智慧海洋

智慧海洋竞赛介绍Task01
• 本赛题围绕“智慧海洋建设，赋能海上安全治理能力现代化”展开。而船舶避碰终端（AIS）、北斗定位终端等通信导航设备的应用，给海上交通和作业带来了极大便利，但同时存在设备信息使用不规范造成的巨大人身和财产损失，给海上安全治理带来了新的挑战。

本赛题基于船舶轨迹位置数据对海上目标进行智能识别和作业行为分析，要求选手通过分析渔船北斗设备位置数据，得出该船的生产作业行为，具体判断出是拖网作业、围网作业还是流刺网作业。同时，希望选手通过数据可视分析，挖掘更多海洋通信导航设备的应用价值。

作业

作业
基础作业：
1.尝试去使用kepler.gl可视化来分析不同类型船舶AIS数据的分布情况，并为接下来的特征工程的提取建立基础
进阶作业：
2.在这个模块中，我们介绍了各种库以及他们常用的方法。如果可以，请同学们尝试在原有剔除异常点的数据（DF）中保留douglas-peucker算法所识别的关键点的数据，删除douglas-peucker未保存的数据，并尝试对这些坐标点进行geohash编码

特征工程

经纬度常规的统计特征对渔船轨迹信息的表征能力其实是有限的，这里我们设计了一种轨迹序列的编码方式（该方法可拓展至其他类型的轨迹序列数据），进一步刻画了轨迹的动态信息和渔船经过的每个点之间的联系。
比较亮点的特征工程

import os
import pandas as pd
from tqdm import tqdm
from sklearn.manifold import TSNE
from sklearn.cluster import KMeans
import matplotlib.pyplot as plt
import matplotlib.colors as cl
import random
from fastdtw import fastdtw
from scipy.spatial.distance import euclidean


# step 1：粗匹配，也可对每个ID对应的lat-lon集合进行采样后再计算交集，效率会大大提升。
def match_ship_in_ais_step1(train_file, ais_file, threshold, cnt_limit=50):
    paths = os.listdir(ais_file)
    match_df = pd.DataFrame()

    # 复赛训练集
    train = pd.read_csv(train_file)
    train.columns = ['ID', 'lat', 'lon', 'speed', 'direction', 'time', 'type']
    train.sort_values(['ID', 'time'], inplace=True)
    train['lat'] = train['lat'].round(2)
    train['lon'] = train['lon'].round(2)
    print('origin train records num:', train.shape[0])
    train['time'] = train['time'].apply(lambda x: x[:4])
    train = train.drop_duplicates(subset=['ID', 'time', 'lat', 'lon'])
    print('after drop duplicates, train records num:', train.shape[0])

    max_lat, min_lat, max_lon, min_lon = train['lat'].max(), train['lat'].min(), train['lon'].max(), train['lon'].min()
    print('*'*100)
    train_agg = train.groupby('ID')['lat', 'lon'].agg(list).reset_index()
    train_cnt = train.groupby(['ID'])['time'].agg({'cnt': 'count'}).reset_index()
    train_cnt = train_cnt[train_cnt['cnt'] > cnt_limit]
    train_agg = train_agg[train_agg['ID'].isin(train_cnt['ID'].unique())]
    train_time = train.groupby('ID')['time', 'type'].agg('min').reset_index()
    train_agg = train_agg.merge(train_time, on='ID', how='left')
    train_agg['lon_lat'] = train_agg.apply(lambda x: {'{}_{}'.format(z[0], z[1]) for z in zip(x['lat'], x['lon'])}, axis=1)
    del train_agg['lat'], train_agg['lon']
    for t in tqdm(range(len(paths))):
        p = paths[t]
        print(f'start match train ship with ais: {p}')
        ais = pd.read_csv(f'{ais_file}/{p}')
        ais.columns = ['ais_speed', 'ais_direction', 'ais_id', 'time', 'ais_lon', 'ais_lat']
        ais.sort_values(['ais_id', 'time'], inplace=True)
        print('origin ais records num:', ais.shape[0])
        ais['time'] = ais['time'].apply(lambda x:x[:4])
        ais['ais_lat'] = ais['ais_lat'].round(2)
        ais['ais_lon'] = ais['ais_lon'].round(2)
        ais = ais.drop_duplicates(subset=['ais_id', 'time', 'ais_lat', 'ais_lon'])
        print('after drop duplicates, ais records num:', ais.shape[0])
        ais = ais[(ais['ais_lat'] > min_lat) & (ais['ais_lat'] < max_lat) & (ais['ais_lon'] > min_lon) & (ais['ais_lon'] < max_lon)]
        print('remove beyond threshold lat and lon, ais records num:', ais.shape[0])
        ais_cnt = ais.groupby(['ais_id'])['time'].agg({'cnt': 'count'}).reset_index()
        ais_cnt = ais_cnt[ais_cnt['cnt'] > cnt_limit]
        ais_time = ais.groupby('ais_id')['time'].agg('min').reset_index()
        ais_agg = ais.groupby('ais_id')['ais_lat', 'ais_lon'].agg(list).reset_index()
        ais_agg = ais_agg[ais_agg['ais_id'].isin(ais_cnt['ais_id'].unique())]
        ais_agg = ais_agg.merge(ais_time, on='ais_id', how='left')
        ais_agg['ais_lon_lat'] = ais_agg.apply(lambda x: {'{}_{}'.format(z[0], z[1]) for z in zip(x['ais_lat'], x['ais_lon'])}, axis=1)
        del ais_agg['ais_lat'], ais_agg['ais_lon']
        all_df = train_agg.merge(ais_agg, on='time', how='inner')
        if all_df.shape[0] == 0:
            continue
        all_df['share_size'] = all_df.apply(lambda x: len(x['lon_lat'] & x['ais_lon_lat']), axis=1)
        all_df['ratio_x'] = all_df.apply(lambda x: x['share_size']/len(x['ais_lon_lat']), axis=1)
        all_df['ratio_y'] = all_df.apply(lambda x: x['share_size']/len(x['lon_lat']), axis=1)
        all_df = all_df[(all_df['ratio_x'] >= threshold) & (all_df['ratio_y'] >= threshold)]
        all_df['ais_file_name'] = p.split('.')[0]
        match_df = match_df.append(all_df[['ID', 'type', 'ais_id', 'share_size', 'ratio_x', 'ratio_y', 'ais_file_name']])
        print('now match:', match_df.shape[0])
        print('*'*100)
    return match_df


# step 2：细匹配并可视化匹配结果
def match_ship_in_ais_step2(ids, train_file, ais_file, threshold):
    map_net = {'刺网': 'cw', '围网': 'ww', '拖网': 'tw'}
    train = pd.read_csv(train_file)
    train = train[train['渔船ID'] == ids[0]].sort_values('time')

    ais = pd.read_csv(f'{ais_file}/{ids[2]}.csv')
    ais = ais[ais['ais_id'] == ids[1]].sort_values('time')

    distance, _ = fastdtw(train[['lat', 'lon']].values, ais[['lat', 'lon']].values, dist=euclidean)
    if distance < threshold:
        fig = plt.figure(figsize=(8, 4))
        plt.subplot(2, 2, 1)
        net_type = train['type'].map(map_net).min()
        plt.scatter(train['lat'], train['lon'], color='blue')
        plt.subplot(2, 2, 2)
        plt.scatter(ais['lat'], ais['lon'], color='red')
        fig.suptitle(f'match ID:{ids[0]}, type: {net_type}, ais_id:{ids[1]}, ais_file:{ids[2]}')
        plt.show()


# 每日小时的渔船count
def stat_hour(ais_file):
    paths = os.listdir(ais_file)
    df = pd.DataFrame()
    for t in tqdm(range(len(paths))):
        p = paths[t]
        ais = pd.read_csv(f'{ais_file}/{p}')
        ais['hour'] = ais['time'].apply(lambda x: x.split(' ')[0]+x.split(' ')[1][:2])
        temp = ais.groupby('hour')['ais_id'].agg('nunique').reset_index()
        df = df.append(temp)
    return df


# 获取指定ais设备的数据
def get_record(ais_file, ids):
    paths = os.listdir(ais_file)
    df = pd.DataFrame()
    for t in tqdm(range(len(paths))):
        p = paths[t]
        ais = pd.read_csv(f'{ais_file}/{p}')
        ais = ais[ais['ais_id'] == ids]
        ais['date'] = int(p[11:19])
        df = df.append(ais)
    return df


# 可视化ais设备轨迹
def plot_record(df):
    n_days = df['date'].nunique()
    rows = n_days//6
    date_list = sorted(df['date'].unique())[:6*rows]
    fig, axs = plt.subplots(rows, 6, figsize=(18, 12))

    for i, j in enumerate(date_list):
        tmp = df[df['date'] == j]
        tmp.sort_values('time', inplace=True)
        axs[i % rows][i//rows].scatter(tmp['lat'], tmp['lon'])
        axs[i % rows][i//rows].set_title(j)
        axs[i % rows][i//rows].set_xticks([])
        axs[i % rows][i//rows].set_yticks([])
    plt.show()


# 线性插值
def liner_insert_values(df):
    df['time'] = df['time'].apply(
        lambda x: '2019-' + x.split(' ')[0][:2] + '-' + x.split(' ')[0][2:] + ' ' + x.split(' ')[1][:-2] + '00')
    df['time'] = pd.to_datetime(df['time'])
    df = df.set_index('time')
    ship_id = df['ID'].unique()[0]
    df = df.resample('10T', closed='left')[['lat', 'lon']].mean().reset_index()
    df['lat'] = df['lat'].interpolate()
    df['lon'] = df['lon'].interpolate()
    df['ID'] = ship_id
    return df


def show_cluster(df, use_cols, n_clusters):
    km = KMeans(n_clusters=n_clusters, random_state=2019)
    df['type'] = km.fit_predict(df[use_cols])
    tsne_model = TSNE(n_components=2, random_state=1024)
    tsne_model.fit_transform(df[use_cols])
    tsne_df = pd.DataFrame(tsne_model.embedding_, index=df.index)
    tsne_df['type'] = df['type']

    color_list = random.sample(cl.cnames.keys(), n_clusters)

    for n, c in zip(range(n_clusters), color_list):
        d = tsne_df[tsne_df['type'] == n]
        plt.plot(d[0], d[1], color=c, marker='.', linestyle='')
    plt.legend(range(n_clusters))
    plt.xticks([])
    plt.yticks([])

    plt.show()


if __name__ == "__main__":
    save_path = 'match.csv'
    ais_path = 'round2_ais_20200310'
    match_result = match_ship_in_ais_step1('train_round2.csv', ais_path, 0.5, cnt_limit=50)
    match_result.to_csv(save_path, index=False)
    for m in match_result[['ID', 'ais_id', 'ais_file_name']].values:
        match_ship_in_ais_step2(m, 'train_round2.csv', ais_path, 50)
        

代码还在理解