Cacher is the code snippet organizer for pro developers

# MSDFF: Multi-Source Deep Feature Framework for Landslide Detection ## Overview This repository contains a PyTorch implementation framework for the **Multi-Source Deep Feature Framework (MSDFF)**, designed for multi-source and multi-temporal landslide segmentation, latent feature extraction, spatial characterization, and cross-dataset evaluation. MSDFF combines complementary information from Sentinel-2 optical imagery, Sentinel-1 synthetic aperture radar (SAR), digital elevation models (DEM), and slope. Modality-specific spatial encoders extract representations from the different input sources. ConvGRU-based temporal encoding and temporal attention are used to represent changes across satellite observations. A gated fusion module integrates the encoded features, and a segmentation decoder produces pixel-level landslide predictions. The repository also includes utilities for preparing aligned patch archives, training the model, evaluating segmentation predictions, and analyzing learned latent representations. **Important:** The scripts in this repository provide a reproducible code framework and define the expected input format. They should be checked against the exact preprocessing records, model checkpoint, and final experimental implementation before being used to claim exact reproduction of reported manuscript results. The repository does not contain the original satellite datasets or pretrained weights. ## Main Components | File | Main functionality | | --- | --- | | `dataset.py` | Reads preprocessed NPZ patches, validates arrays, creates deterministic data splits, and builds PyTorch DataLoaders. | | `preprocessing.py` | Provides patch validation, band selection, temporal date matching, validity masks, training-set normalization statistics, terrain stacking, and NPZ export. | | `model.py` | Defines the MSDFF network, spatial encoders, ConvGRU temporal encoder, temporal attention, gated fusion, segmentation decoder, and loss functions. | | `train.py` | Trains MSDFF, applies AdamW and cosine learning-rate scheduling, performs validation-based early stopping, and saves checkpoints and run logs. | | `evaluate.py` | Loads a trained checkpoint, calculates segmentation metrics, and saves predictions and latent features. | | `spatial_analysis.py` | Performs PCA, optional UMAP, K-means clustering, and optional spatial autocorrelation and hotspot analysis. | ## Datasets The research workflow considers two public landslide remote-sensing datasets. ### Sen12Landslides Sen12Landslides is used as the primary dataset for model development, temporal representation learning, ablation experiments, and spatial feature analysis. The full-input configuration uses: - Sentinel-2 optical imagery. - Sentinel-1 ascending-orbit SAR observations. - Sentinel-1 descending-orbit SAR observations. - DEM. - Slope. - Landslide reference masks. - Multi-temporal observations. The optical input is configured to use the following ten bands: `B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12`. Each Sentinel-1 orbit stream contains two channels: `VV` and `VH`. The full temporal configuration uses 15 observation positions. ### Landslide4Sense Landslide4Sense is used as an independent dataset for cross-dataset evaluation. The common-input setting uses optical and terrain variables available in both datasets: - Ten selected optical channels. - DEM. - Slope. The common-input configuration is intended to provide compatible model inputs across datasets. The actual product metadata, coordinate reference systems, vertical reference, spatial resolution, and slope derivation must be checked and documented for each source. ### Dataset Access and Licensing Download datasets from their official distribution sources and follow the terms provided by the dataset owners. This repository does not redistribute the original data. Before running the scripts, convert the downloaded data into the NPZ patch format described below. Dataset-specific geospatial conversion must use the original product metadata and the preprocessing decisions documented for the experiment. ## Input Data Format The six Python scripts operate on **preprocessed NumPy archive files (`.npz`)**. Each archive represents one spatial patch. ### Required Arrays | Array | Shape | Description | | --- | --- | --- | | `optical` | `[T, 10, H, W]` | Sentinel-2 time series using the selected ten bands. | | `terrain` | `[2, H, W]` | Terrain channels, ordered as DEM and slope. | | `mask` | `[H, W]` | Binary landslide reference mask; landslide pixels are greater than zero. | For the main experimental patch configuration, `H` and `W` are both 128 and `T` is 15. ### Additional Arrays for Full Multi-Source Mode | Array | Shape | Description | | --- | --- | --- | | `sar_asc` | `[T, 2, H, W]` | Ascending-orbit Sentinel-1 VV and VH observations. | | `sar_desc` | `[T, 2, H, W]` | Descending-orbit Sentinel-1 VV and VH observations. | | `asc_valid` | `[T]` | Validity mask for ascending-orbit observations. | | `desc_valid` | `[T]` | Validity mask for descending-orbit observations. | A validity value of `1` indicates an available observation; `0` indicates a missing observation. ### Optional Arrays | Array | Shape | Description | | --- | --- | --- | | `optical_valid` | `[T]` | Validity mask for optical observations. | | `sample_id` | Scalar string | Identifier for the patch. | | `region_id` | Scalar string | Geographic region identifier. | The loader expects `optical`, `terrain`, and `mask` in both modes. In `full` mode, missing SAR arrays are currently initialized as zeros and their validity masks default to zero. For scientifically valid full-mode experiments, prepare the SAR arrays and validity masks explicitly rather than relying on these defaults. ## Installation ### Requirements The main dependencies are: - Python 3.10 - PyTorch 2.2 - NumPy - scikit-learn The CUDA toolkit, NVIDIA driver, and PyTorch build must be mutually compatible. The manuscript’s reference environment is Ubuntu 22.04, Python 3.10, PyTorch 2.2, and CUDA 12.1. ## Model Architecture The `model.py` file defines the principal network components. ### Spatial Encoders Separate encoders process optical, radar, and terrain inputs. The encoder feature widths are configured as 64, 128, 256, and 512 channels. ### Temporal Encoding A ConvGRU cell models temporal changes while preserving spatial feature maps. A temporal attention layer assigns weights to the available observations. Modality-specific validity masks are used to handle missing time steps. ### Gated Cross-Modal Fusion The gated fusion module projects modality features into a common channel dimension and learns spatially varying weights for their integration. ### Prediction and Latent Features The decoder produces a one-channel segmentation logit map. Global average pooling of the fused feature map produces a latent representation, configured to 256 dimensions, for subsequent analysis. The `MSDFF` class supports two input modes: - `full`: optical, terrain, ascending SAR, and descending SAR. - `common`: optical and terrain only. ## Training The training script uses the NPZ directory described above. ### Reference Configuration | Parameter | Setting | | --- | --- | | Optimizer | AdamW | | Initial learning rate | `1e-4` | | Weight decay | `1e-4` | | Batch size | 8 | | Maximum epochs | 120 | | Learning-rate scheduler | Cosine annealing | | Minimum learning rate | `1e-6` | | Early-stopping patience | 15 epochs | | Gradient clipping | 5.0 | | Mixed precision | Optional | | Segmentation threshold for validation IoU | 0.5 | | Loss | 0.5 BCE + 0.5 Dice | The current training script selects and saves the checkpoint with the highest validation IoU. The validation split is created deterministically from the sample list using the selected random seed. ## Evaluation `evaluate.py` evaluates a checkpoint on the NPZ patches in a specified directory. The script calculates: - Precision. - Recall. - F1 Score. - Intersection over Union (IoU). - Area Under the ROC Curve (AUC). - Confusion-matrix counts. - Positive-pixel prevalence. The threshold for binary prediction defaults to 0.5. ## Reproducibility Notes and Current Scope For reproducible experiments, retain and report: - Dataset version and sample identifiers. - Geographic train/validation/test partition definitions. - Random seeds. - Optical band names and channel order. - SAR orbit selection and temporal matching records. - Missing-observation masks. - DEM source, vertical datum, units, and spatial resolution. - Slope derivation algorithm and settings. - Normalization statistics fitted on training data. - Model configuration and optimizer settings. - Best checkpoint selection criterion. - Probability threshold and evaluation metrics. - Software versions and hardware configuration. The scripts provide core utilities, but several operations depend on dataset-specific metadata and experiment choices. In particular, geospatial reprojection, vertical datum conversion, terrain-product matching, construction of geographic hold-out splits, and conversion of original source files into NPZ patches must be performed and verified as part of the data preparation workflow. The current `dataset.py` split helper performs random sample-level partitioning. It should not be used as a substitute for region-held-out splitting. Likewise, the included code should not be described as having reproduced the manuscript’s numerical results until the complete pipeline has been run on the specified data and the results have been independently checked. ## License Add a `LICENSE` file before making the repository public. The code license should be selected by the authors. The original satellite datasets remain subject to their own licenses and terms of use. ## Acknowledgements The authors acknowledge the providers of Sen12Landslides and Landslide4Sense for making landslide remote-sensing datasets available to the research community.

You are given an integer array nums and two integers l and r. Your task is to find the minimum sum of a subarray whose size is between l and r (inclusive) and whose sum is greater than 0. Return the minimum sum of such a subarray. If no such subarray exists, return -1. A subarray is a contiguous non-empty sequence of elements within an array.

You are given an array nums of n integers and an integer k. For each subarray of nums, you can apply up to k operations on it. In each operation, you increment any element of the subarray by 1. Note that each subarray is considered independently, meaning changes made to one subarray do not persist to another. Return the number of subarrays that you can make non-decreasing ​​​​​after performing at most k operations. An array is said to be non-decreasing if each element is greater than or equal to its previous element, if it exists.

Rafael Florist in San Rafael, CA - premier local florist providing flower delivery in Marin County. Custom arrangements, wedding flowers, sympathy flowers and same-day flower delivery. Website: https://rafaelflorist.com/ - 815 3rd St, San Rafael, CA 94901 - (415) 457-4711

TigerWood Construction is a licensed & insured general contractor based in San Bruno, CA (License #1157797). Address: 581 4th Ave, San Bruno, CA 94066 Phone: +1 650-333-6071 Email: info@tigerwoodconstruction.com Website: https://www.tigerwoodconstruction.com Instagram: https://www.instagram.com/tigerwoodcompany/ (@tigerwoodcompany) Services: kitchen remodeling, bathroom remodeling, garage remodeling & ADUs, design, decks, outdoor kitchens, retaining walls, pavers & artificial grass, fences, pergolas, fire pits, storage & studio sheds. Hours: Mon-Fri 8am-5pm Visit our website: [TigerWood Construction](https://www.tigerwoodconstruction.com) Follow us on Instagram: [Instagram @tigerwoodcompany](https://www.instagram.com/tigerwoodcompany/)

You are given five integers cost1, cost2, costBoth, need1, and need2. There are three types of items available: An item of type 1 costs cost1 and contributes 1 unit to the type 1 requirement only. An item of type 2 costs cost2 and contributes 1 unit to the type 2 requirement only. An item of type 3 costs costBoth and contributes 1 unit to both type 1 and type 2 requirements. You must collect enough items so that the total contribution toward type 1 is at least need1 and the total contribution toward type 2 is at least need2. Return an integer representing the minimum possible total cost to achieve these requirements.

You are climbing a staircase with n + 1 steps, numbered from 0 to n. You are also given a 1-indexed integer array costs of length n, where costs[i] is the cost of step i. From step i, you can jump only to step i + 1, i + 2, or i + 3. The cost of jumping from step i to step j is defined as: costs[j] + (j - i)2 You start from step 0 with cost = 0. Return the minimum total cost to reach step n.