_dir()
+        if not work_dir:
+            return ""
+
+        wd = Path(work_dir)
+        found = []
+
+        # 1) 10_WaterIndex_CSV 下所有 csv（Step 10 输出）
+        idx_dir = wd / "10_WaterIndex_CSV"
+        if idx_dir.is_dir():
+            found.extend(idx_dir.glob("*.csv"))
+
+        # 2) 7_Water_Quality_Indices 下与采样/测试相关的 csv
+        qa_dir = wd / "7_Water_Quality_Indices"
+        if qa_dir.is_dir():
+            for pattern in ("*sampling*.csv", "*test*.csv", "*predict*.csv"):
+                found.extend(qa_dir.glob(pattern))
+
+        # 3) work_dir 直接子树下含 indices/wqi 关键词的 csv
+        for keyword in ("*indices*.csv", "*wqi*.csv", "*WQI*.csv"):
+            found.extend(wd.rglob(keyword))
+
+        # 4) 启发式：> 60 列的 csv（50 波段 + 至少 10 个指数）
+        try:
+            for csv in wd.rglob("*.csv"):
+                if csv in found:
+                    continue
+                try:
+                    head = pd.read_csv(csv, nrows=0)
+                    if head.shape[1] > 60:
+                        found.append(csv)
+                except Exception:
+                    pass  # 读取失败就跳过，不影响其它候选
+        except Exception:
+            pass
+
+        if not found:
+            return ""
+
+        # 去重 + 按 mtime 倒序排
+        uniq = {p.resolve(): p for p in found}.values()
+        sorted_paths = sorted(uniq, key=lambda p: p.stat().st_mtime, reverse=True)
+        return str(sorted_paths[0]).replace('\\', '/')
+
     def update_from_config(self, work_dir=None, pipeline=None):
         if work_dir: self.work_dir = work_dir
 
         main_window = self.window()
         factory = getattr(main_window, '_panel_factory', None) if main_window else None
-        if not factory: return
 
-        # 1. 拿第 4 步的采样光谱
+        # 1. 智能挑选采样 CSV：优先"含 WQI 指数的测试集"（防止特征维度与训练时不匹配）
-        step4_panel = factory.get_panel('step4_sampling')
+        #    修复目标：用户在 Step 8 用 95 维 (50+45 WQI) 训练 → Step 9 默认读 50 维 raw sampling
-        if step4_panel and hasattr(step4_panel, 'output_file'):
+        #    时 inference_batch.preprocess_spectra 会触发"自动特征补全"逻辑；但若用户已经在
-            path = step4_panel.output_file.get_path()
+        #    Step 10/手工把指数算到 CSV 里了，应该直接用那个文件（少走内存补全、避免 band 列顺序漂移）
-            if path: self.sampling_csv_file.set_path(path)
+        wqi_test_csv = self._resolve_latest_wqi_test_csv()
+        if wqi_test_csv:
+            self.sampling_csv_file.set_path(wqi_test_csv)
+        elif factory:
+            # 兜底：拿第 4 步的纯原始采样光谱（旧行为保留）
+            step4_panel = factory.get_panel('step4_sampling')
+            if step4_panel and hasattr(step4_panel, 'output_file'):
+                path = step4_panel.output_file.get_path()
+                if path: self.sampling_csv_file.set_path(path)
 
         # 2. 拿第 8 步的模型目录
-        step8_panel = factory.get_panel('step8_ml_train')
+        if factory:
-        if step8_panel and hasattr(step8_panel, 'output_path'):
+            step8_panel = factory.get_panel('step8_ml_train')
-            path = step8_panel.output_path.get_path()
+            if step8_panel and hasattr(step8_panel, 'output_path'):
-            if path: self.models_dir_file.set_path(path)
+                path = step8_panel.output_path.get_path()
+                if path: self.models_dir_file.set_path(path)
 
         # 3. 生成第 9 步的输出目录
         if hasattr(self, 'work_dir') and self.work_dir: