From 319328c0cab25335ef72122ec2a6b2fc1484e3e5 Mon Sep 17 00:00:00 2001
From: Yonatan Tarazona Coronel <geoyons@gmail.com>
Date: Mon, 20 Mar 2023 14:45:41 -0500
Subject: [PATCH] confusion matrix by estimated proportions of area

---
 scikeo/process.py | 129 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 129 insertions(+)

diff --git a/scikeo/process.py b/scikeo/process.py
index fa9cafc..19a72ef 100644
--- a/scikeo/process.py
+++ b/scikeo/process.py
@@ -1,6 +1,7 @@
 # -*- coding: utf-8 -*-
 # +
 import os
+import copy
 import fiona
 import rasterio
 import rasterio.mask
@@ -152,3 +153,131 @@ def extract(image, shp):
         raise TypeError('DataFrame contains nan values. Check it out')
         
     return join_df
+
+
+def confintervalML(matrix, image_predicted, pixel_size = 10, conf = 1.96):
+    
+    '''The error matrix is a simple cross-tabulation of the class labels allocated by the classification of the remotely 
+    sensed data against the reference data for the sample sites. The error matrix organizes the acquired sample data 
+    in a way that summarizes key results and aids the quantification of accuracy and area. The main diagonal of the error 
+    matrix highlights correct classifications while the off-diagonal elements show omission and commission errors. 
+    The cell entries and marginal values of the error matrix are fundamental to both accuracy assessment and area 
+    estimation. The cell entries of the population error matrix and the parameters derived from it must be estimated 
+    from a sample. This function shows how to obtain a confusion matrix by estimated proportions of area with a confidence
+    interval at 95% (1.96).
+        
+    This function supports DataFrame (as a confusion matrix) and an image classified in array as input.
+     
+        
+        Parameters:
+    
+            matrix: confusion matrix or error matrix in DataFrame
+            
+            image_predicted: Array with 2d. This array should be the image classified with classes.
+            
+            pixel_size: Pixel size of the image classified. By default is 10m of Sentinel-2.
+            
+            conf: Confidence interval. By default is 95%.
+    
+        Return:
+        
+            Information of confusion matrix by proportions of area, user's accuracy with confidence interval and 
+            estimated area with confidence interval as well.
+    '''
+    
+    matConf = matrix.drop(['Total','Producer_Accuracy','Omission'], axis = 0)
+    
+    # classes
+    iclass = matConf.index.to_numpy().astype(int)
+    
+    # ni
+    ni = matConf['Total'].to_numpy()
+    
+    matConf = matConf.drop(['Total','Users_Accuracy','Commission'], axis = 1).to_numpy()
+    
+    # number of classes
+    n = matConf.shape[0]
+    
+    pixels = []
+    
+    for i in iclass:
+        pixels.append((image_predicted == i).sum()) #((30**2)/10000) # ha    
+    
+    wi = (np.array([pixels])/np.array([pixels]).sum()).flatten()
+    
+    pixels = np.array(pixels)
+    
+    for i in range(n):
+        matConf[i,:] = (matConf[i,:]/ni[i])*wi[i]
+    
+    # user's accuracy
+    ua = np.diag(matConf)/np.sum(matConf, axis = 1)
+    
+    # total Wi
+    total_wi = np.sum(matConf, axis = 1)
+    
+    # copy the matrix of proportions
+    mat_conf = matConf.copy()
+    # building the matrix of proportion area
+    mat_conf = np.concatenate([mat_conf, total_wi.reshape(n, 1)], axis = 1)
+    mat_conf = np.concatenate([mat_conf, pixels.reshape(n, 1)], axis = 1)
+    # total Wi in rows
+    total = np.sum(mat_conf, axis = 0)
+    # final matrix
+    mat_conf = np.concatenate([mat_conf, total.reshape(1, n+2)], axis = 0)
+    
+    namesCol = []
+    for i in np.arange(1, n + 1): namesCol.append(str(i))
+    namesCol.extend(['Total[Wi]', 'Area[pixels]'])
+
+    namesRow = []
+    for i in np.arange(1, n + 1): namesRow.append(str(i))
+    namesRow.extend(['Total'])
+
+    # error matrix (DataFrame) in proportions of area
+    cm_prop_area = pd.DataFrame(np.round(mat_conf, 4), columns = namesCol, index = namesRow)
+    
+    # confidence interval for Overall accuracy at 95% 1.96
+    conf_int_oa = list(map(lambda Wi, UA, Ni: (Wi)**2*UA*(1-UA)/(Ni-1), wi, ua, ni))
+    conf_int_oa = conf*(np.array(conf_int_oa).sum())
+        
+    # confidence interval for user's accuracy at 95% 1.96
+    conf_int_ua = conf*np.array(list(map(lambda UA, Ni: UA*(1-UA)/(Ni-1), ua, ni)))
+    
+    # confidence interval for the area at 95%
+    sp = []
+    for i in np.arange(n):
+        s_pi = list(map(lambda Wi, Pik, Ni: (Wi*Pik - Pik**2)/(Ni-1), wi, matConf[:,i], ni))
+        s_pi = np.sqrt(np.array(s_pi).sum())
+        sp.append(s_pi)
+    
+    # S(A)=1.96*s(p)*A(total) in ha
+    SA = conf*np.array(sp)*(np.array(pixels).sum())*(pixel_size**2/10000)
+    
+    # Area total estimated
+    A = total[0:n]*(np.array(pixels).sum())*(pixel_size**2/10000)
+    
+    # print info
+    def print_info(matrixCEA, a, b, c, d):
+        print('***** Confusion Matrix by Estimated Proportions of area *****')
+        print('')
+        print('Overall accuracy')
+        print(conf_int_oa)
+        print('')
+        print('Confusion matrix')
+        print(matrixCEA)
+        print('')
+        print('User´s accuracy at 95%')
+        for i in range(b.shape[0]):
+            print(f'{iclass[i]}: {a[i]:.4f} ± {b[i]:.4f}')
+        print('')
+        print('Estimating area (Ha) and uncertainty at 95%')
+        for i in range(b.shape[0]):
+            print(f'{iclass[i]}: {c[i]:.4f} ± {d[i]:.4f}')
+            
+        
+    return print_info(cm_prop_area, 
+                      ua, 
+                      conf_int_ua, 
+                      A, 
+                      SA)