## Prediction Pipeline (Projection)
This file contains the prediction stage and consumes artifacts produced by `bio-oracle-learning.qmd`.

## Load required R packages
```{r}
library(tidyr)
library(dplyr)
library(terra)
library(mregions2)
library(biooracler)
library(stringr)
library(tibble)
library(catboost)
library(sf)
```

## Load shared helpers and define run configuration
```{r}
source("R/shared-utils.R")

config <- list(
    range_shapefile = "data/iucn/Pagophilus_groenlandicus.shp",
    bbox_expand_degrees = 5,
    artifacts = list(
        dynamic_layers = "dynamic_layers.rds",
        subset_layer_names = "subset_baseline_layer_names.rds",
        seal_range_df = "seal_range_df.rds",
        seal_range_raster = "seal_range_raster.tif",
        model = "cat_model.cbm",
        manifest = "artifacts-manifest-learning.csv"
    )
)
```

## Recreate spatial constraints
These bounds are needed for downloading future Bio-ORACLE slices.

```{r}
study_bounds <- make_study_bounds(
    range_shapefile = config$range_shapefile,
    expand_degrees = config$bbox_expand_degrees
)
seal_range <- study_bounds$seal_range
lon_range <- study_bounds$lon_range
lat_range <- study_bounds$lat_range
```

## Load shared artifacts from learning stage
```{r}
required_artifacts <- unlist(config$artifacts[c("dynamic_layers", "subset_layer_names", "seal_range_df", "seal_range_raster", "model")])
assert_required_files(required_artifacts)

cat_model <- catboost.load_model(config$artifacts$model)
subset_baseline_layer_names = readRDS(config$artifacts$subset_layer_names)
seal_range_df = readRDS(config$artifacts$seal_range_df)
seal_range_raster = rast(config$artifacts$seal_range_raster)
dynamic_layers = readRDS(config$artifacts$dynamic_layers)
```

## Shared artifacts manifest (optional inspection)
```{r}
if (file.exists(config$artifacts$manifest)) {
    artifacts_manifest <- utils::read.csv(config$artifacts$manifest)
    artifacts_manifest
}
```

## Prediction function
```{r}
get_prediction = function(ssp_code, decade) {
    ssp_slice = download_biooracle_slice_subset(
        dynamic_layers = dynamic_layers,
        scenario_value = ssp_code,
        decade_start = decade,
        layers_to_download = subset_baseline_layer_names,
        lon_range = lon_range,
        lat_range = lat_range
    )

    ssp_slice_brick = rast(ssp_slice)
    ssp_slice_brick = set_brick_names_with_depth(ssp_slice_brick)

    ssp_slice_df = ssp_slice_brick |>
        as.data.frame(cells = TRUE, xy = TRUE)

    ssp_slice_features = ssp_slice_df |> select(-cell, -x, -y)
    ssp_slice_pool <- catboost.load_pool(data = ssp_slice_features)

    preds_prob <- catboost.predict(cat_model, ssp_slice_pool, prediction_type = "Probability")
    preds_class <- ifelse(preds_prob > 0.5, 1, 0)

    ssp_slice_prediction = ssp_slice_df |>
        mutate(prediction = preds_class) |>
        select(cell, prediction)

    ssp_slice_diff = seal_range_df |>
        left_join(ssp_slice_prediction, by = "cell") |>
        mutate(diff = 2 * target + prediction)

    r = rast(ssp_slice_brick)
    r[ssp_slice_diff$cell] = ssp_slice_diff$diff

    writeRaster(r[[1]], paste0(ssp_code, "-", decade, ".tif"), overwrite = TRUE)

    png(filename = paste0(ssp_code, "-", decade, ".png"), width = 800, height = 800)
    plot(
        r[[1]],
        type = "classes",
        col = c("grey", "green", "red", "purple"),
        levels = c("00", "01", "10", "11"),
        main = paste0(ssp_code, "-", decade)
    )
    dev.off()
}
```

## Example runs
```{r}
get_prediction("ssp585", 2020)
```

```{r}
sapply(seq(2050, 2050, by = 10), function(decade) {
    get_prediction("ssp585", decade)
})
```