app.py

import os
import streamlit as st
import zipfile
import torch
from utils import *
import matplotlib.pyplot as plt
from matplotlib import colors

if not hasattr(st, 'paths'):
    st.paths = None
if not hasattr(st, 'daily_model'):
    best_model_daily_file_name = "best_model_daily.pth"
    best_model_annual_file_name = "best_model_annual.pth"

    first_input_batch = torch.zeros(71, 9, 5, 48, 48)
    # first_input_batch = first_input_batch.view(-1, *first_input_batch.shape[2:])
    st.daily_model = FPN(opt, first_input_batch, opt.win_size)
    st.annual_model = SimpleNN(opt)

    if torch.cuda.is_available():
        st.daily_model = torch.nn.DataParallel(st.daily_model).cuda()
        st.annual_model = torch.nn.DataParallel(st.annual_model).cuda()
        st.daily_model = torch.nn.DataParallel(st.daily_model).cuda()
        st.annual_model = torch.nn.DataParallel(st.annual_model).cuda()
    else:
        st.daily_model = torch.nn.DataParallel(st.daily_model).cpu()
        st.annual_model = torch.nn.DataParallel(st.annual_model).cpu()
        st.daily_model = torch.nn.DataParallel(st.daily_model).cpu()
        st.annual_model = torch.nn.DataParallel(st.annual_model).cpu()

    print('trying to resume previous saved models...')
    state = resume(
        os.path.join(opt.resume_path, best_model_daily_file_name),
        model=st.daily_model, optimizer=None)
    state = resume(
        os.path.join(opt.resume_path, best_model_annual_file_name),
        model=st.annual_model, optimizer=None)
    st.daily_model = st.daily_model.eval()
    st.annual_model = st.annual_model.eval()

# Load Model
# @title Load pretrained weights


st.title('In-season and dynamic crop mapping using 3D convolution neural networks and sentinel-2 time series')
st.markdown(""" Demo App for the model presented in the paper:
```
@article{gallo2022in_season,
  title = {In-season and dynamic crop mapping using 3D convolution neural networks and sentinel-2 time series},
  journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
  volume = {195},
  pages = {335-352},
  year = {2023},
  issn = {0924-2716},
  doi = {https://doi.org/10.1016/j.isprsjprs.2022.12.005},
  url = {https://www.sciencedirect.com/science/article/pii/S0924271622003203},
  author = {Ignazio Gallo and Luigi Ranghetti and Nicola Landro and Riccardo {La Grassa} and Mirco Boschetti},
}
```
""")

file_uploaded = st.file_uploader(
    "Upload",
    type=["zip"],
    accept_multiple_files=False,
)
sample_path = None
if file_uploaded is not None:
    with zipfile.ZipFile(file_uploaded, "r") as z:
        z.extractall("uploaded_samples")
    sample_path = "uploaded_samples/" + file_uploaded.name[:-4]
st.markdown('or use a demo sample')
if st.button('sample 1'):
    sample_path = 'demo_data/lombardia'

paths = None
if sample_path is not None:
    st.markdown(f'elaborating {sample_path}...')

    validationdataset = SentinelDailyAnnualDatasetNoLabel(
        sample_path,
        opt.years,
        opt.classes_path,
        opt.sample_duration,
        opt.win_size,
        tileids=None)
    validationdataloader = torch.utils.data.DataLoader(
        validationdataset, batch_size=opt.batch_size, shuffle=False, num_workers=opt.workers)

    st.markdown(f'predict in progress...')

    out_dir = os.path.join(opt.result_path, "seg_maps")
    if not os.path.exists(out_dir):
        os.makedirs(out_dir)
    for i, (x_dailies, dates, dirs_path) in enumerate(validationdataloader):
        with torch.no_grad():
            # x_dailies, dates, dirs_path = next(iter(validationdataloader))
            # reshape merging the first two dimensions
            x_dailies = x_dailies.view(-1, *x_dailies.shape[2:])
            if torch.cuda.is_available():
                x_dailies = x_dailies.cuda()

            feat_daily, outs_daily = st.daily_model.forward(x_dailies)
            # return to original size of batch and year
            outs_daily = outs_daily.view(
                opt.batch_size, opt.sample_duration, *outs_daily.shape[1:])
            feat_daily = feat_daily.view(
                opt.batch_size, opt.sample_duration, *feat_daily.shape[1:])

            _, out_annual = st.annual_model.forward(feat_daily)
            pred_annual = torch.argmax(out_annual, dim=1).squeeze(1)
            pred_annual = pred_annual.cpu().numpy()
            # Remapping the labels
            pred_annual_nn = ids_to_labels(
                validationdataloader, pred_annual).astype(numpy.uint8)

            for batch in range(feat_daily.shape[0]):
                # _, profile = read(os.path.join(dirs_path[batch], '20191230_MSAVI.tif'))  # todo get the last image
                _, tmp_path = get_patch_id(validationdataset.samples, 0)
                dates = get_all_dates(
                    tmp_path, validationdataset.max_seq_length)
                last_tif_path = os.path.join(tmp_path, dates[-1] + ".tif")
                _, profile = read(last_tif_path)
                profile["name"] = dirs_path[batch]

                pth = dirs_path[batch].split(os.path.sep)[-3:]
                full_pth_patch = os.path.join(
                    out_dir, pth[1] + '-' + pth[0], pth[2])

                if not os.path.exists(full_pth_patch):
                    os.makedirs(full_pth_patch)
                full_pth_pred = os.path.join(
                    full_pth_patch, 'patch-pred-nn.tif')
                profile.update({
                    'nodata': None,
                    'dtype': 'uint8',
                    'count': 1})
                with rasterio.open(full_pth_pred, 'w', **profile) as dst:
                    dst.write_band(1, pred_annual_nn[batch])

                # patch_predictions = None
                for ch in range(len(dates)):
                    soft_seg = outs_daily[batch, ch, :, :, :]
                    # transform probs into a hard segmentation
                    pred_daily = torch.argmax(soft_seg, dim=0)
                    pred_daily = pred_daily.cpu()
                    daily_pred = ids_to_labels(
                        validationdataloader, pred_daily).astype(numpy.uint8)
                    # if patch_predictions is None:
                    #     patch_predictions = numpy.expand_dims(daily_pred, axis=0)
                    # else:
                    #     patch_predictions = numpy.concatenate((patch_predictions, numpy.expand_dims(daily_pred, axis=0)),
                    #                                           axis=0)

                    # save GT image in  opt.root_path
                    full_pth_date = os.path.join(
                        full_pth_patch, dates[ch][batch] + f'-ch{ch}-b{batch}-daily-pred.tif')
                    profile.update({
                        'nodata': None,
                        'dtype': 'uint8',
                        'count': 1})
                    with rasterio.open(full_pth_date, 'w', **profile) as dst:
                        dst.write_band(1, daily_pred)

    st.markdown('End prediction')

    folder = "demo_data/results/seg_maps/example-lombardia/2"
    st.paths = os.listdir(folder)

if st.paths is not None:
    folder = "demo_data/results/seg_maps/example-lombardia/2"
    file_picker = st.selectbox("Select day predict (annual is patch-pred-nn.tif)",
                               st.paths, index=st.paths.index('patch-pred-nn.tif'))

    file_path = os.path.join(folder, file_picker)
    # print(file_path)
    target, profile = read(file_path)
    target = np.squeeze(target)
    target = [classes_color_map[p] for p in target]

    fig, ax = plt.subplots()
    ax.imshow(target)

    markdown_legend = ''
    for c, l in zip(color_labels, labels_map):
        # print(colors.to_hex(c))
        markdown_legend += f'<div style="color:gray;background-color: {colors.to_hex(c)};">{l}</div><br>'

    col1, col2 = st.columns(2)
    with col1:
        st.pyplot(fig)
    with col2:
        st.markdown(markdown_legend, unsafe_allow_html=True)

st.markdown(""" 
    ## Lombardia Dataset
    You can download other patches from the original dataset created and published on 
    [Kaggle](https://www.kaggle.com/datasets/ignazio/sentinel2-crop-mapping) and used in our paper.
    
    ## How to build an input file for the Demo
    Using a daily FPN and giving a zip that contains 30 tiff with 7 channels, correctly named you can reach prediction of 
    crop mapping og the area...
    """)