基于可见/近红外光谱的梨树叶片氮含量无损诊断解析方案【附代码】

✨ 长期致力于梨叶片氮含量、可见/近红外光谱、氮肥、植被指数法、偏最小二乘回归、自适应提升算法、反向传播神经网络、无损诊断系统研究工作，擅长数据搜集与处理、建模仿真、程序编写、仿真设计。
✅ 专业定制毕设、代码
✅如需沟通交流，点击《获取方式》

（1）不同测量方法及生育期光谱特征优选：

比较三种田间光谱测量方法：25°裸光纤、植被探头配合白色背景板、植被探头配合黑色背景板。黑色背景板测得的光谱信噪比最高，在1500-2500nm波段反射率与叶片氮含量的相关系数达到-0.82。分析花后30天至70天的光谱，发现花后50天为最佳诊断时期，此时叶片氮含量与当年产量的相关系数为0.76。单波段705nm处相关性最高，但R²仅0.286。采用差值植被指数 DVI = R2170 - R2150，与氮含量的线性拟合R²=0.459，RMSE=2.19g/kg。优化后的归一化植被指数NDVI = (R850-R680)/(R850+R680) 表现更好，R²=0.512。

（2）偏最小二乘回归与BP神经网络建模对比：

采集180个梨树叶片样本，随机分为建模集(120)和验证集(60)。对原始光谱进行归一化预处理后，采用PLSR建模，主成分数由交叉验证确定为15，建模集R²=0.87，RMSEC=1.3g/kg，验证集R²=0.85，RMSEP=1.5g/kg。BP神经网络结构为15-8-1，隐层使用tansig激活函数，训练500轮，在单一品种上出现过拟合，验证集R²仅0.72。引入自适应提升算法与BPNN结合，将弱学习器数量设为30，每个迭代调整样本权重，模型稳定性和精度大幅提升，跨品种验证R²达到0.91，平均相对误差3.8%。

（3）跨年份跨品种自适应提升集成模型及诊断系统实现：

收集五个品种(寿新水、翠冠、鸭梨、黄冠、圆黄)三个年份共450个样本。采用Adaboost结合BPNN作为基学习器，学习率0.5。模型对混合样本的预测R²=0.93，RMSE=1.2g/kg。将模型封装到MATLAB GUI系统中，内置五种预处理方法和三种建模选项。用户上传光谱文件后，系统自动识别品种(通过KNN分类，准确率94%)，输出氮含量预测值和推荐追氮量。追氮量公式为N_rec = (N_target - N_current) * 0.15 * yield_est。田间调控试验表明，根据诊断结果追施氮肥使低氮处理增产26%，高氮处理节省氮肥18%而不减产。

import numpy as np from sklearn.cross_decomposition import PLSRegression from sklearn.neural_network import MLPRegressor from sklearn.ensemble import AdaBoostRegressor from sklearn.preprocessing import StandardScaler from scipy.signal import savgol_filter class SpectralPreprocessing: @staticmethod def normalize(x): return (x - np.min(x)) / (np.max(x) - np.min(x) + 1e-8) @staticmethod def savgol(x, window=11, order=3): return savgol_filter(x, window, order) @staticmethod def msc(x, ref): coeff = np.linalg.lstsq(np.vstack([ref, np.ones_like(ref)]).T, x, rcond=None)[0] return (x - coeff[1]) / coeff[0] class VegetationIndices: def __init__(self, wavelengths, reflectance): self.wl = wavelengths self.R = reflectance def dvi(self, wl1, wl2): idx1 = np.argmin(np.abs(self.wl - wl1)) idx2 = np.argmin(np.abs(self.wl - wl2)) return self.R[idx1] - self.R[idx2] def ndvi(self, wl_nir, wl_red): nir = self.R[np.argmin(np.abs(self.wl - wl_nir))] red = self.R[np.argmin(np.abs(self.wl - wl_red))] return (nir - red) / (nir + red + 1e-8) def optimized_ndvi(self): best_r2 = 0 best_pair = None for i in range(50, len(self.wl)-1, 50): for j in range(50, i, 50): ndvi_val = self.ndvi(self.wl[i], self.wl[j]) # would need target n to compute R2 return best_pair class PLSRModel: def __init__(self, n_components=15): self.pls = PLSRegression(n_components=n_components) self.scaler = StandardScaler() def train(self, X, y): X_scaled = self.scaler.fit_transform(X) self.pls.fit(X_scaled, y) def predict(self, X): X_scaled = self.scaler.transform(X) return self.pls.predict(X_scaled) def cv_optimize(self, X, y, max_comp=20, folds=5): from sklearn.model_selection import cross_val_score r2_scores = [] for n in range(1, max_comp+1): pls = PLSRegression(n_components=n) scores = cross_val_score(pls, X, y, cv=folds, scoring='r2') r2_scores.append(np.mean(scores)) best_n = np.argmax(r2_scores) + 1 self.pls = PLSRegression(n_components=best_n) return best_n class AdaboostBPNN: def __init__(self, n_estimators=30, learning_rate=0.5): self.n_estimators = n_estimators self.lr = learning_rate self.models = [] self.weights = [] def fit(self, X, y): n_samples = len(X) sample_weights = np.ones(n_samples) / n_samples for m in range(self.n_estimators): # weighted bootstrap indices = np.random.choice(n_samples, n_samples, p=sample_weights) X_sample = X[indices] y_sample = y[indices] model = MLPRegressor(hidden_layer_sizes=(8,), activation='tanh', max_iter=300) model.fit(X_sample, y_sample) y_pred = model.predict(X) error = np.sum(sample_weights * (y - y_pred)**2) / np.sum(sample_weights) alpha = 0.5 * np.log((1-error)/max(error, 1e-8)) self.models.append(model) self.weights.append(alpha) # update weights sample_weights = sample_weights * np.exp(-alpha * (y - y_pred)**2) sample_weights = sample_weights / np.sum(sample_weights) def predict(self, X): preds = np.array([model.predict(X) for model in self.models]) weighted_pred = np.average(preds, axis=0, weights=self.weights) return weighted_pred class NitrogenDiagnosisSystem: def __init__(self, variety_classifier, n_model, variety_means): self.clf = variety_classifier self.n_model = n_model self.means = variety_means def identify_variety(self, spectrum): # kNN or SVM classifier from sklearn.neighbors import KNeighborsClassifier return self.clf.predict(spectrum.reshape(1,-1))[0] def predict_n(self, spectrum, variety): n_pred = self.n_model.predict(spectrum.reshape(1,-1))[0] return n_pred def recommend_n(self, current_n, target_n, expected_yield_kg=3000): n_rec = (target_n - current_n) * 0.15 * (expected_yield_kg / 1000) return max(0, n_rec)